How Do Rural Households Achieve Poverty Alleviation? Identification and Characterization of Development Pathways Using Explainable Machine Learning

Abstract

Exploring the dynamic mechanisms of household poverty alleviation is crucial for achieving sustainable poverty reduction and preventing relapse into poverty. However, existing research is often constrained by a static perspective, failing to integrate poverty states with transition processes, and lacking the methodological tools to decipher the nonlinear heterogeneity and spatial dependence inherent in household pathways. This study addresses three critical questions: How can we conceptualize and quantify the dynamic trajectories of household poverty alleviation? What are the key mechanisms that drive households from poverty to stable sustainability? And how do these pathways vary across different spatial contexts? Our analysis, based on an explainable machine learning framework applied to longitudinal data from 107,637 households, yields several key findings. First, household pathways are strongly predicted by their initial typology. Those with heavy burdens and limited labor capacity ($S_{I-4}$) predominantly remained in unstable states (62.5%), while households with human capital advantages ($S_{I-3}$, $S_{I-6}$) achieved stable poverty alleviation directly at rates of 84.9% and 100%, respectively. Second, the transition from instability to stability follows discernible bridging mechanisms, where pathways reliant on skill upgrading prove more decisive for long-term stability than those dependent solely on short-term subsidies. Third, pathways are intrinsically shaped by spatial context, creating a geography of opportunity and risk—from policy compensation in mountainous areas, to resource-institutional synergy in agricultural plains, and labor-market stabilization in mining and peri-urban regions. In conclusion, sustainable poverty alleviation hinges on interventions precisely aligned with both initial household profiles and regional contexts. The central policy implication is to move beyond one-size-fits-all approaches by balancing protective safety nets with capacity-building investments, thereby creating equitable development pathways across diverse geographies.

1. Introduction

Poverty remains a pressing global challenge that undermines sustainable development and social stability [1,2]. Although significant progress has been made, recent global crises have reversed long-term declines in extreme poverty. These crises—including conflict, climate change, and rising food and energy prices—have also heightened vulnerability among marginalized populations [3,4]. In this context, China’s 2020 declaration of absolute poverty eradication marked a historic achievement at the national scale. Yet households that have just exited poverty and low-income groups remain highly susceptible to relapse or new poverty entry. At the same time, international experience shows that diverse strategies can reduce poverty and improve welfare, including agricultural reforms, industrialization, social security, and ecological compensation [5,6]. The central challenge in the post-2020 era, however, lies in consolidating these achievements and systematically distilling successful practices to address the reality that some households have not yet achieved stable and sustainable poverty alleviation, while others continue to face significant risks of falling back into poverty [7].

Early poverty studies primarily defined poverty by setting poverty lines based on the monetary cost of meeting basic consumption needs and minimum economic welfare thresholds [8,9]. While this approach was intuitive, it overlooked the social dimensions of poverty [10,11]. With the deepening of research, the connotation of poverty was expanded to non-economic aspects such as social exclusion [11,12,13]. Sen’s capability approach conceptualizes poverty as deprivation of the capabilities needed to improve circumstances and manage risks. This view helped establish a multidimensional consensus covering education, health, quality of life, and access to public services [14,15,16]. As research perspectives evolved, the focus shifted from poverty alleviation toward poverty prevention, re-entry into poverty control, and stable poverty reduction. Theories of vulnerability, resilience, and sustainable livelihoods were therefore widely introduced into rural poverty studies, laying the foundation for linking poverty reduction to the broader agenda of sustainable development, and have been applied to analyze the risk factors, adaptive capacities, and livelihood strategies of specific groups such as relocated households, mountain farmers, disabled individuals, and landless households [17,18]. At the regional level, research has frequently employed a spatial differentiation perspective to uncover distinct geographical patterns in poverty reduction effectiveness and vulnerability to returning to poverty. Evidence indicates that socioeconomic and geographical disadvantages are primary factors leading to spatial clustering of risks, whereas income stability, healthcare accessibility, and adequate housing constitute essential safeguards for sustainable poverty alleviation, which is central to preventing poverty relapse and achieving long-term development outcomes [19,20,21]. At the household scale, research has emphasized individual and family heterogeneity, showing that re-entry risks are concentrated in areas with fragile natural conditions and weak socioeconomic foundations, while highlighting education, health, social security, and family characteristics as critical influencing factors [7,22,23]. Such studies provide empirical support for differentiated anti-poverty policies. In summary, most existing studies rely on prior theoretical frameworks to construct indicator systems and evaluation standards, using them to predict poverty risks and identify their determinants. However, these studies often remain at a static level and fail to uncover the underlying mechanisms of poverty dynamics. In reality, stable poverty alleviation usually follows a continuous process—from poverty identification and characteristic mitigation to capability enhancement. Yet current research tends to focus on single stages, such as re-entry into poverty, poverty exit, or chronic poverty, while lacking systematic exploration of the sequential changes across the full trajectory and the complex interactions among variables.

From the perspective of methodological applications, poverty research has generally revolved around four core questions: feature identification (what), intensity measurement (how), causal explanation (why), and spatial distribution (where) [24,25]. Within a multidimensional analytical framework, research efforts have predominantly concentrated on defining and quantifying the characteristics and severity of poverty [26,27]. A common methodology involves operationalizing this concept through the construction of an indicator system, which is developed based on the underlying theoretical framework, the specific attributes of the target population, and the study’s overarching goals. This system subsequently enables the identification and measurement of poverty through the defined value thresholds and carefully calibrated weighting of individual indicators. The Alkire–Foster (AF) method is one of the most widely applied multidimensional poverty measurement models, identifying poverty status through a unified threshold [28]. Some extensions introduce multiple stage-specific thresholds and align them with distinct poverty alleviation states, thereby enabling differentiated distance analysis to enhance identification accuracy [29]. With respect to weights, subjective schemes—assigning importance according to normative judgments—are used to facilitate cross-regional comparability [30]. Objective schemes, in contrast, estimate weights from the statistical properties of the data to limit researcher discretion [7]. However, most studies still adopt fixed weight settings. This approach struggles to reflect the heterogeneity in influence mechanisms across different poverty states and their dynamic changes [31]; consequently, it fails to adequately characterize the nonlinear features of continuous state evolution and complex interaction mechanisms in poverty reduction and development trajectories.

Geographically, the occurrence of poverty is closely associated with “geographical capital,” which mainly refers to the natural environment, resource endowments, locational conditions, and infrastructure [32,33]. When geographical capital is persistently deficient or its improvement is constrained, regions are prone to fall into a “spatial poverty trap” [34,35]. Building on this understanding, geographical research has employed spatial distribution analysis, typology, and causal diagnostics to address three fundamental questions: Where is poverty located? What types of poverty exist? And why does poverty occur in particular places? [25]. In China’s poverty alleviation practices over the past four decades, these approaches have been extensively widely applied to delineate poverty areas and classify types, and to diagnose drivers and guide targeted interventions [24]. Rural households’ poverty reduction efforts are also heavily shaped by geographical factors [36,37,38]. The interactions between rural households and their surrounding environment jointly determine diverse pathways out of poverty [39,40,41]. However, while previous studies have largely focused on static associations between poverty distribution and geographical capital, they have paid insufficient attention to systematically examining the spatial dynamics of poverty reduction processes. In particular, little is known about the interaction mechanisms between heterogeneous spatial resource endowments and rural households’ livelihood development. This gap limits the ability to precisely identify dynamic poverty alleviation pathways and assess their sustainability, thereby hindering the design of targeted and durable poverty reduction policies.

The field already provides usable support for anti-poverty action—multidimensional frameworks clarify what to measure, AF-based thresholds and weighting schemes improve comparable identification, and spatial analyses inform policy targeting. Yet three bottlenecks persist at the household scale: (i) evidence and tools remain largely stage-specific, limiting a path-dependent, sequential account from identification and constraint mitigation to capability enhancement; (ii) unified thresholds and fixed weights, though helpful for comparability, cannot capture state-dependent, nonlinear heterogeneity across households and stages; and (iii) pathway-level interactions between heterogeneous geographical endowments and livelihood trajectories over time remain under-identified, constraining fine-grained, resilience-oriented planning across diverse regional contexts.

Machine learning methods, with their powerful nonlinear fitting capabilities and multi-source data fusion features, have significantly improved the efficiency and accuracy of poverty prediction and identification [42]. Building on this progress, CNNs applied to satellite imagery have broadened data sources and offer a credible alternative to costly field surveys in data-scarce settings [43]. In the poorest or rural areas, nighttime lights often offer limited discrimination. By contrast, daytime high-resolution imagery can directly capture visible assets such as roof materials, roads, and infrastructure. When combined with heterogeneous data (e.g., POIs, transport, and administrative or household records), this approach delivers stronger predictive performance and spatial alignment, especially for physically observable “hard” indicators [44]. For multi-source structured data, tree-based models (random forests, gradient boosting, decision trees) effectively capture nonlinearities and interactions and typically outperform traditional regression baselines in accuracy and robustness [45,46]. Nevertheless, most existing models are largely confined to identifying static poverty statuses; they lack adequate analysis of individual heterogeneity and changes in poverty reduction development characteristics, thereby limiting our capacity to reveal the micro-level mechanisms of impoverishment and poverty alleviation.

To understand predictions and identify key poverty drivers, recent studies have introduced interpretable methods. Beyond local techniques such as occlusion and feature-attribution analyses, counterfactual (minimal-change) approaches illustrate which feature adjustments could trigger status transitions. However, these methods often assume feature independence or local linearity, and, without constraining interventions to actionable features, may yield impractical recommendations [47,48]. More importantly, they rarely integrate domain knowledge or systematically test the consistency among predictive outputs, sustainable poverty-alleviation goals, and established poverty theories. Among model-agnostic local tools, LIME and SHAP are the most widely used. LIME approximates complex decision boundaries with local linear surrogates, but its explanations are sensitive to neighborhood sampling and perturbation strategies [49,50]. By contrast, SHAP—grounded in cooperative game theory—decomposes predictions into marginal feature contributions; its explanations satisfy axioms such as local accuracy and consistency, and admit exact solutions for tree models (e.g., random forests, gradient boosting) via TreeSHAP [51,52,53]. Cross-domain applications further show that SHAP reliably identifies key drivers and supports mechanism-oriented communication [54], capturing individual-level heterogeneity as well as nonlinearities and interactions [55,56]. Even so, most prior work applies SHAP primarily for pointwise, local interpretation. Using SHAP attribution vectors as features to cluster in the explanation space enables characterization of group-level differences. We therefore posit—and empirically test—that grouping by “patterns of feature influence,” rather than by raw feature values, more effectively aggregates cases driven by similar mechanisms, thereby providing more interpretable evidence on transitions of poverty states and development pathways.

Drawing on the above research progress and addressing these gaps, we advance a path-aware, explainable agenda at the household scale, linking conceptual framing, model development, and empirical validation. Accordingly, we set out three research objectives, which also constitute this study’s contributions to the literature: First, it introduces the concept of household poverty alleviation and development pathways, integrating both “poverty states” and “poverty processes” into a unified framework that enables a dynamic and systematic understanding of poverty alleviation. Second, it constructs an explainable machine learning framework that combines XGBoost with SHAP to identify heterogeneous state–process transitions, thereby offering both methodological rigor and interpretability. Third, it empirically applies this framework to large-scale, fine-grained monitoring data from Pingdingshan City, uncovering diverse household pathways and their spatial heterogeneity, and generating insights for differentiated poverty prevention and sustainable rural development strategies. The remainder of this paper is structured as follows: Section 2 outlines the conceptual framework and methodological design, Section 3 reports the empirical results, Section 4 discusses implications, and Section 5 concludes.

2. Method

2.1. Conceptual Definition and Analytical Framework

Building on dynamic poverty studies, this research defines household poverty alleviation and development pathways as the dynamic trajectories through which rural households evolve from initial poverty toward sustainable development. These pathways are conceptualized as a combination of states (S), representing the household’s condition at a given point in time, and processes (P), describing the transitions between these states. This perspective draws on Sen’s capability approach [14], the multidimensional poverty framework [26], and recent studies on poverty dynamics [17,21,23], while highlighting the need to analyze both initial conditions and sequential changes.

States and processes are distinct but interdependent. States capture discrete household conditions, including whether a household remains in a Poverty State ($S_P$), has reached a Stable Poverty Alleviation State ($S_S$), or remains in an Unstable Poverty Alleviation State ($S_U$). The Initial State ($S_I$) refers to the household′s condition when first included in the monitoring and support system. In this study, we hypothesize that a household′s pathway comprises a sequence of these annual states and the transitions between them.

However, states alone cannot capture the complexity of pathway dynamics. Therefore, building on recent advances in poverty dynamics [20,26,57,58] and responding to the post-2020 challenge of consolidating achievements and preventing regression [3,7], this study introduces processes to trace the mechanisms of change. We conceptualize processes not only as movements between states but also as the underlying mechanisms that either enable households to consolidate gains into stable poverty alleviation or leave them in precarious, reversible alleviation.

To frame the possible evolution pathways and their driving mechanisms, we further postulate that household trajectories are jointly determined by the initial multidimensional poverty profile and the subsequent livelihood development dynamics, which are themselves shaped by spatial contexts. Specifically, we hypothesize that:

H1: 

The effectiveness of the pathway hinges on the synergistic alignment between poverty alleviation interventions (e.g., reducing living burdens, improving housing & infrastructure) and livelihood development strategies (e.g., enhancing human capital, facilitating industry assistance).

H2: 

The spatial heterogeneity of pathways is systematically influenced by regional resource endowments and market structures, which moderate the efficacy of poverty alleviation mechanisms across mountainous, agricultural, and mining/peri-urban areas.

These hypotheses allow us to conceptualize typical pathways as state-process sequences (see

Table 1. Symbols, states, and processes in household poverty alleviation and development pathways.

Symbol	Type	English Term	Definition
$S_P$	State	Poverty State	Household has not yet escaped poverty
$S_S$	State	Stable Poverty Alleviation State	Household has escaped poverty and is unlikely to fall back
$S_U$	State	Unstable Poverty Alleviation State	Household has escaped poverty but remains at risk of returning
$S_I$	State	Initial State	Household’s status when first included in monitoring and support system
$P^S$	Process	Stable Poverty Alleviation Process	Transition from $S_P$ or $S_U$ to $S_S$, with sustained stability
$P^R$	Process	Return to Poverty Process	Transition from $S_S$ back to $S_P$ or $S_U$ due to internal or external shocks
$P^U$	Process	Unstable Poverty Alleviation Process	Fluctuation between $S_P$ and $S_U$ without reaching $S_S$

for a complete typology), such as the progressive stabilization of $S_P-P^U-P^S$. This formulation provides a testable framework for analyzing how the interplay of initial conditions, dynamic processes, and spatial mechanisms jointly determines pathway outcomes, thereby directly informing the differentiated policy implications discussed in Section 4.

2.2. Study Area and Data

Pingdingshan City, located in Henan Province of central China, serves as the study area. The analysis focuses on rural households located in non-municipal districts of Pingdingshan City, excluding the urban municipal districts. These areas represent the administrative units under county-level jurisdiction, where rural poverty alleviation policies were primarily implemented. The city spans mountainous regions in the west, hilly zones in the center, and flat plains in the east, with mineral resources concentrated in the central and western subregions. Location and village-level maps (Figure 1) illustrate the city’s administrative position as well as variations in slope, cultivated land, and the poverty population. These spatial contrasts highlight the pronounced heterogeneity of natural conditions, resource endowments, and settlement patterns, which shape diverse household livelihoods and offer an appropriate context for analyzing differentiated poverty alleviation pathways.

The empirical analysis is based on longitudinal monitoring data from 2013 to 2024, covering 111,468 rural households, of which 3831 were excluded due to deregistration or migration, resulting in 107,637 valid samples. The dataset was collected through the official government poverty monitoring system, which continuously tracked all registered poor households and related low-income groups. While not nationally representative, it provides comprehensive census-style coverage at the prefecture level. To ensure data quality, households with severe missing information or inconsistent identifiers were excluded, thereby reducing potential bias. The dataset records annual household-level information on demographics, education, health, housing, production resources, infrastructure, social protection, and income. Its large scale, long time span, and fine-grained resolution support robust statistical inference but also provide a rare empirical foundation for applying explainable machine learning, enabling the systematic identification of household states, processes, and pathways across heterogeneous spatial environments.

2.3. Indicator System

This study develops the indicator system to operationalize and empirically test the analytical framework outlined in Section 2.1. The selection and organization of variables across seven dimensions are explicitly designed to measure the core constructs of the initial multidimensional poverty profile and subsequent livelihood development dynamics. This design enables a direct examination of H1 (pathway effectiveness) and facilitates the exploratory analysis of H2 (spatial heterogeneity) through post hoc visualization and comparative case analysis.

The system is grounded in multidimensional poverty theory, the capability approach, and dynamic poverty research. It also incorporates the practical logic of China’s Targeted Poverty Alleviation and Rural Revitalization policies. Consequently, it captures households′ multidimensional characteristics and integrates the processes of poverty reduction and capacity building for development to comprehensively capture heterogeneity across rural households.

The dimensions are strategically grouped to reflect the synergistic mechanisms posited in H1:Dimensions A (Living Burdens) and B (Housing & Infrastructure) primarily quantify the critical constraints and basic living conditions that are the direct targets of poverty alleviation interventions. Dimensions C (Human Capital), D (Production Resources), F (Industry-assisted Measures), and G (Labor & Income) collectively gauge the household′s intrinsic livelihood development capacity and the external supports it receives. Dimension E (Social Protection Measures) captures the compensatory or safety-net mechanisms that underpin both poverty reduction and stability. According to poverty identification criteria, some dimensions are derived from household living conditions and demographic burdens (A. Living Burdens, B. Housing & Infrastructure, C. Human Capital, D. Production Resources), while other dimensions originate from protective measures and targeted assistance in poverty alleviation and rural revitalization (E. Social Protection Measures, F. Industry-assisted Measures, G. Labor & Income). Specifically, the indicator system comprises seven dimensions: A. Living Burdens, reflecting household pressures from illness, education, and elderly care, which often constitute critical constraints to escaping poverty; B. Housing & Infrastructure, indicating residential safety and access to essential public services, representing the basic conditions for alleviation; C. Human Capital, measuring household structure and the quality of labor force, which determines development potential; D. Production Resources, covering land, transportation, and production assets as the basis of livelihood; E. Social Protection Measures, reflecting institutional support such as minimum living allowance, public service positions, and ecological compensation; F. Industry-assisted Measures, representing context-specific industrial support and opportunities for income generation; and G. Labor & Income, capturing employment and income outcomes, which embody the combined effects of internal conditions and external interventions. The complete list of indicators and measurement methods is provided in Appendix A.

In terms of data processing, the indicators include categorical, ratio, and interval variables. All categorical variables were encoded as discrete numerical values to suit the tree-based model: binary variables (e.g., housing safety) used 0/1 encoding, while multi-category variables (e.g., family structure, access road type) were mapped to integers reflecting their logical ordering or complexity. Following this encoding, the processed feature set—comprising standardized continuous variables and encoded ordinal/nominal variables—was assembled for model training. This processing approach is well-suited for tree-based models and provides a solid foundation for subsequent XGBoost classification and SHAP-based interpretability analysis. This multidimensional system provides a comprehensive depiction of household poverty characteristics and provides a solid foundation for subsequent XGBoost classification and SHAP-based interpretability analysis.

2.4. Analytical Framework and Methods

To effectively identify the characteristics of household poverty alleviation and development paths, this study introduces the Rural Household Poverty Alleviation and Development Path Identification Framework (Figure 2). Building on the strengths of the SHAP (Shapley Additive Explanations) method in providing interpretable insights, the framework utilizes multi-stage data processing, model training, and feature attribution analysis to uncover both the key influencing factors and the dynamic characteristics embedded in households′ poverty reduction and development paths. The framework consists of four key components. First, it involves constructing the database and calculating SHAP-based indicator indices. Second, it focuses on identifying the initial household states. Third, it includes recognizing the poverty reduction processes. Finally, it integrates state identification and process recognition to perform a comprehensive analysis of path features, thereby revealing the full structure of household poverty reduction and development paths.

2.4.1. Database Construction and SHAP Value Estimation of Indicators

Based on the proposed indicator system for household poverty alleviation and development paths, the original monitoring data were first cleaned and preprocessed. This stage involved handling missing values, removing outliers, and normalizing variables, which together resulted in the creation of the Feature Database. Missing values were handled through listwise deletion to maintain data integrity. Outliers in numerical variables (e.g., income) were treated using winsorization at the 1st and 99th percentiles. All continuous numerical variables (e.g., income) were standardized to a mean of 0 and standard deviation of 1 to mitigate the influence of varying scales.

Building on this foundation, we employed the XGBoost algorithm for multi-class classification of household poverty states. To benchmark its performance, we also implemented Random Forest and Decision Tree classifiers using scikit-learn. The hyperparameters for all models were optimized through an iterative process of manual tuning combined with cross-validation, aimed at maximizing the macro-average AUC. Hyperparameters were tuned via stratified cross-validation with early stopping, using Macro-F1 on validation folds for model selection; complete configurations and search ranges for XGBoost, Random Forest, and Decision Tree are reported in Appendix A 

Table A2. Hyperparameter Configurations for All Models.

Model	Key Hyperparameters	Value	Description
XGBoost	objective	multi:softprob	Multi-class classification with probability outputs
	num_class	3	Three poverty states
	learning_rate	0.05	Step size shrinkage
	max_depth	8	Maximum tree depth
	n_estimators	200	Number of boosting rounds
	early_stopping_rounds	20	Early stopping patience
	subsample	0.8	Data sampling ratio
	colsample_bytree	0.8	Feature sampling ratio
	lambda	1	L2 regularization term
	tree_method	hist	Histogram-based algorithm
	device	cuda	GPU acceleration
Random Forest	n_estimators	100	Number of trees in the forest
	max_depth	15	Maximum tree depth
	min_samples_split	5	Minimum samples required to split
	random_state	42	Random seed for reproducibility
Decision Tree	max_depth	10	Maximum tree depth
	min_samples_leaf	4	Minimum samples at leaf node
	random_state	42	Random seed for reproducibility

Note: All models were implemented in Python 3.11 (Python Software Foundation, Wilmington, DE, USA). The XGBoost classifier used XGBoost v2.1.1; baseline models (Random Forest and Decision Tree) used scikit-learn v1.3.2; interpretability analysis employed SHAP v0.44.0; clustering was conducted using scikit-learn-extra v0.3.0 (K-Medoids algorithm); and visualization used Plotly v5. Computations were performed on a workstation equipped with an NVIDIA RTX 4090 GPU (NVIDIA Corporation, Santa Clara, CA, USA).

. A comparative analysis of all models′ performance is provided in Section 3.1.

Feature attributions were then obtained from the optimized XGBoost model using the SHAP TreeExplainer with feature_perturbation = “interventional” to calculate exact SHAP values, capturing both the magnitude and direction of each feature′s contribution. In this study, each household is assessed annually and assigned to one of three possible states: $S$ (where $\mathrm{S}\in S_S,S_U{,S}_P$). For a given household m in year t, we obtain three sets of SHAP values for each indicator, denoted as $\left\{{\mathrm{S}\mathrm{H}\mathrm{A}\mathrm{P}}_{i,t}^{(m,S)}\right\}$. These values reflect the relative contribution of individual indicators to the classification of each state, thereby providing a quantitative foundation for subsequent feature identification and clustering analysis.

2.4.2. Identification of Initial State Characteristics

The initial state $S_I$ represents the point at which rural households are first incorporated into the monitoring and assistance system. The poverty characteristics observed at this stage determine the type of policy interventions that households may receive, and they also play a decisive role in shaping subsequent outcomes and the trajectory toward stable poverty alleviation.

For the purpose of initial-state identification, this study focuses on the poverty-class state ($S_P$). Specifically, for household $m$ at the initial time point $t_0$, we select the corresponding SHAP value vector $\left\{{\mathrm{S}\mathrm{H}\mathrm{A}\mathrm{P}}_{i,t_0}^{(m,S_p)}\right\}$. This set of values provides the quantitative basis for identifying and analyzing the key characteristics of households’ starting poverty conditions, which in turn supports the subsequent recognition of development paths.

First, for each indicator, we calculate the Indicator SHAP Influence Coefficient ${\alpha }_i^{\left(S_I\right)}$ under the initial state. This coefficient is obtained by normalizing the absolute SHAP values while retaining their original signs, thereby ensuring that the direction of influence is consistent with the sign of the SHAP attribution. The formula is expressed as follows:

$${\alpha }_i^{\left(S_I\right)}={\displaystyle \frac{{\mathrm{S}\mathrm{H}\mathrm{A}\mathrm{P}}_i^{\left(S_P\right)}}{\sum _{j=1}^n \left|{\mathrm{S}\mathrm{H}\mathrm{A}\mathrm{P}}_j^{\left(S_P\right)}\right|}},i =\mathrm{1,2},\dots ,n$$

(1)

where ${\mathrm{S}\mathrm{H}\mathrm{A}\mathrm{P}}_i^{\left(S_P\right)}$ denotes the SHAP value of the $i$-th indicator under the poverty state $S_P$, and $n$ is the total number of indicators. This formulation normalizes the SHAP values across all indicators, thereby quantifying the relative influence of each indicator on the model’s classification of household states, while preserving the directional information encoded in the sign. A positive coefficient indicates that the indicator tends to classify the household as being in poverty, reflecting poverty-specific characteristics, whereas a negative coefficient suggests that the indicator is associated with poverty alleviation, highlighting a trend toward development.

Subsequently, we compute the Dimension-level SHAP Influence Coefficient ${\beta }_k^{\left(S_I\right)}$ for each dimension. For a given dimension $k$, the coefficient is defined as the sum of the absolute values of the indicator-level influence coefficients within that dimension:

$${\beta }_k^{\left(S_I\right)}={\sum }_{i\in D_k} \left|{\alpha }_i^{\left(S_I\right)}\right|,k =\mathrm{1,2},\dots ,K$$

(2)

where $D_k$ denotes the set of indicators belonging to dimension $k$, and $K$ represents the total number of dimensions. This coefficient effectively captures the relative importance of each dimension in the identification of initial-state poverty characteristics.

After obtaining the indicator influence coefficients ${\alpha }_i^{\left(S_I\right)}$ at the initial state, we applied the K-Medoids clustering algorithm to group the household samples. K-Medoids is a widely used unsupervised learning method that is particularly suitable for high-dimensional datasets containing multiple features. The cluster medoids derived from the algorithm represent the typical poverty characteristics of households within each group. To determine the optimal number of clusters, we employed a combination of the Elbow Plot and the Silhouette Coefficient, complemented by domain knowledge, to ensure a robust and substantively meaningful clustering solution. To further assess robustness, we conducted bootstrap resampling and calculated the Adjusted Rand Index (ARI), which confirmed the consistency of cluster assignments.

Finally, we applied a cumulative selection rule by ranking the absolute values of the indicator influence coefficients in descending order and retaining indicators until the cumulative contribution reached 0.9. For each cluster medoid, we then ranked the dimensions according to their dimension-level influence coefficients, and further ordered the indicators within each dimension by the absolute values of their indicator-level coefficients. Based on these rankings, we constructed heatmaps to visualize the differences in initial-state poverty characteristics across household categories and dimensions. In combination with spatial visualization, this approach allows for a more comprehensive analysis of the spatial distribution of households with distinct initial poverty profiles.

2.4.3. Identification of Process Dynamics Feature

Poverty reduction and development paths primarily reflect the dynamic transitions that households undergo across multiple stages in the pursuit of stable poverty alleviation. To analyze these processes, this study relies on the set of SHAP values corresponding to the stable poverty alleviation state $S_S$, denoted as $\left\{{\mathrm{S}\mathrm{H}\mathrm{A}\mathrm{P}}_{i,t}^{(m,S_S)}\right\}$, which serves as the data foundation for process-level identification. For each household, the complete path is first divided into distinct process stages, with both the starting and ending time points of each stage clearly defined. We then calculate the change in SHAP values for each indicator between the beginning and end of a given process, expressed as $\mathsf{\Delta }{\mathrm{S}\mathrm{H}\mathrm{A}\mathrm{P}}_i^{(m,P)}$. This change measures the variation in the relative contribution of indicator $i$ during process $P$. A positive change indicates that the indicator facilitates the household’s transition toward stable poverty alleviation, whereas a negative change reflects an inhibiting effect that heightens the risk of falling back into poverty. Such differential analysis highlights the increasing or decreasing influence of key features along the evolution of poverty reduction and development paths, thereby providing interpretable evidence to support the examination of dynamic processes across different path types.

To facilitate both cross-sectional comparison and longitudinal trend analysis, we further calculated the indicator-level SHAP influence coefficients ${\alpha }_i^{\left(P\right)}$ and the dimension-level SHAP influence coefficients ${\beta }_k^{\left(P\right)}$, following the same procedure used in the identification of initial-state features. A larger ${\alpha }_i^{\left(P\right)}$ indicates that indicator $i$ plays a more critical role in supporting households’ transition toward stable poverty alleviation during the corresponding stage, whereas a smaller or negative value implies a constraining effect. In parallel, ${\beta }_k^{\left(P\right)}$ captures the dominant role of dimension $k$ within that process stage, thereby providing a clearer understanding of how different dimensions contribute to the evolution of household poverty reduction and development paths.

Based on the indicator-level influence coefficients ${\alpha }_i^{\left(P\right)}$, we applied the K-Medoids clustering algorithm to classify the process-level changes. The number of clusters was determined through a combined assessment using the Elbow Plot, the Silhouette Coefficient, and expert domain knowledge. Following the same ranking principles as in the initial-state analysis, we ordered the dimensions according to their aggregated influence coefficients and then ranked the indicators within each dimension by the magnitude of their ${\alpha }_i^{\left(P\right)}$. On this basis, we constructed heatmaps to visualize the characteristic differences across process categories and further mapped their spatial distributions. This procedure enables a clear identification of how key dimensions and indicators vary dynamically during different poverty reduction and development processes, while also providing insights into the spatial patterns of diverse household trajectories.

2.4.4. Recognition and Analysis of Poverty Reduction and Development Path Features

By combining the results of initial-state feature identification with the analysis of process-level dynamics, this study obtains complementary information on the key characteristics of households at both the starting point and during subsequent transitions. Integrating these two sources of information enables us to systematically capture the full set of feature patterns that characterize the evolution from the initial poverty state through all intermediate processes to the attainment of stable poverty alleviation. This comprehensive recognition of poverty reduction and development path features not only provides a holistic understanding of the dynamic mechanism of household poverty alleviation, but also establishes a solid foundation for subsequent path classification and the design of differentiated policy measures.

To provide an intuitive representation of path transitions, this study employs the Sankey Diagram as the primary visualization tool to depict the flows and proportions of transitions across different states. On the left side of the diagram, the initial state categories of households at the time of their inclusion in the poverty reduction monitoring system are displayed, while the subsequent process types are shown sequentially from left to right. The width of each flow line directly reflects the magnitude of transitions between categories, thereby revealing the dominant trends through which households move from their initial states into different processes. Building on this overall visualization of transition patterns, we further select several representative paths and conduct spatial distribution analyses. This approach allows for a more detailed deconstruction of household poverty reduction and development path features, as well as a deeper understanding of their spatial differentiation.

3. Results

3.1. Overall Characteristics of Key Indicators Across the Three States

The household poverty reduction and development state classification model was developed using the XGBoost algorithm, which demonstrated superior performance (macro-average AUC = 0.97) compared to Random Forest (0.95) and Decision Tree (0.91) benchmarks. The model exhibited strong performance in distinguishing among the three states. Specifically, the area under the ROC curve (AUC) for the poverty state ($S_P$) is 0.99, indicating highly precise identification. The stable poverty alleviation state ($S_S$) and the unstable poverty alleviation state ($S_U$) achieved AUC values of 0.96 and 0.94, respectively. Although slightly lower, these values still reflect robust classification performance. The identification of absolute poverty is based on a well-defined national poverty standard, which provides clear and consistent criteria. In contrast, the classification of stable and unstable poverty alleviation states depends more heavily on the subjective judgment and practical experience of local officials and field workers, which increases similarity between these two categories in the feature space. Nevertheless, the model maintains high recognition accuracy across all states, enabling us to reliably map states and processes that are directly relevant to sustainable poverty alleviation.

Building on this validated performance, we further employed the SHAP method to examine feature importance. Figure 3 presents the ten most influential indicators for each state, along with the distribution of SHAP values and the corresponding effects of different feature values. Specifically, the upper X-axis in each figure denotes the absolute mean SHAP contribution of each indicator across all samples, while the lower X-axis illustrates the distribution of SHAP values associated with different feature levels. This visualization highlights both the relative importance of indicators under different states and the directional nature of their contributions, providing an intuitive understanding of the feature differences that characterize poverty reduction and development states and how these differences underpin (un)sustainable trajectories.

In the poverty state ($S_P$), access to safe drinking water (B4) emerges as a distinctive indicator relative to the stable ($S_S$) and unstable ($S_U$) poverty alleviation states. Regarding the distribution of SHAP values, both per capita net income (G2) and residential electricity access (B2) display wide variation in both positive and negative directions. Indicators such as social assistance participation (E1), household size (C1), and village poverty exit status (D6) show broadly positive SHAP distributions, whereas household size (C1) and access road type (D3) exhibit stronger negative contributions. For most indicators, SHAP values change monotonically with feature levels, but disability burden (A3) and access road type (D3) present more complex, non-monotonic patterns. Overall, the classification of poverty states relies primarily on clearly defined indicators linked to basic living burdens (Category A), emphasizing fundamental livelihood needs whose remediation constitutes the necessary baseline for any sustainable poverty alleviation.

In the unstable poverty alleviation state ($S_U$), several policy-related support indicators are uniquely important, including photovoltaic project participation (E3), agro-processing assistance (F2), and the dilapidated housing renovation project (B3). In terms of SHAP value distributions, the allowance for needy families (E1) shows wide variation in both directions. Public welfare employment participation (E2), photovoltaic project participation (E3), and household size (C1) display higher contributions on the positive side, whereas residential electricity access (B2), per capita net income (G2), and access road type (D3) show stronger negative contributions. Notably, indicators such as public welfare employment participation (E2), allowance for needy families (E1), agro-processing assistance (F2), household size (C1), and the dilapidated housing renovation project (B3) exhibit complex and heterogeneous relationships with SHAP values. To promote stable poverty alleviation among households in the unstable poverty reduction state, the provision of public welfare employment opportunities and industrial support during periods of unstable development or insecure income sources provides the necessary buffer conditions that bridge unstable conditions toward sustainable poverty alleviation, especially when livelihoods are intermittently insecure.

In the stable poverty alleviation state ($S_S$), distinctive indicators include production electricity access (D5), gender ratio (C3), and the education level of the labor force (C5). The access road type (D3) rises sharply in importance, ranking highest with a broad positive contribution range. Meanwhile, per capita net income (G2) shows wide variation in both positive and negative directions, while residential electricity access (B2) exhibits more extensive negative distributions. Moreover, indicators such as disability burden (A3), public welfare employment participation (E2), and gender ratio (C3) display diverse and non-linear contribution trends as their feature values increase. The determination of the stable poverty alleviation state is shaped mainly by indicators of the production environment (Category D) and human capital (Category C), suggesting that strengthening both human resources and production conditions is essential to secure lasting and sustainable poverty alleviation outcomes by enhancing household resilience to shocks.

Overall, per capita net income (G2), residential electricity access (B2), access road type (D3), and village poverty exit status (D6) consistently rank among the top ten indicators across all three states in terms of SHAP values. From the perspective of indicator composition, as households transition from the poverty state toward the stable poverty alleviation state, the top ten indicators shift in emphasis—from those primarily reflecting housing facilities and living burdens to those highlighting development-support measures and production resources.

3.2. Identification of Initial State and Dynamic Process Features

The optimal cluster numbers for the initial state and the three poverty alleviation processes were determined by jointly considering the Elbow Plot, the Silhouette Coefficient, and the interpretability of the results. The analysis identified six clusters for the initial state, seven for the stable poverty alleviation process, three for the return-to-poverty process, and four for the unstable poverty alleviation process.

As shown in Figure 4 and Figure 5, the cluster centers of the initial state and their spatial distributions are presented. In the heatmaps, positive SHAP values indicate features contributing to the classification of households as being in the poverty state ($S_p$), whereas negative values represent advantageous features that facilitate poverty alleviation and signal higher potential for sustainability if maintained or strengthened.

Regarding group size, $S_{I-2}$ and $S_{I-5}$ account for the largest shares, at 28% and 26% respectively, while $S_{I-6}$ represents the smallest cluster. In terms of overall feature differences, $S_{I-2}$ is clearly distinct from the others: most of its indicator influence coefficients are negative, suggesting relative advantages in poverty alleviation, although it is still identified as part of the unstable poverty alleviation state. The key poverty-related features of this cluster are primarily associated with G1: Average Annual Work-Months and D3: Access Road Type.

For the remaining clusters, the most influential features are concentrated in the Housing & Infrastructure dimension (B2: Residential Electricity Access, B4: Safe Drinking Water Access), the Production Resources dimension (D5: Production Electricity Access, D6: Village Poverty Exit Status), and the Labor & Income dimension (G1: Average Annual Work-Months, G2: Per Capita Net Income). More specifically:

• $S_{I-1}$ shows pronounced poverty-related characteristics in the Human Capital dimension (C1: Household Size, C2: Family Structure Type).
• $S_{I-3}$ demonstrates relative advantages within the Social Protection Measures dimension (E1: Allowance for Needy Families) and several indicators of the Living Burdens dimension (A1: Chronic Disease Burden, A9: Dependency Ratio), yet still exhibits a poverty-related feature in B4: Safe Drinking Water Access (Housing & Infrastructure).
• $S_{I-4}$ and $S_{I-5}$ are both strongly associated with E1: Allowance for Needy Families (Social Protection Measures). However, $S_{I-4}$ shows additional poverty-related characteristics in C2: Family Structure Type (Human Capital) and A3: Disability Burden (Living Burdens), whereas $S_{I-5}$ is characterized by poverty in C1: Household Size (Human Capital) and B4: Safe Drinking Water Access (Housing & Infrastructure), but displays an alleviation advantage in A1: Chronic Disease Burden (Living Burdens).
• $S_{I-6}$ exhibits advantages across multiple dimensions, yet still shows a relatively high poverty-related feature in B1: Housing Safety (Housing & Infrastructure).

From a spatial perspective, households in all clusters are predominantly concentrated in mountainous and resource-dependent areas. Within this distribution, $S_{I-1}$ and $S_{I-4}$ are more prevalent in regions with abundant arable land resources, while $S_{I-3}$, $S_{I-5}$, and $S_{I-6}$ exhibit more spatially clustered patterns.

As shown in Figure 6 and Figure 7, the cluster centers of the stable poverty alleviation process ($P^S$) and their spatial distributions are illustrated. In terms of group size, $P^{S-1}$ (20%), $P^{S-5}$ (38%), and $P^{S-6}$ (18%) constitute the major pathways through which households achieve stable poverty alleviation. Regarding feature dynamics, all cluster centers exhibit positive changes in their dominant dimensions. Among them, $P^{S-2}$, $P^{S-3}$, $P^{S-4}$,and $P^{S-7}$ are primarily characterized by improvements in the Production Resources dimension (Category D), while $P^{S-1}$, $P^{S-5}$, and $P^{S-6}$ display more significant changes in the Labor & Income dimension (Category G). With respect to specific indicator composition:

• $P^{S-1}$ shows marked changes in D1: Cultivated Land Area (Production Resources), C1: Household Size (Human Capital), and C2: Family Structure Type (Human Capital).
• Both $P^{S-2}$ and $P^{S-4}$ exhibit positive changes in D2: Other Agricultural Land Area (Production Resources), indicating that households in these clusters primarily rely on diversified farmland resources—such as forestry, orchards, and aquaculture—for achieving stable poverty alleviation. Beyond this common foundation, their feature dynamics diverge: $P^{S-2}$ is mainly characterized by improvements in F2: Agro-processing Assistance (Industry-Assisted Measures), C1: Household Size (Human Capital), and C3: Gender Ratio (Human Capital), whereas $P^{S-4}$ shows positive changes in D5: Production Electricity Access (Production Resources), C6: Skilled Labor Share (Human Capital), and C8: Participation in Education Aid Programs (Human Capital).
• $P^{S-3}$ is characterized by improvements in D4: Distance to Main Village Road (Production Resources), D6: Village Poverty Exit Status (Production Resources), G2: Per Capita Net Income (Labor & Income), and C2: Family Structure Type (Human Capital).
• $P^{S-5}$ is more strongly associated with positive changes in G2: Per Capita Net Income (Labor & Income), G1: Average Annual Work-Months (Labor & Income), and C2: Family Structure Type (Human Capital).
• Both $P^{S-6}$ and $P^{S-7}$ exhibit substantial positive changes in the Social Protection Measures dimension (Category E), underscoring the importance of policy-driven support in sustaining poverty alleviation. Their specific feature dynamics, however, differ: $P^{S-6}$ is characterized by improvements in E1: Allowance for Needy Families and E2: Public Welfare Employment Participation, accompanied by gains in G2: Per Capita Net Income (Labor & Income), whereas $P^{S-7}$ is distinguished by notable changes in E3: Photovoltaic Project Participation and E4: Ecological Compensation Participation, along with a significant improvement in B1: Housing Safety (Housing & Infrastructure). This comparison suggests that while both clusters rely heavily on social protection programs, $P^{S-6}$ reflects a pathway centered on livelihood security through cash and employment support, whereas $P^{S-7}$ highlights ecological and infrastructural interventions as key drivers of stable poverty alleviation with co-benefits for sustainable development.

From a spatial perspective, households in $P^{S-1}$ to $P^{S-4}$ are relatively more concentrated in areas with abundant arable land resources or within the plains, compared with other clusters, whereas the distributions of $P^{S-5}$ to $P^{S-7}$ broadly correspond to the overall settlement pattern of poor households in the study area.

Figure 6. Heatmap of Cluster Centers in the Stable Poverty Alleviation Process.

Figure 6. Heatmap of Cluster Centers in the Stable Poverty Alleviation Process.

Figure 7. Spatial Distribution of Clustering Results in the Stable Poverty Alleviation Process.

Figure 7. Spatial Distribution of Clustering Results in the Stable Poverty Alleviation Process.

As shown in Figure 8 and Figure 9, the cluster centers of the return-to-poverty process ($P^R$) and their spatial distributions are illustrated. Overall, households experiencing re-poverty are primarily associated with disadvantages in the Living Burdens dimension (Category A), the Labor & Income dimension (Category G), and the Housing & Infrastructure dimension (Category B). Among these, $P^{R-2}$ is the dominant type, accounting for 65% of the households undergoing this process.

In terms of structural characteristics, $P^{R-1}$ is mainly characterized by vulnerabilities in housing and infrastructure conditions, particularly B2: Residential Electricity Access, B1: Housing Safety, and D3: Access Road Type. The re-poverty features of $P^{R-2}$ are reflected in strong fluctuations in employment and income indicators, namely G1: Average Annual Work-Months and G2: Per Capita Net Income. By contrast, $P^{R-3}$ is closely associated with disadvantages in living burdens, especially A3: Disability Burden.

From a spatial perspective, households in $P^{R-1}$ are relatively more concentrated in mountainous and hilly areas, whereas those in $P^{R-2}$ and $P^{R-3}$ are more dispersed across the study area. The results suggest that re-poverty risks exhibit pronounced spatial patterns in relation to housing and infrastructure conditions, whereas the spatial differentiation of other livelihood-related factors is relatively weak.

As shown in Figure 10 and Figure 11, the cluster centers of the unstable poverty alleviation process ($P^U$) and their spatial distributions are presented. Overall, with the exception of $P^{U-3}$, most cluster centers are dominated by positive changes in indicators from the Production Resources dimension (Category D) and the Labor & Income dimension (Category G). In contrast, $P^{U-3}$ is mainly characterized by re-poverty features associated with increases in the Living Burdens dimension (Category A, e.g., A1: Chronic Disease Burden; A3: Disability Burden) and reductions in G1: Average Annual Work-Months (Labor & Income).

In terms of structural differences, $P^{U-1}$ and $P^{U-2}$ are distinguished by the positive role of Social Protection Measures (Category E) combined with improvements in the Living Burdens dimension (Category A) in facilitating transitions toward stable poverty alleviation. Specifically, $P^{U-1}$ is characterized by contributions from E2: Public Welfare Employment Participation together with improvements in disease-related burdens (A1: Chronic Disease Burden; A3: Disability Burden), whereas $P^{U-2}$ is more strongly associated with E4: Ecological Compensation Participation, alongside contributions from G2: Per Capita Net Income and G1: Average Annual Work-Months. By contrast, $P^{U-4}$ highlights improvements in the Human Capital dimension (Category C), particularly C2: Family Structure Type and C8: Participation in Education Aid Programs, reflecting enhanced labor quality and educational support. Among the $P^U$ types, $P^{U-1}$ and $P^{U-2}$ serve as policy-enabled transitions toward sustainability, whereas $P^{U-3}$ flags medical and dependency burdens that erode sustainability unless addressed.

From a spatial perspective, households in $P^{U-2}$ and $P^{U-4}$ are relatively more concentrated in mountainous and hilly regions, whereas those in $P^{U-1}$ and $P^{U-3}$ are more scattered, indicating stronger spatial heterogeneity in unstable poverty alleviation pathways.

3.3. Identification of Poverty Alleviation and Development Path Features

The clustering results of the initial state ($S_I$) and different process types ($P^S$, $P^R$, $P^U$) reveal the key characteristics and categorical differences in households’ poverty alleviation and development. However, feature extraction based on a single state or a single process is insufficient to fully uncover the underlying mechanisms of the complete trajectory. Instead, households’ poverty alleviation and development paths are composed of dynamic combinations of multiple states and processes, exhibiting distinct stage-specific patterns and household-level heterogeneity. Therefore, building on the preceding identification results, this section integrates households’ initial states and dynamic processes, constructs path representations and typologies, and identifies characteristic patterns of path variation.

Figure 12 illustrates the overall distribution of different household poverty reduction and development path types. In general, the trajectories can be classified into four categories: $S_I-P^S$ (53.5%), $S_I-P^U$ (25.1%), $S_I-P^U-P^S$ (21.2%), and $S_I-P^S-P^R-P^S$ (0.2%). Examining the preferences across different initial states, the $S_I-P^S$ path type is most prevalent among households in $S_{I-6}$ (100%), $S_{I-3}$ (84.9%), $S_{I-5}$ (64.5%), and $S_{I-1}$ (42.8%), indicating a greater tendency for these groups to achieve stable poverty alleviation directly. In the $S_I-P^U-P^S$ category, households in $S_{I-1}$ (30.2%) and $S_{I-2}$ (27.9%) account for the largest shares, suggesting that these groups tend to pass through an unstable poverty alleviation stage before attaining stability. By contrast, within the $S_I-P^U$ type, $S_{I-4}$ represents the dominant group (62.5%), which remains in an unstable poverty alleviation state at present.

Figure 13 illustrates the variation within the $S_I-P^S$ path type. In terms of distribution across $P^S$ subcategories, $P^{S-5}$ accounts for the largest share (35.4%), followed by $P^{S-1}$ (21.6%) and $P^{S-6}$ (16.9%). Across different initial states, households in $S_{I-2}$, $S_{I-4}$, and $S_{I-6}$ tend to rely on a single $P^S$ subtype to achieve stable poverty alleviation, whereas those in $S_{I-1}$, $S_{I-3}$, and $S_{I-5}$ display more diversified pathways toward stability. The predominance of $P^{S-5}$ and $P^{S-1}$ within $S_I-P^S$ indicates that stable income and capability accumulation are central to sustainable exits.

Figure 14 illustrates the variation relationships within the $S_I-P^U$ and $S_I-P^U-P^S$ path types. Regarding the correspondence from the initial state to the unstable poverty alleviation process, 88.5% of $S_{I-4}$ households transition into $P^{U-1}$, while the majority of $S_{I-5}$ (68.9%) transition into $P^{U-2}$. Most households in $S_{I-2}$ (89.6%) and $S_{I-3}$ (93.9%) are associated with $P^{U-4}$, whereas $S_{I-1}$ is more evenly distributed between $P^{U-2}$ (38.8%) and $P^{U-4}$ (43.7%). In terms of transitions from the unstable poverty alleviation process to the stable poverty alleviation process, both $P^{U-3}$ and $P^{U-4}$ primarily shift to $P^{S-5}$ (51.3% and 57.8%, respectively) and $P^{S-1}$ (22.3% and 13.0%). $P^{U-1}$ is predominantly linked to $P^{S-6}$ (85.7%). By contrast, $P^{U-2}$ shows a more diverse transition pattern, with 32.7% shifting to $P^{S-6}$, 26.0% to $P^{S-1}$, and 21.3% to $P^{S-5}$. Overall, the preferences of different initial states in selecting $P^S$ categories for achieving stable poverty alleviation are broadly consistent with the patterns observed in the direct ${\mathrm{S}}_{\mathrm{I}}-{\mathrm{P}}^{\mathrm{S}}$ pathways. Transitions from $P^{U-1}$, $P^{U-2}$, and $P^{U-4}$ into $P^{S-5}$, $P^{S-1}$, and $P^{S-6}$ reveal concrete bridging routes to sustainability, while $P^{U-3}$ requires burden-reduction policies to avoid relapse.

Within the $S_I-P^S-P^R-P^S$ path type (Figure 15), a total of 264 households were identified. Among them, $S_{I-5}-P^{S-5}-P^{R-2}-P^{S-5}$ constitutes the largest group (57%), followed by $S_{I-4}-P^{S-6}-P^{R-1}-P^{S-6}$ (18%), which together represent the dominant patterns of this trajectory. In terms of structural characteristics, a notable feature of this path type is that the two $P^S$ subcategories at the beginning and end remain identical, indicating a relatively stable configuration of poverty alleviation processes and suggesting that re-establishing the same P^S subtype can effectively reconsolidate sustainability after shocks.

3.4. Characteristics of Typical Poverty Alleviation and Development Paths

The formation of household poverty reduction and development paths is influenced by multiple factors, including initial characteristics, natural–social environments, and external policy and industrial conditions. Consequently, these paths show high complexity and heterogeneity. To capture their essential patterns, this section focuses on representative groups with significant features, analyzing their main path variations and spatial distribution characteristics.

The $S_{I-4}$ and $S_{I-5}$ groups, both identified at the initial state as households receiving minimum living allowances, exhibit markedly different outcomes in their subsequent poverty reduction and development paths. $S_{I-4}$ forms the largest cluster still trapped in the unstable poverty alleviation process ($P^U$), with 62.5% of households yet to achieve stability. In contrast, $S_{I-5}$ demonstrates a relatively high level of stability, with 78.5% of households having successfully transitioned into the stable poverty alleviation state ($S_s$). Among the remaining groups, $S_{I-1}$ and $S_{I-3}$ display pronounced differences in the Human Capital dimension (Category C), which are reflected in their divergent path outcomes: while 99.5% of households in $S_{I-3}$ have achieved stable poverty alleviation, 27% of those in $S_{I-1}$ remain in the unstable state. Accordingly, these four initial-state groups are selected as representative cases to contrast sustainability potentials, dependencies, and failure risks across pathways for further analysis of poverty reduction and development path features.

From the perspective of poverty reduction and development paths, both $S_{I-4}$ (37%) and $S_{I-5}$ (31.2%) primarily rely on the $P^{S-6}$ pathway to achieve stable poverty alleviation. However, the $S_{I-5}$ group also attains stability through more diversified routes, including $P^{S-5}$ (19.3%), $P^{S-1}$ (15.8%), and $P^{S-2}$ (7.2%), reflecting a broader range of path choices. With regard to unstable poverty alleviation processes, 71.4% of $S_{I-4}$ households transitioned through $P^{U-1}$, of which 16.6% ultimately reached stable poverty alleviation; in contrast, 24.1% of $S_{I-5}$ households experienced $P^{U-2}$, with 9.6% eventually achieving stability. Overall, the $S_{I-4}$ group depends heavily on policy-oriented measures such as public welfare employment (E2) and social assistance (E1), indicating a poverty reduction and development path that is strongly reliant on external support and therefore exhibits weaker sustainability due to limited endogenous capacity. By comparison, the $S_{I-5}$ group, mainly located in mountainous areas (Figure 5), demonstrates relative advantages in human resources and is thus better positioned to integrate local resource endowments and achieve stable poverty alleviation through industry-driven development that strengthens endogenous resilience and long-term sustainability.

In terms of the pathways leading to stable poverty alleviation, $S_{I-1}$ households are more likely to experience the $P^U$ process before achieving stability, with a relatively balanced distribution between $S_{I-1}-P^S$ (42.8%) and $S_{I-1}-P^U-P^S$ (30.2%). By contrast, $S_{I-3}$ households predominantly achieve stability directly through $S_I-P^S$ (84.9%). Regarding the choice of $P^S$ subtypes, $S_{I-1}$ mainly follows $P^{S-1}$ (24.0%) and $P^{S-5}$ (21.4%), whereas $S_{I-3}$ relies more heavily on $P^{S-1}$ (43.2%) and $P^{S-5}$ (16.0%), with additional contributions from $P^{S-2}$ (17.7%) and $P^{S-3}$ (17.4%). In terms of livelihood strategies, crop cultivation and off-farm wage employment constitute the predominant options for most poor households. Benefiting from stronger human capital, however, $S_{I-3}$ households are able to combine policy support with local resources to pursue more diversified development pathways. For $P^U$ subtype selection across initial states, $S_{I-3}$ shows a preference for $P^{U-4}$ (14.1%), while $S_{I-1}$ is more inclined toward $P^{U-2}$ (22.2%) and $P^{U-4}$ (25.0%). Notably, the share of $S_{I-1}-P^{U-2}-P^S$ (10.8%) is substantially lower than that of $S_{I-1}-P^{U-4}-P^S$ (14.2%) and $S_{I-3}-P^{U-4}-P^S$ (13.7%), suggesting that improvements in labor skills play a more decisive role in promoting stable poverty alleviation and in converting fragile transitions into sustainable, non-relapse outcomes.

From the perspective of the spatial distribution of pathways to stable poverty alleviation across the four initial states (Figure 16 and Figure 17), $P^{S-1}$, $P^{S-2}$, $P^{S-3}$ and $P^{S-6}$ show similar spatial patterns despite differences in their realization rates among household types. In contrast, the spatial distribution of $P^{S-5}$ exhibits a certain degree of randomness. Moreover, the distribution of rural livelihoods dominated by agriculture and sideline industries is largely shaped by regional resource endowments. The spatial distribution of households’ livelihood development pathways reflects the same pattern.

From the perspective of spatial differences in the number of pathways to stable poverty alleviation, $S_{I-3}$, $S_{I-4}$, and $S_{I-5}$ generally exhibit more significant spatial clusters of high values across different pathways, whereas $S_{I-1}$ shows relatively fewer. $S_{I-3}$ and $S_{I-5}$, with relative advantages in human capital and lighter household burdens, often enable poor households within resource-rich villages and under policy support to achieve stable poverty alleviation through similar pathways. $S_{I-4}$, which achieves stable poverty alleviation mainly through $P^{S-6}$, relies heavily on government protective support policies and regional public welfare positions. In mining areas and ecological protection zones in particular, these regions provide more welfare positions and financial assistance, thereby generating higher numbers of households lifted out of poverty through this specific pathway. In contrast, $S_{I-1}$, which is more spatially dispersed, commonly faces serious human capital constraints. Since the conditions of these households largely fall outside the coverage of protective support policies and they lack competitiveness, their stable poverty alleviation is more often achieved through crop cultivation and off-farm wage employment. Spatial clusters where policy-led $P^{S-6}$ dominates signal sustainability contingent on institutional support, whereas clusters with income-/capability-led $P^{S-1}$ and $P^{S-5}$ indicate stronger endogenous sustainability.

4. Discussion and Implications

4.1. Mechanisms of Pathways to Stable Poverty Alleviation

(1)

The process of achieving stable poverty alleviation consistently centers on the fundamental goal of eradicating poverty (Figure 18). Accurate identification of households’ initial poverty characteristics is a prerequisite for formulating effective policies and implementing targeted interventions. During China’s stage of Targeted Poverty Alleviation, the standard of poverty identification was established as the basic criterion for evaluating household poverty status [59,60]. This standard consisted of two major aspects: the minimum guarantee of basic living conditions, including food and essential daily necessities; and the assurance of access to fundamental public services, particularly in education, healthcare, and housing.

In terms of initial states, most households face two major constraints: deficits in Housing and Infrastructure and burdens in Living Burdens. Deficits in Housing and Infrastructure are mainly reflected in inadequate access to B4. Safe Drinking Water Access, B2. Residential Electricity Access, and B1. Housing Safety. With the implementation of the New Rural Construction Program, these problems have been effectively alleviated or largely resolved for many households [61,62], a pattern also reflected in Korea’s Saemaul Undong, where improvements in rural housing and infrastructure were closely linked with community mobilization to address structural disadvantages and promote sustainable development [63]. Nevertheless, some households ($S_{I-6}$) still rely on protective support measures—most notably relocation-based poverty alleviation projects and large-scale dilapidated housing renovation programs that targeted over 7.9 million rural households ($P^{S-7}$)—to achieve stable poverty alleviation [64]. However, such reliance on relocation and large-scale housing renovation indicates a policy-dependent form of stability, raising questions about the long-term sustainability of these outcomes once intensive interventions are withdrawn. To secure long-term resilience for these households, subsequent efforts should integrate an asset-building approach into project designs. For instance, pairing housing improvements with skills training for home-based businesses or support for courtyard economies can help transform fixed assets into sustainable income-generating capabilities.

Living Burdens are primarily associated with medical, educational, and dependency-related pressures. Within this dimension, A5/A6/A8: Education Expenditure Burden and A2: Major Illness Burden are less dominant features of households’ initial poverty states, largely due to China’s nine-year compulsory education policy and catastrophic-illness insurance, which have substantially reduced these costs [65,66]. During the targeted poverty alleviation stage, for instance, more than 20 million impoverished patients received medical treatment, significantly reducing the medical burden on poor households [67]. By contrast, A1: Chronic Disease Burden, A3: Disability Burden, and A9: Dependency Burden remain significant for many households. In addition, household demographic characteristics within the Human Capital dimension—such as C1: Household Size, C2: Family Structure Type, and C3: Gender Ratio—further reinforce these burdens. A subset of households (notably $S_{I-4}$ and $S_{I-6}$) remain dependent on protective policy measures—such as E1: Allowance for Needy Families and E2: Public Welfare Employment Participation—to achieve stable poverty alleviation. Nationwide, nearly 20 million poor people were covered by subsistence allowances and extreme poverty support during the targeted poverty alleviation stage, underscoring the critical role of protective transfers for these vulnerable households [68]. For this group, future support should maintain safety-net continuity while building “capability escalators.” A practical approach is to redesign public-welfare jobs (E2) as paid apprenticeships that integrate targeted vocational training and create clear pathways into the open labor market.

(2)

Achieving stable poverty alleviation not only requires mitigating or eliminating poverty itself, but also relies on stable Livelihood to sustain continuous improvements in living standards. From the perspective of livelihood development, two types can be distinguished: Social Protection Measures and Industry-Assisted Measures.

Pathways under Social Protection Measures ($P^{S-6}$) are generally characterized by reliance on E1: Allowance for Needy Families and E2: Public Welfare Employment Participation, which provide essential assistance for groups facing heavy livelihood burdens and limited labor capacity [59,68]. Comparable to Ethiopia’s Productive Safety Net Program (PSNP), these social protection pathways highlight the critical role of short-term safety nets in preventing immediate deprivation, while also revealing their limited capacity to ensure long-term resilience [69,70]. Development trajectories of the $P^{S-7}$ type highlight the importance of locally distinctive resource-based initiatives—such as E3: Photovoltaic Project Participation and E4: Ecological Compensation Participation—in supporting poverty alleviation in disadvantaged mountainous and ecological protection areas [71,72,73]. In practice, more than 1.1 million poor individuals were employed as ecological forest rangers nationwide, providing stable income opportunities for groups with limited labor capacity [74]. These pathways highlight the irreplaceable role of safety nets, but also underscore that without parallel capacity-building, their sustainability remains fragile. Unstable poverty alleviation processes ($P^{U-1}$, $P^{U-2}$, $P^{U-4}$) further reveal the transitional role of short-term employment and subsidy support, combined with improvements in labor skills (C5: Education Level of Labor Force, C6: Skilled Labor Share), in facilitating households’ eventual movement from unstable to stable poverty alleviation through the $S_I-P^U-P^S$ pathway. These approaches remain highly dependent on policy interventions and exhibit limited resilience to external shocks, thereby retaining unstable characteristics. Similar forms of dependence on protective support and locally embedded resource utilization have also been documented in other poverty alleviation studies, underscoring the broader relevance of these mechanisms beyond the local context [75,76,77]. This finding underscores the need to establish clear “graduation pathways.” These pathways would bundle conditional transfers with mandatory skills upgrading, setting clear timelines and expectations to guide households from temporary assistance toward stable employment or entrepreneurship.

Industry-Assisted Development pathways primarily target households with development potential, who make livelihood choices based on their production resources and environmental conditions. These choices include crop cultivation, agro-processing activities, rural tourism and leisure agriculture, and off-farm employment through rural-to-urban labor transfer [78,79,80]. For the majority of poor households, improvements in C5: Education Level of Labor Force and C6: Skilled Labor Share enable them to achieve stable poverty alleviation through a combination of crop cultivation and nearby wage employment ($P^{S-1}$, $P^{S-5}$). Under the household contract responsibility system, land provides households with essential livelihood and income security. In the absence of a comprehensive rural social security system, land income—derived from either self-cultivation or land transfer—also serves as a substitute form of social protection [81,82]. For households with limited farmland resources, alternative livelihood pathways are often pursued, including cash crop cultivation (e.g., fruit trees), sideline agricultural activities, and rural tourism programs ($P^{S-2}$, $P^{S-3}$, $P^{S-4}$). However, these pathways generally require both industry assistance and improvements in labor skills as prerequisites. This resonates with the Philippines’ land reform experience, which demonstrated that land tenure security—rather than redistribution alone—was the critical factor shaping households’ productive use of land resources and ensuring the sustainability of poverty reduction [83,84]. This highlights that industry-assisted development is not only a growth pathway, but also a resilience-building mechanism that underpins long-term sustainability. In terms of stability, groups such as $S_{I-2}$ that rely primarily on wage employment can rapidly increase their income and achieve poverty alleviation, but they are simultaneously more vulnerable to external shocks (e.g., pandemics), revealing a tension between rapid exit and sustainable exit. Building resilience for these households requires income diversification (for example, local agro-processing during agricultural slack periods) and risk buffers tailored to flexible workers, such as emergency micro-credit and more inclusive social-insurance products.

(3)

From an overall perspective (Figure 18), the internal mechanisms through which rural households achieve stable poverty alleviation can be broadly categorized into two dimensions: “poverty alleviation” and “livelihood development.” First, poverty alleviation is guided by multidimensional poverty standards, which provide the basis for targeted interventions. Its primary focus is on reducing household burdens—particularly those related to healthcare, education, and dependency—and on improving basic living conditions such as safe drinking water and housing security. Second, livelihood development represents the core pathway for sustained income growth and long-term stability. This includes Social Protection Measures for households with special needs or disadvantaged environments, as well as Industry-Assisted Measures that support diversified pathways in agriculture, agro-processing, and rural services. Overall, while targeted poverty alleviation reduces immediate burdens, livelihood development determines households’ ability to sustain poverty reduction. The interaction of these two dimensions—protective transfers and endogenous development—constitutes the fundamental mechanism through which stable poverty alleviation evolves into sustainable, long-term resilience.

4.2. Spatial Heterogeneity and Underlying Mechanisms of Household Poverty Alleviation Pathways

Although previous studies have deepened our understanding of the spatial distribution and drivers of poverty, they have paid little attention to whether poverty alleviation pathways themselves exhibit spatial dependence and heterogeneity. In this study, spatial visualization of household poverty alleviation pathways reveals distinct patterns across regions. These patterns also demonstrate how local environmental and socioeconomic contexts shape households’ capacity to achieve not only poverty exit but also sustained poverty alleviation. Building on these findings, we explore how the characteristics of poverty alleviation pathways interact across mountainous areas, agricultural plains, and mining or peri-urban regions.

Mountainous areas, where natural resource development is constrained and infrastructure remains weak, have long been the focus of China’s poverty alleviation policies [85,86]. Due to limited regional endowments, some poor households ($S_{I-5}$, $S_{I-6}$) are included in the minimum living guarantee system even though they do not exhibit typical poverty traits such as severe labor shortages or heavy dependency burdens. This phenomenon reflects structural disadvantages commonly observed in mountainous regions, including scarce arable land, fragmented terrain, frequent natural disasters, and poor transportation access, which collectively hinder the development of sustainable livelihoods and stable employment opportunities. As a result, housing security and inadequate infrastructure have emerged as the primary manifestations of poverty in these areas [87,88]. In the poverty alleviation process, targeted policies have played a decisive role. On one hand, employment-oriented programs have facilitated labor transfer; on the other hand, public welfare positions, including ecological protection schemes and photovoltaic projects, have provided sustainable income sources for groups with weaker labor capacity. Together, these measures compensate for households’ disadvantages in the labor market and enable them to gradually achieve stable poverty alleviation. This highlights a policy-driven compensation mechanism that offsets the environmental constraints characteristic of mountainous regions. However, because such pathways rely heavily on external transfers rather than endogenous resource advantages, their long-term sustainability remains contingent on continuous policy input. Consequently, the strategic focus should shift from consumption subsidies to fostering endogenous industrial drivers. Developing localized value chains around specialty products (e.g., medicinal herbs, high-mountain tea) and ecotourism can provide sustainable livelihoods that leverage—rather than are limited by—the local environment.

Agricultural plains are generally characterized by abundant farmland and diverse livelihood opportunities, which have provided households with favorable conditions for development. Consequently, most households in these areas have successfully escaped poverty. However, the regional environment also exerts a selective effect. A subset of households ($S_{I-4}$) with limited labor supply or low labor efficiency, compounded by heavy structural burdens such as childcare responsibilities or disabled members, have struggled to achieve stable poverty alleviation. For these households, livelihood security relies primarily on social assistance and minimum living guarantee schemes ($P^{S-6}$). At the same time, the level of collective economic development shapes the pace of poverty alleviation. In some agricultural regions, township enterprises and village-level dividends have created a “shared dividends” mechanism that provides additional economic support to the most disadvantaged households [89,90]. This reflects a resource-advantaged but selectively differentiated pathway, in which institutional compensation mechanisms play a crucial role in sustaining vulnerable households. In the long run, the abundance of farmland and collective economic resources provides stronger conditions for sustainable poverty alleviation, yet the persistence of structurally disadvantaged groups also reveals the limits of resource advantages without inclusive institutional arrangements. Hence, plains regions should emphasize inclusion. Village collective economic organizations can use their revenues to pay medical insurance premiums for $S_{I-4}$ households, provide universal childcare services, or offer land trusteeship for households with severe disability—turning regional resource advantages into precise support for the most vulnerable.

Households in mining and peri-urban areas ($S_{I-2}$), whose livelihoods rely heavily on migrant labor, are particularly vulnerable to fluctuations in the labor market. Their income is highly sensitive to external shocks such as economic downturns or the COVID-19 pandemic, which can trigger both a reduction in employment opportunities and a decline in wages. These dual pressures often lead to fewer working hours or outright unemployment, eventually pushing households back into poverty. Compared with households in agricultural or mountainous areas, this group can achieve income growth more quickly in the short term, but their earnings are far more volatile. Although they lack a stable agricultural resource base, they do possess certain labor advantages. As a result, their pathway to stable poverty alleviation typically requires intermediate stages of vocational training and targeted policy support ($P^{U-2}$), which ultimately enable them to secure steady employment or sustainable livelihoods. This reflects a labor-based but market-vulnerable pathway, in which policy interventions in skill training and employment support are essential for sustaining poverty alleviation. Because households in these regions lack secure agricultural assets, their sustainable poverty reduction depends critically on stabilizing non-farm employment and building resilience against labor market fluctuations. Priority should be given to strengthening labor-market resilience. This can be achieved by developing localized job-matching platforms to reduce search costs, providing advanced training for shock-resistant sectors (e.g., logistics, equipment repair), and improving the portability of social benefits for migrant workers to facilitate smoother labor mobility.

The comparative analysis of mountainous, agricultural, and mining or peri-urban areas reveals that, across different spatial contexts, household poverty alleviation pathways are shaped by three distinctive mechanisms: policy compensation for environmental constraints, institutional support to offset resource-based differentiation, and policy-assisted stabilization of labor-dependent but market-vulnerable livelihoods. Together, these mechanisms underscore the inherent spatial heterogeneity of poverty alleviation processes, driven by the interplay of environmental conditions, resource endowments, and labor market dynamics. Achieving sustainable poverty alleviation therefore hinges on accurately matching household livelihood demands with local industrial and employment opportunities. Future policy should move beyond generalized strategies toward region-specific approaches that integrate spatial heterogeneity with household characteristics. In particular, mountainous regions require sustained policy transfers, agricultural plains demand inclusive use of collective resources, and mining/peri-urban areas need labor-market stabilization measures—together illustrating how spatial characteristics shape the sustainability of household poverty alleviation pathways.

4.3. Research Limitations and Future Directions

While this study demonstrates the utility of explainable machine learning in mapping household poverty pathways, several limitations should be acknowledged. First, the findings are based on household-level data from a single prefecture in central China. Although the dataset is substantial, its regional specificity limits external validity and the direct generalizability of the results to other socioeconomic or policy contexts. Caution is therefore warranted when extrapolating the identified pathways and mechanisms to regions with markedly different geographic conditions, resource endowments, or institutional settings.

Future research could extend this work in several promising directions. To enhance external validity, subsequent studies should expand the sample to cover diverse geographic environments, resource bases, and developmental conditions. This would help build a more representative knowledge system of poverty alleviation and development. Furthermore, while the unsupervised clustering approach effectively identified broad typologies of household pathways, its capacity for fine-grained analysis of complex feature interactions remains limited. To address this, future work could incorporate more advanced techniques, such as graph representation learning, to better model the intricate relationships between household characteristics and their environments.

Beyond methodological refinements, a critical next step involves deepening the spatial and dynamic dimensions of the analysis. This study primarily relied on visualization and descriptive analysis to explore spatial patterns. A more rigorous approach would be to formally integrate multi-source spatial data—such as topography, infrastructure, and resource accessibility—into the modeling framework to construct a spatially-aware, context-responsive analysis. Finally, as dynamic data become increasingly available, future research should also explore predictive and simulation modeling to better understand poverty relapse risks and forecast the long-term evolution of household pathways under different policy scenarios.

5. Conclusions

Understanding the dynamic mechanisms of household poverty alleviation and development provides theoretical support for achieving stable and sustainable poverty reduction and for preventing poverty relapse. These pathways are shaped not only by household characteristics but also by the joint influences of natural–social environments and policy interventions. Effectively uncovering these processes is therefore of great practical significance and theoretical value for promoting resilient and sustainable rural development.

To address this, the study develops an explainable machine learning framework that integrates multidimensional poverty theory with a dynamic poverty perspective. By combining XGBoost and SHAP, the framework identifies and interprets household poverty alleviation and development pathways, using monitoring data from 107,637 households in Pingdingshan City over a 12-year period for empirical validation.

The main contributions of this study are as follows:

• Introducing the concept of household poverty alleviation and development pathways, which integrates both “poverty states” and ‘process dynamics’. This conceptualization addresses the temporal fragmentation problem in existing studies that analyze either states or processes in isolation, and enables a more comprehensive understanding of the dynamic evolution of poverty alleviation and development;
• Constructing an explainable machine learning framework that combines XGBoost and SHAP methods to identify the evolutionary features of household “state–process” transitions, providing a methodological foundation for analyzing the mechanisms of poverty alleviation and development;
• Empirically validating the framework with large-scale household monitoring data to reveal fine-grained differences in state transitions and pathway choices;
• Analyzing the role of policy measures during China’s targeted poverty alleviation phase, thereby linking empirical findings with theoretical insights.
• Building on this foundation, the analysis yields several key findings regarding pathway dynamics:

First, household pathways are strongly predicted by their initial typology. Our results reveal distinct trajectory patterns: households initially characterized by heavy burdens and limited labor capacity ($S_{I-4}$) predominantly remained in unstable states (62.5%), relying on social protection ($P^{S-6}$). In contrast, those with relative human capital advantages ($S_{I-3}$, $S_{I-6}$) achieved stable poverty alleviation directly at rates of 84.9% and 100%, respectively. Households dependent on migrant labor ($S_{I-2}$) were highly vulnerable to market shocks, with 33.8% following unstable pathways ($P^{U-2}$).

Second, the transition from instability to stability follows discernible bridging mechanisms. Pathways reliant on skill upgrading (e.g., from $P^{U-4}$ to $P^{S-5}$/$P^{S-1}$) proved more decisive in securing long-term stability than those dependent solely on short-term subsidies. This underscores that enhancing human capital and employability is critical for converting fragile transitions into resilient outcomes.

Third, pathways are intrinsically shaped by spatial context, creating a geography of opportunity and risk. In mountainous areas, stability was achieved primarily through policy compensation mechanisms ($P^{S-7}$). Agricultural plains, benefiting from better resources, supported more diversified pathways but required inclusive institutional arrangements to support vulnerable groups. Mining and peri-urban regions exhibited high sensitivity to labor market fluctuations, necessitating targeted stabilization policies.

In summary, this study demonstrates that sustainable poverty alleviation hinges on aligning interventions with both initial household profiles and regional contexts. The central policy implication is that geographic disadvantages and development opportunities coexist. Effective strategies must therefore not only address immediate constraints but also actively transform local conditions, balancing safety nets with capacity-building to create equitable development pathways across diverse regions.