Practice Primacy: Revisiting the Knowledge–Action Gap in Pro-Environmental Behavior with eXplainable AI

Abstract

Against the backdrop of an escalating global environmental crisis, bridging the “knowledge–action gap” in the pro-environmental behavior (PEB) of university students has become a key challenge for sustainable development education, aligning with SDG 12 (Responsible Consumption and Production) and SDG 13 (Climate Action). Traditional linear models often struggle to capture the complex non-linearities and interaction effects when explaining this gap. To overcome this limitation, this study introduces an integrated “prediction-plus-explanation” framework using eXplainable Artificial Intelligence (XAI). Based on survey data from 463 university students in China, we constructed a high-precision PEB prediction model (Accuracy = 93.55%) using the CatBoost algorithm and conducted an in-depth analysis of its internal decision-making mechanisms with the SHAP (SHapley Additive exPlanations) framework. The results reveal that a “Practice Primacy” model plays a dominant role in driving PEB: the formation of environmental habits, participation in environmental practices, and the investment of related resources are the overwhelmingly dominant factors in predicting individual behavior, with their cumulative contribution far exceeding that of traditional cognitive and attitudinal variables. Furthermore, heterogeneity analysis revealed significant group differences in these driving mechanisms: the behavioral decisions of male students tend to be more “value-driven,” while lower-division students are more susceptible to external educational interventions. By quantifying the non-linear effects and relative importance of each driver, this study offers a new “Action-to-Cognition” perspective for bridging the knowledge–action gap and provides robust, data-driven support for universities to design precise and differentiated intervention strategies, thus contributing to the achievement of SDGs.

1. Introduction

Amidst the escalating global environmental crisis, promoting pro-environmental behavior (PEB) among the public has become a central pathway to addressing climate change, protecting biodiversity, and achieving sustainable development [1,2,3]. However, a persistent challenge in the fields of environmental education and behavioral science is the well-documented “knowledge–action gap,” which describes the discrepancy where enhanced environmental knowledge and positive attitudes do not consistently or directly translate into tangible environmental actions [4]. This gap is a critical bottleneck that undermines the effectiveness of many environmental policies and educational programs [5,6,7]. This is particularly relevant in the context of the Sustainable Development Goals (SDGs), which are integral to guiding global efforts toward a more sustainable future. In particular, SDG 12 (Responsible Consumption and Production) and SDG 13 (Climate Action) underscore the need for individuals to make sustainable choices and adopt behaviors that mitigate environmental impact. Among various social groups, university students represent an ideal sample for investigating this phenomenon. As a highly educated cohort whose values and behavioral habits are in a crucial formative stage, they are poised to become future decision-makers, innovators, and primary consumers [8,9]. If a significant knowledge–action gap exists even within this demographic, uncovering its underlying drivers holds substantial theoretical and practical importance for the broader societal transition toward sustainability. Therefore, this study focuses on university students in China to deeply investigate the key factors influencing their pro-environmental behavior.

To explain pro-environmental behavior, academia has developed established theoretical frameworks, with the Theory of Planned Behavior (TPB) and the Value-Belief-Norm (VBN) Theory being among the most influential [8,9]. These theories provide critical analytical dimensions for understanding behavioral intentions, confirming that environmental cognition (knowledge and beliefs) [10], environmental attitude [11], subjective norms, and personal norms [12] are significant antecedents. Nevertheless, these traditional models, which are often based on linear assumptions, have shown limitations in explaining the conversion from intention to actual behavior and possess relatively modest predictive power [13]. Research indicates that behavioral decision-making is a complex psychological process, shaped by a multitude of factors that often interact in non-linear and interdependent ways [14]. For instance, the impact of a specific piece of knowledge may depend on an individual’s emotional state, while the execution of an environmental intention can be constrained by contextual factors such as convenience and cost [15]. Traditional statistical models struggle to fully capture these complex dependencies, which limits our ability to identify the “behavioral levers” that are genuinely decisive in real-world settings.

To overcome the limitations of traditional methods, this study advocates for a paradigm shift from “theory-testing” to “data-driven prediction,” introducing eXplainable Artificial Intelligence (XAI) as a core analytical tool [16,17]. Unlike conventional models that rely on strict a priori assumptions, machine learning algorithms—particularly tree-based models like CatBoost—possess a powerful capacity to automatically learn complex non-linear relationships and higher-order interaction effects from high-dimensional data, with the primary goal of maximizing predictive accuracy [18]. In recent years, the maturation of interpretability techniques, exemplified by SHAP (SHapley Additive exPlanations), has successfully opened the “black box” of machine learning [18]. This allows researchers not only to build high-precision predictive models but also to gain deep insights into the model’s decision-making mechanisms by clearly quantifying the marginal contribution of each predictor to individual outcomes. This “prediction-plus-explanation” combination provides an unprecedentedly powerful toolkit for identifying and ranking the key drivers of pro-environmental behavior [14,18].

This study does not aim to test or revise any specific theory. Instead, it draws upon the key dimensions of influence identified by classic theories such as the TPB and VBN to construct a comprehensive analytical framework, which serves as the input feature set for our machine learning model. This framework systematically incorporates potential predictors across multiple dimensions, including demographic characteristics, environmental cognition, environmental attitudes and willingness, and past practical experiences, to comprehensively examine their collective impact on actual pro-environmental behavior. Specifically, this study will employ the high-performance CatBoost, a gradient boosting decision tree model, for predictive modeling and integrate it with the SHAP framework for in-depth interpretation. The research objectives are as follows: (1) to identify the key variables that most accurately predict the level of pro-environmental behavior among university students; (2) to quantify the relative importance of these key predictors and reveal their patterns of influence on behavior; and (3) to explore whether group heterogeneity exists in these influencing factors based on demographic characteristics. Through this research, we hope to offer a new methodological perspective for tackling the knowledge–action gap and to provide solid empirical evidence for designing more targeted and effective environmental education intervention strategies for universities and policymakers.

2. Background and Related Work

In recent years, an increasing body of research has emphasized the integration of Sustainable Development Goals (SDGs) in education, particularly in higher education [19,20,21]. Studies have shown that incorporating SDGs into educational curricula can significantly enhance students’ environmental awareness and promote pro-environmental behaviors, such as recycling and energy conservation, which are crucial for advancing SDG 12 (Responsible Consumption and Production) [22]. The integration of SDGs in educational practices encourages students to adopt behaviors that contribute to climate action, aligning with SDG 13 (Climate Action) [23,24]. The role of social identity in pro-environmental behavior varies across cultures, and fostering environmental responsibility through education contributes to the realization of SDG 11 (Sustainable Cities and Communities) [25]. Therefore, these studies suggest that education focused on SDGs, and promoting sustainable behaviors is essential for achieving SDG targets. Despite these findings highlighting the close relationship between environmental education and behavioral change, a critical issue remains—the knowledge–action gap, which refers to how environmental knowledge is translated into actual pro-environmental behavior.

2.1. The Knowledge–Action Gap: An Enduring Puzzle

A central and persistent challenge in the field of environmental behavior research is the widely recognized “knowledge–action gap,” also referred to as the “attitude-behavior gap” [26,27,28]. This phenomenon describes the discrepancy where an individual’s environmental knowledge and positive environmental attitudes often fail to translate effectively into stable and consistent pro-environmental actions [27]. A substantial body of research indicates that while environmental education can significantly enhance public environmental literacy and concern, this cognitive and affective shift does not share a simple linear relationship with improvements in actual behavior [18]. This gap is not only a frontier theoretical issue in environmental psychology but also a critical real-world bottleneck that diminishes the effectiveness of many environmental policies, campaigns, and educational programs (

Table 1. Literature Review on the “Knowledge–Action Gap”.

Authors (Year)	Theory	Method	Sample	Key Findings
Frederiks, E.R., Stenner, K., & Hobman, E.V. (2015) [29]	Behavioral economics, psychology	Review of household energy consumption behaviors	The specific sample size is not mentioned, the study draws from various behavioral economics concepts related to household energy use.	Consumers’ behaviors often contradict their knowledge and intentions due to cognitive biases and motivational factors. Despite awareness, household energy conservation is difficult due to complex psychological barriers, not just financial incentives.
Knutti, R. (2019) [30]	None specifically mentioned, focusing on the gap in climate change	Commentary	Not applicable (review study)	Despite the rapid increase in climate change knowledge, global emissions continue to rise due to inadequate policy implementation. The gap between knowledge and action in climate change is exacerbated by political and economic inertia.
Casiás, B., & Faria, J. (2021) [31]	Intention-Behavior (I-B) Gap in Ethical Consumption	Mediation and moderation analysis using survey data	364 respondents	Ethical behavior is mediated by plans and habits and moderated by commitment and sacrifice. The paper presents four consumer profiles based on their ethical priorities and suggests interventions to close the I-B gap by promoting ethical consumption as a social norm.
Meyer, K.B., & Simons, J. (2021) [32]	Food Choice Process Model	Ethnographic fieldwork	Six families in Germany	Despite positive attitudes towards sustainable food, actual behavior was influenced by household realities, conflicting personal and situational factors, and the complexity of decision-making processes.
Tawde, S., Kamath, R., & Husain, S.R.V. (2022) [33]	TPB, Self-efficacy Theory	Survey-based empirical research using moderated mediation analysis	674 green consumers	When consumers are provided with cues to implement intentions and have higher self-efficacy, their intentions are more likely to translate into green behavior.
Essiz, O., Yurteri, S., Mandrik, C., & Senyuz, A. (2023) [34]	Cognitive psychology, Green value-action gap	Moderated moderation model using survey data	328 respondents	Risk aversion and subjective knowledge are important moderators of the green value-action gap. The study also finds gender differences, with women showing greater consistency between their green values and actions compared to men.
Mooney, M.E. et al. (2022) [35]	None specified, focusing on educational interventions	Survey on climate knowledge and behavior change among university students	71 undergraduate students from the University of Wisconsin–Madison	Climate education led to significant behavior changes in students regarding carbon footprints. A positive attitude from education could lead to more sustainable behaviors.
Wut, T.M., Lee, D., & Lee, S.W. (2023) [36]	Theory of planned behavior	Literature review and research agenda	Not applicable (review study)	In sustainable tourism, there is a significant gap between attitude and behavior due to practical issues like convenience, time constraints, and moral values conflicting with sustainability intentions.
Colombo, S.L., et al. (2023) [37]	Comprehensive-Action-Determination Model	Literature review on environmental knowledge–action gap	Not applicable (review study)	Environmental knowledge alone does not lead to action. Psychological and cognitive factors such as self-regulation, executive functions, and emotional responses are crucial in bridging the knowledge–action gap.
Räsaänen, A., et al. (2024) [28]	Knowledge co-production theory	Case study on transdisciplinary project in land use management	Various researchers, local stakeholders, and administrative actors involved in a project on sustainable land use in the Kiiminkijoki river catchment, northern Finland.	Knowledge–action gaps persist not just because of incomplete knowledge but due to the complex, iterative, and coevolutionary nature of knowledge production. Bridging this gap requires ongoing interaction and co-production between diverse stakeholders.
Sinha, R., & Annamdevula, S. (2024) [38]	Knowledge-Attitude-Behavior Theory	Parallel and serial mediation effects using PROCESS macro (Models 4 and 6)	395 youth from three different cities in India, collected via purposive sampling	The study reveals that environmental concern, green perceived value, and green attitude act as both parallel and sequential mediators between environmental knowledge and green purchase intentions. The direct impact of environmental knowledge on green purchase intentions was found to be insignificant. The model emphasizes that environmental knowledge, when combined with environmental concern and green perceived value, plays a significant role in shaping attitudes that foster green purchase intentions.
Recio-Román, A., et al. (2024) [27]	Knowledge-Attitude-Behavior model, Attitude-Behavior-Context theory	Mediation path analysis using Eurobarometer data	26,630 EU citizens	The gap between positive environmental attitudes and actual behavior is mediated by factors like ecolabel knowledge, trust, and environmental concern. Despite a favorable attitude, actual ecolabel adoption remains low.

).

Consequently, a deeper understanding of this gap and the exploration of pathways to bridge it are of paramount importance [15]. The academic community has explained the origins of the knowledge–action gap from multiple perspectives, with its causes often interacting in complex and interdependent ways to impede the conversion from intention to action. These factors can be systematically categorized into the following dimensions:

Internal Psychological Barriers: These barriers originate from an individual’s internal cognitive and affective processes. Common psychological obstacles include cognitive dissonance, where individuals selectively ignore environmental information to alleviate the discomfort arising from inconsistencies between their beliefs and actions; optimism bias, the belief that environmental problems pose a greater threat to others than to oneself; the powerful inertia of habit, where automated daily routines are difficult to alter through conscious environmental intentions; and limited cognitive and emotional resources, leading to feelings of helplessness or apathy when faced with vast amounts of environmental information [14,39,40]. Furthermore, environmental intentions are often suppressed when a goal conflict arises between environmental objectives and other personal goals, such as the pursuit of convenience, comfort, or economic benefits.

External Contextual Barriers: These barriers are constraints imposed by an individual’s physical and social environment. Structural and physical obstacles are among the most direct factors, including high economic costs (e.g., the price premium for green products), limited time, a lack of convenience (e.g., insufficient or distant recycling facilities), and inadequate infrastructure [41,42]. The information environment is also crucial, as information overload, conflicting sources, or uncertainty can lead to decision paralysis or delayed action [43].

Socio-cultural Barriers: Individual behavior is deeply embedded within social networks and cultural norms. Social norms play a pivotal role; when individuals perceive that their peers, family, or community members are not widely adopting pro-environmental behaviors, their own willingness to act is significantly diminished due to conformity [44,45]. Moreover, a lack of social trust in the environmental commitments of governments, corporations, or environmental organizations can also weaken public motivation to participate [46,47]. Cultural values, such as the differences between individualism and collectivism, can similarly moderate the internalization of environmental norms and subsequent behavioral performance [48].

2.2. Theoretical Frameworks and Key Predictors of Pro-Environmental Behavior

2.2.1. Contributions from Classic Theories: Identifying Core Dimensions of Influence

To systematically investigate the multifaceted factors influencing PEB, this study draws upon two of the most influential classic theoretical frameworks in environmental psychology and behavioral science to guide the subsequent selection of predictors.

The Theory of Planned Behavior (TPB) is one of the most widely applied theories for explaining and predicting human social behavior [49,50]. The theory posits that an individual’s actual behavior is primarily determined by their behavioral intention. This intention, in turn, is shaped by three core psychological constructs: (1) Attitude toward the Behavior, an individual’s positive or negative evaluation of performing the behavior; (2) Subjective Norms, the perceived social pressure from significant others or groups to perform or not perform the behavior; and (3) Perceived Behavioral Control (PBC), an individual’s perception of the ease or difficulty of performing the behavior, which reflects both their belief in their own capabilities (self-efficacy) and their assessment of external resources and opportunities [10,11,51]. The TPB provides a foundational analytical framework for understanding the personal, social, and perceived situational factors in individual decision-making.

The Value-Belief-Norm (VBN) Theory is a theoretical model developed specifically to explain pro-environmental behavior [52,53]. The VBN theory constructs a causal chain from macro-level values to specific behaviors, emphasizing the driving role of intrinsic moral factors. According to this theory, an individual’s PEB originates from their activated personal norms—a sense of moral obligation to take pro-environmental action. The activation of these personal norms is, in turn, influenced by a series of antecedents, including: (1) an individual’s core values (e.g., altruism, ecocentrism); (2) general Beliefs about environmental issues, encompassing one’s environmental worldview and cognition of the consequences of environmental degradation; and (3) the perception of being able to alleviate the threat through personal action (awareness of consequences and ascription of responsibility) [12]. The VBN theory deepens our understanding of the cognitive, value-based, and moral motivations that drive pro-environmental behavior.

Positioning and Adaptation in This Study: Although the TPB and VBN have made outstanding contributions to explaining behavioral intentions [54], the purpose of this study is not to test or compare the structural validity of these theoretical models using methods such as Structural Equation Modeling (SEM) [55,56,57]. Instead, we regard these widely validated theories as a rich “theoretical toolbox.” We draw upon the core dimensions of influence that these theories have repeatedly identified and confirmed—namely, environmental cognition, environmental attitudes, personal norms and responsibility, and perceived behavioral control—to guide our selection of a comprehensive and theoretically grounded set of candidate predictors for our subsequent predictive models. In this way, our research is both rooted in established theoretical traditions and capable of exploring the relative importance and complex patterns of these factors in predicting actual behavior through a data-driven approach.

2.2.2. An Integrated Framework and Predictors for This Study

Guided by the classic theories discussed above, this study constructs a multidimensional, integrated framework that systematically incorporates potential predictors of university students’ pro-environmental behavior (PEB). These factors collectively form the input feature set for our machine learning model, aiming to uncover key behavior-driving patterns from the data. This framework comprises the following four core dimensions:

Environmental Cognition. This dimension corresponds to the “Environmental Beliefs” construct in VBN theory and constitutes the rational foundation upon which individuals understand and form judgments about environmental issues. In this study, environmental cognition is operationalized through two key variables: (1) Environmental Knowledge, which encompasses an individual’s basic understanding of environmental facts, causes, consequences, and relevant laws and regulations [58]. A solid knowledge base is a necessary precondition for forming accurate risk perceptions and positive environmental attitudes [59]. (2) Risk Perception, which refers to an individual’s subjective assessment of the severity of environmental problems and their relevance to personal life [59,60]. Substantial research has shown that when individuals perceive a higher level of environmental risk, their motivation to take protective actions is significantly enhanced.

Environmental Attitude and Willingness. This dimension represents the affective and motivational core of an individual’s environmental decision-making, corresponding primarily to “Attitude” in the TPB and “Personal Norms” in the VBN theory. It is measured through the following variables: (1) Willingness to Participate, defined as an individual’s general evaluative tendency toward participating in various environmental protection activities. This directly reflects the core of attitude in the TPB and serves as a foundation for forming specific behavioral intentions. (2) Ascribed Responsibility, which measures whether an individual attributes the cause of environmental problems to human activities and consequently perceives a personal or collective moral obligation [61]. According to VBN theory, this activated sense of responsibility is a key intrinsic motivation driving pro-social and pro-environmental behaviors [61].

Perceived Behavioral Control and Past Experience. This dimension primarily corresponds to “Perceived Behavioral Control” (PBC) in the TPB, which is an individual’s perception of the ease or difficulty of performing a behavior [62]. Direct measurement of PBC can be complex across different contexts, and research suggests that past behavioral experience is one of the best proxy variables for reflecting and shaping an individual’s PBC [62]. Therefore, this study indirectly measures this construct by examining students’ past practical experiences, including whether they have systematically participated in environmental education courses. Individuals with such prior experiences typically possess higher self-efficacy, have acquired necessary skills and information, and consequently perceive fewer barriers to action, making them more likely to translate environmental intentions into actual behavior. By including these variables as predictors, we aim to explore whether past “small actions” can effectively predict future “larger actions” [63].

Demographic and Contextual Factors. Individual decision-making does not occur in a vacuum but is deeply embedded in social and contextual settings. Consequently, this study incorporates a series of demographic and contextual variables as important control and moderating factors. These variables include gender, academic year, family background (urban/rural), and information acquisition channels, among others [25,64]. Although these variables are not typically considered direct psychological drivers, they can indirectly shape PEB by influencing an individual’s knowledge acquisition, attitude formation, access to opportunities, and resource constraints. Including these variables in the model not only helps us control for potential confounding effects but also provides the foundation for subsequent exploration of heterogeneity in the drivers of PEB across different groups.

By constructing this integrated framework, which encompasses cognitive, attitudinal, control, and contextual factors, this study aims to move beyond the limitations of any single theory. It seeks to systematically examine how different types of predictors jointly operate to more comprehensively and accurately predict the pro-environmental behavior of university students.

3. Methods

The research process was divided into three stages: study design and data collection, variable selection and measurement, and model selection and data analysis. To provide a clearer overview of the methodological framework, Figure 1 presents a schematic summary of the entire research process.

3.1. Study Design, Participants, and Data Collection

This study employed a questionnaire survey method to systematically identify the factors influencing the PEB of university students. The questionnaire was designed based on a review of existing research and was adapted to the context of Chinese universities. It consisted of four sections: the first section collected basic demographic information and external contextual variables, such as gender, academic year, family background, participation in environmental courses, information acquisition channels, and preferences for pathways to enhance environmental awareness. The second section measured cognitive variables, including basic environmental knowledge, legal awareness, and relational cognition. The third section assessed attitudinal and value-oriented variables, such as value-based cognition, sense of responsibility and crisis, and willingness to participate. The final section measured the specific practices of students’ PEB, such as active engagement time, environmental habit formation, and resource and effort investment.

The data for this study were collected from a large-scale questionnaire survey conducted between February and March 2025 at six universities in Zhejiang Province, China. These institutions included comprehensive, normal, and science and engineering universities, making the findings applicable to student populations with similar institutional backgrounds. To ensure sample representativeness, a stratified sampling strategy was employed, with sample allocation based on key dimensions such as university type, gender, and academic year. The survey was administered anonymously online via the “Wenjuanxing” platform. To control data quality, we implemented several technical measures, including IP address restrictions, response time thresholds, and logical consistency checks. After excluding invalid responses due to logical conflicts, excessively short completion times, or duplicate entries, a final sample of 463 valid responses was obtained.

During the questionnaire administration phase, the research team implemented multiple quality control measures to ensure the scientific rigor and reliability of the data. These measures included setting “trap” questions to screen for inattentive responses, preventing duplicate submissions from the same device, establishing a minimum response time to filter out random answers, and using a standardized distribution portal and instructions to minimize operational errors and information bias. Furthermore, this study strictly adhered to ethical standards for research in the humanities and social sciences. All participation was based on informed consent, with the research purpose, scope of data use, and confidentiality principles clearly stated on the first page of the questionnaire. Respondents participated voluntarily and anonymously, and the data were used exclusively for academic research, ensuring full ethical compliance.

To assess the measurement quality of the survey instrument, reliability and validity analyses were conducted on the collected data using SPSS 27.0. For reliability, the overall Cronbach’s alpha for the questionnaire was 0.781, indicating good internal consistency. For validity, the Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy was 0.864, and Bartlett’s test of sphericity was significant (p < 0.001), suggesting that the sample data were suitable for factor analysis. However, the scale for Environmental Cognition Evaluation (EC2) demonstrated low reliability within the university student sample (α = 0.313). This may be related to the heterogeneity in students’ cognitive levels and interest in environmental issues. Specifically, some items in this construct might involve higher-order abstract cognition, and the questionnaire design may not have accurately captured the true cognitive responses of this demographic. Future research could therefore consider redesigning the items for this construct to more clearly distinguish its sub-dimensions or employ alternative measurement approaches to enhance its reliability and validity. Additionally, future work could explore extending this scale to different populations or cultural contexts to verify its adaptability and stability.

3.2. Variable Measurement and Feature Engineering

To comprehensively investigate the factors influencing individual pro-environmental behavior, this study selected 15 independent variables and one dependent variable to construct the analytical model, based on the literature review, theoretical logic, and the content of the survey questionnaire. The selection of variables was guided by three core principles: first, theoretical support, requiring that the variables have been repeatedly validated in existing research on environmental behavior or behavioral science; second, data availability, ensuring that the variables had high response rates and acceptable reliability and validity in our survey; and third, variable representativeness, aiming to cover multiple dimensions including individual cognition, motivation, behavior, resources, and background.

The dependent variable in this study is Pro-Environmental Behavior (PEB). This variable was assessed by measuring the frequency and willingness of individuals to take environmental actions in their daily lives. The original data were scores on a five-point Likert scale. To adapt this variable for the prediction task of our classification models, we dichotomized it: individuals with scores greater than or equal to the sample median were classified into the “High-PEB group” (coded as 1), while those with scores below the median were assigned to the “Low-PEB group” (coded as 0).

Furthermore, to systematically present the composition and measurement of the independent variables, this study categorized the explanatory variables based on the psychological mechanisms and behavioral dimensions they reflect. They were grouped into four main categories: (1) individual background characteristics, (2) environmental cognition evaluation, (3) responsibility and willingness, and (4) practice and capacity investment. Each category corresponds to a key node in the theoretical pathways and was measured using standardized items from the questionnaire. The specific variable definitions and coding methods are detailed in

Table 2. Assessment of Questionnaire Reliability and Factorability.

Statistic	Environment Cognition Evaluation	Attitudes and Values	Practice and Ability Investment	Overall
Cronbach’s α	0.313	0.753	0.744	0.781
KMO	0.630	0.835	0.864	0.864
Bartlett	p < 0.001	p < 0.001	p < 0.001	p < 0.001
Number of Items	9	15	10	34

.

As a demographic group in a critical period for value shaping and habit formation, university students exhibit distinct characteristics in their cognitive abilities, emotional responses, and acceptance of social norms. Students with high value-based cognition and a strong willingness to participate often demonstrate more active engagement in actual PEB. Therefore, systematically incorporating the aforementioned variables into our modeling framework is not only supported by a solid theoretical foundation but also aligns with the psychological and behavioral traits of the university student population (

Table 3. Selection, Definition, and Measurement of Independent Variables.

Variable Dimension	Variable Name	Variable Definition	Specific Features
Individual Background Characteristics	Gender		Male = 1, Female = 0
	Grade		Freshman = 1, Sophomore = 2, Junior = 3, Senior = 4, Fifth-Year Undergraduate = 5, Postgraduate = 6.
	Home Location (HL)	Whether the respondent is from city or countryside.
Environment Cognition Evaluation (EC)	Basic Knowledge (EC1)	The factual understanding of the current environmental situation, risks, and causality [65].	Exposure to nature is beneficial for both physical and mental well-being.
			In the past 12 months, has your place of residence been affected by climate-related issues?
	Perceived correlation (EC2)	The strength of the perceived correlation between environmental issues and the utilization of oneself, society, and resources [66].	Environmentally friendly consumption can only minimize environmental damage to a limited extent. (Reverse-coded item)
			Organic products are healthier and taste better.
			Only by cherishing and utilizing resources efficiently and responsibly can we protect nature.
	Value-based Cognition (EC3)	Rational judgment on the trade-off between nature conservation, economic development, and individual costs [65].	You enjoy outdoor recreational activities such as hiking, birdwatching, swimming, and skiing.
			Protecting biodiversity requires sacrificing industrial agricultural production.
			Infrastructure development is more important than environmental protection.
			Industrial growth is more important than environmental protection.
	Awareness of Relevant Laws (EC4)	Whether the respondent is familiar with laws related to environmental protection.	Are you familiar with the relevant laws regarding environmental protection?
	Participation in Environmental Courses (EC5)	Whether the respondent has previously taken courses related to environmental protection.	Have you participated in any courses related to the environment?
Attitudes and Values (AV)	Subjective Evaluation (AV1)	Overall self-assessment of the adequacy/effectiveness of environmental education and personal changes [7].	Environmental education has an impact on changing environmental attitudes and behaviors.
			Schools should provide environmental education.
			Environmental education in schools is currently well implemented.
			The relationship between humans and nature should not be one of harmonious coexistence. (Reverse-coded item)
			Just as humans have the right to exist, plants and animals, as living beings on Earth, also have their own rights.
	Responsibility and Crisis Awareness (AV2)	Depth of understanding of the severity of environmental problems and their impacts on humanity. Views on the environmental impacts caused by human activities [66,67].	The rapid depletion of natural resources will not threaten the future of humanity. (Reverse-coded item)
			Climate change and environmental pollution jeopardize living conditions, food security, and animal health.
			The Earth’s ecosystems are highly fragile.
			Humans are severely polluting the environment.
			The Earth’s population growth has reached the limit of its carrying capacity.
	External Impact (AV3)	Belief that environmentally unfriendly behaviors are mainly attributable to the lack of environmental education programs in schools. Views on the extent to which economic growth models, innovation, and technological progress can improve environmental outcomes, as well as on the self-regulating capacity of ecological systems [67].	Environmental harm is primarily due to the lack of environmental education in schools.
			If the current economic growth model continues unchanged, humanity may face an environmental disaster.
			Human innovation and technological progress can improve the environment.
			Ecological balance has self-regulating and restorative capabilities to cope with the impacts of human-induced pollution.
			Regardless of human special abilities, we are still subject to the laws of nature.
	Intention to Join Environmental Organizations (AV4)	Whether the respondent is willing to join environmental organizations	Would you be willing to join an environmental protection organization and actively participate in environmental activities?
Practice and Ability Investment (PA)	Active Engagement in Practice (PA1)	Participation in organizations and activities (related to environmental protection). Practical involvement in environmental protection activities [66].	Are you willing to lower your standard of living to protect the environment?
			Do you watch documentaries on environmental issues?
			Do you discuss the importance of the environment with others?
	Environmental Habit Formation (PA2)	Frequency of everyday pro-environmental behaviors. Pro-environmental consumption choices [66,68].	Do you prefer choosing single-use bottled beverages?
			Before purchasing products, do you read labels to check if they are environmentally friendly?
			Out of concern for environmental protection, do you continuously switch to eco-friendly products?
			How often do you engage in environmental behaviors in your daily life, such as reusing items, saving water and electricity, and using eco-friendly transportation?
	Resource and Effort Investment (PA3)	Investment of financial resources. Investment of time and effort [67].	To reduce environmental issues, do you actively engage in self-learning?
			Do you actively participate in recycling activities organized by your school?
			If you discover environmental crimes or invasive species, do you actively report them to the relevant government departments?
Pro-Environmental Behavior (PEB)	Energy Conservation	Energy-saving behaviors adopted by individuals in daily life, such as reducing energy consumption and using energy-efficient devices.	You avoid using air conditioners or heaters during peak hours and regularly check household appliances for energy wastage.
	Resource Consumption Reduction	Conservation behaviors exhibited by individuals during resource consumption, such as avoiding waste and promoting reuse.	You try to avoid using single-use plastic products (such as disposable cups, utensils, etc.) and minimize food waste by purchasing according to need.
	Eco-friendly Consumption	Individuals prioritize the environmental characteristics and sustainability of products when purchasing and consuming goods.	When purchasing and consuming goods, you prioritize the environmental characteristics and sustainability of the products.

). These variables have not only been widely confirmed in prior research to possess strong theoretical explanatory power and behavioral predictive capabilities but are also highly compatible with the machine learning methods used in this study, featuring clear quantification standards and practical measurability. By systematically modeling these internal factors and ranking their importance, this study aims to more deeply uncover the psychological mechanisms underlying students’ PEB and provide micro-level empirical support for bridging the “knowledge–action gap.”

3.3. Analytical Strategy

3.3.1. Model Comparison and Selection

Given that PEB is a complex behavior driven by multiple factors, its underlying mechanisms may involve intricate non-linearities and interaction effects that fall beyond the explanatory scope of traditional linear models. To enhance the accuracy and robustness of behavioral prediction, this study systematically introduces eight mainstream supervised classification algorithms for comparative analysis. These algorithms span both traditional and ensemble learning frameworks and include: Support Vector Machine (SVM) [69], Neural Network (NN), Decision Tree, Random Forest (RF) [70,71], K-Nearest Neighbors (KNN), LightGBM, XGBoost [72], and CatBoost [73]. To ensure the adequacy of the sample size for the machine learning methods used in this study, a post hoc power analysis was conducted using G*Power 3.1. Given the total sample size of 463 participants, with 15 predictors and an effect size of f² = 0.15 (medium effect), the analysis revealed a high achieved power of 0.9999. This indicates that the sample size was more than sufficient to detect significant effects and robustly support the applied machine learning techniques. The critical F value for this analysis was 1.6888, confirming that the study had sufficient power to detect the anticipated effect size with high confidence.

To ensure the model selection process was both representative and scientifically sound, we first partitioned the entire dataset into a training set and a test set at an 80:20 ratio. The training of all models was conducted exclusively on the training set to maintain the objectivity and generalizability of the results on the test set. Subsequently, under a unified data input and parameter optimization strategy, we compared the performance of each model on the test set across four core metrics: Accuracy, Precision, Recall, and F1-score.

3.3.2. Interpretability Analysis

To move beyond mere predictive performance and gain a deeper understanding of the internal mechanisms of the model’s decision-making, this study employs a state-of-the-art eXplainable AI (XAI) framework: SHAP (SHapley Additive exPlanations) [18,74]. The SHAP method is rooted in cooperative game theory, and its primary advantage lies in its ability to fairly attribute a model’s individual predictions to each input feature. Specifically, a SHAP value (or Shapley Value) precisely quantifies the marginal contribution of each feature in pushing a prediction from the “average prediction of the entire sample” to the “specific prediction for the current instance.” This is currently the only attribution method that simultaneously satisfies the three desirable properties of local accuracy, missingness, and consistency, ensuring the reliability and fairness of the model’s interpretation [75].

This study applies the SHAP framework to conduct a multi-level attribution analysis to address our core research questions. First, global feature importance: by calculating the mean absolute SHAP value for each feature, we can obtain a global ranking of importance. This not only allows us to identify the most predictive drivers of PEB but also to precisely quantify their relative importance. Second, feature effect direction and non-linearity: we will utilize the SHAP bee swarm summary plot for visualization. This plot reveals not only the global importance of each feature but also, through color-coding (high/low feature values) and horizontal positioning (positive/negative SHAP values), the direction, distribution, and potential non-linear relationships of each feature’s impact on the model’s output at different value levels. Third, exploration of group heterogeneity: by calculating and comparing the SHAP feature importance rankings for different subgroups (e.g., partitioned by gender or academic year), we can quantify and uncover differences in the key mechanisms driving PEB across these groups.

This study uses SHAP to decompose feature contributions. In the presence of feature correlation or shared variance, SHAP attributions exhibit “path-dependent” properties, meaning that if multiple variables statistically represent the same underlying construct, their contributions will be distributed among these correlated variables. Therefore, this paper interprets the SHAP “importance ranking” as an indicator of association within the model, rather than as a causal explanation. Furthermore, priority is given to interpreting the results at a thematic level (e.g., Practice and Investment [PA], Attitude and Willingness [AV], Environmental Cognition [EC]), rather than over-interpreting individual variables within the same cluster. All SHAP results reported in this paper were obtained under the same data partitioning and evaluation settings to ensure comparability.

3.3.3. Group Heterogeneity Analysis

To move beyond an examination of “average effects” and delve into the structural differences in the drivers of PEB across various subgroups, this study further incorporates a group heterogeneity analysis. This analysis is designed to address our third research objective: to investigate whether the impacts of key drivers vary systematically according to individual characteristics.

The heterogeneity analysis in this study will be conducted based on three key grouping variables: gender (male vs. female), academic year (lower-division vs. upper-division students), and level of legal cognition (high vs. low). The analytical procedure is as follows: first, we will partition the entire sample into several mutually exclusive subgroups based on the aforementioned variables. Second, for each subgroup, we will independently train a new machine learning model using the same feature set, parameter tuning, and cross-validation strategies as the main model to ensure strict comparability across the sub-models. Finally, we will extract and compare the SHAP global feature importance rankings for each sub-model. By contrasting the changes in the ranks and weights of feature importance across different groups, we can identify key differences in the mechanisms driving PEB. This approach provides data-driven support for developing targeted and differentiated intervention strategies.

Compared to traditional methods such as interaction-term regression or multi-group structural equation modeling, this machine learning-based approach to heterogeneity analysis offers two core advantages. First, it possesses greater flexibility and capacity to capture non-linearities, enabling the identification of complex mechanism shifts in a high-dimensional feature space without relying on strict a priori assumptions (such as the functional form of interaction terms). Second, it constructs a customized predictive model for each subgroup. The results of this analysis not only reveal differences but also constitute more powerful predictive tools tailored to specific populations.

4. Results

4.1. Descriptive Statistics

This study collected a total of 463 valid questionnaire responses. The demographic characteristics of the sample are detailed in Figure 2. Specifically, the sample consisted of 59.40% females (N = 275) and 40.60% males (N = 188). In terms of academic year distribution, lower-division students (first and second year) constituted the majority of the sample (65.22%), with first-year students accounting for the largest proportion (52.05%, N = 241). Upper-division students (third year and above) and graduate students comprised 22.90% and 11.88% of the sample, respectively. Regarding family background, the proportions of students from urban and rural areas were nearly balanced, at 50.11% and 49.89%, respectively. Overall, the distribution of the sample across key demographic characteristics is well-structured, providing a solid foundation for the subsequent heterogeneity analysis.

Regarding the channels for acquiring environmental knowledge, the overall proportions are illustrated in Figure 3. The most frequently selected channel was “school” (N = 365, 78.83%), highlighting the pivotal role of higher education institutions in disseminating environmental knowledge. This was followed by “government agencies” (N = 228, 49.24%) and “environmental organizations” (N = 201, 43.41%), reflecting the significant role of formal institutions in environmental education. The “family” was a relatively less chosen channel, selected by only 137 respondents (29.59%). Meanwhile, “other” channels (such as social media platforms and independent media not covered by the preceding categories) accounted for a substantial proportion (N = 207, 44.71%), indicating that information acquisition pathways are increasingly diversified.

Concerning the perceived effective ways to enhance environmental awareness, the overall choices are also presented (Figure 4). An overwhelming majority of respondents selected “online media” (N = 433, 93.35%), suggesting that digital platforms are the primary pathway for university students to form their environmental cognition and attitudes. This was followed by “volunteer activities” (N = 288, 62.20%), “course-based education” (N = 237, 51.19%), and “books and magazines” (N = 218, 47.08%), indicating a high level of engagement in both practical activities and systematic learning. In contrast, the proportions for “academic reports” (N = 130, 28.08%), “exhibitions and competitions” (N = 123, 26.57%), and “newspapers” (N = 104, 22.46%) were relatively low. This suggests that the penetration of traditional media and formal academic formats is comparatively limited. Other methods (such as public science popularization and legal constraints not belonging to the above categories) accounted for a smaller share.

In summary, the sample for this study is well-structured in terms of gender, academic year, and urban-rural distribution, demonstrating strong representativeness. Concurrently, the respondents predominantly acquire environmental knowledge through school education and online channels, and they tend to enhance their environmental awareness through media information and practical activities. This phenomenon reflects a high level of attention and potential for engagement in environmental issues among contemporary university students, providing a solid foundation for the subsequent construction and interpretation of the pro-environmental behavior model.

4.2. Model Performance Comparison

To ensure the scientific rigor and reproducibility of this study, the modeling and evaluation process was conducted following a standardized machine learning analytical workflow. First, consistent with the common practice in most existing research [76,77,78,79,80], the entire dataset was randomly partitioned into a training set and a test set at an 80:20 ratio. All model training and hyperparameter tuning were performed exclusively on the training set, with the test set reserved as “unseen” data solely for evaluating the final model’s generalization performance. Second, on the training set, we employed 5-fold cross-validation (K = 5) to train and evaluate the CatBoost model. This approach helps to reduce the volatility of model performance and guides the selection of optimal hyperparameters. The specific operational procedure is illustrated in Figure 5.

Table 4. Performance Comparison across Different Machine Learning Models.

Model	Accuracy	Recall	Precision	F1-Score
SVM	89.247%	89.247%	90.160%	0.89
NN	89.247%	89.247%	89.224%	0.89
Decision Tree	89.247%	89.247%	89.224%	0.89
RF	91.398%	91.398%	91.399%	0.91
KNN	78.495%	78.495%	78.597%	0.78
CatBoost	93.548%	93.548%	93.548%	0.94
LightGBM	90.323%	90.323%	90.297%	0.90
XGBoost	89.247%	89.247%	89.224%	0.89

presents the performance of all evaluated models on the testing set. The results clearly indicate that the CatBoost model outperformed all other algorithms across every evaluation metric. It achieved an Accuracy of 93.55%, with Precision, Recall, and F1-Score all reaching 0.94. This performance was notably superior to that of other models, including Random Forest (F1 = 0.91), LightGBM (F1 = 0.90), and Support Vector Machine (F1 = 0.89). The exceptional performance of CatBoost can likely be attributed to its built-in optimizations for handling categorical features and its symmetric tree growth strategy, which provides a natural defense against overfitting. This aligns well with the characteristics of our dataset, which contains numerous categorical variables (such as school and information channels).

Given its superior predictive accuracy and robustness, the CatBoost model was selected as the final model for the subsequent explainability analysis using SHAP. The specific hyperparameter settings for the final model are detailed in

Table 5. Parameter Settings of the Model.

Parameter Name	Parameter Value
Training Set Proportion	0.8
Loss_Function	Auto
Iterations	500
Learning_rate	0.1
Depth	6
RSM	1.0
L2_leaf_reg	3.0

.

4.3. Feature Variable Importance Analysis

Using the feature importance analysis tool built into CatBoost, this study attributed and identified the key factors influencing individual pro-environmental behavior. As shown in Figure 6, the weights of the three variables—“Environmental Habit Formation,” “Active Practice Engagement,” and “Resource and Effort Investment”—accounted for 22.22%, 21.22%, and 20.20%, respectively, totaling 63.64%. This proportion significantly surpasses that of other variables, indicating that the establishment of individual habits, the extent of practical experience, and the degree of time/resource commitment are the primary drivers of pro-environmental behavior.

Building upon this, variables such as Relational Cognition (5.23%), Sense of Responsibility and Crisis (5.02%), and External Factors Influence (4.00%) also showed certain influence, suggesting that the depth of an individual’s cognition regarding environmental issues and their sense of responsibility play important roles in behavioral decision-making. Conversely, the weights of variables like gender, family location, and course participation were relatively low, potentially indicating weaker predictive power of these background variables on behavior.

4.4. SHAP Value Analysis

Based on the average SHAP values (mean (|SHAP value|)) calculated by the CatBoost model, this metric reflects the average marginal contribution of different feature variables to the prediction output of the student PEB model. From the overall results, the variable EB1 had the most significant impact on the model’s prediction outcome, with its average SHAP value exceeding 1.4, indicating its crucial role in distinguishing between individuals with high and low PEB levels. Closely following were PA3 and PA2, whose average SHAP values were also noticeably higher than other variables, confirming the core predictive importance of metrics related to actual practice. Among variables of moderate importance, EC3 and HL demonstrated certain predictive contributions, suggesting that an individual’s perceived relational cognition of the environment to their own life, along with their background and living context, may influence their environmental behavior to some extent. In contrast, the importance of environmental attitude indicators (such as “AV1” and “AV2”), legal cognition (“EC4”), basic environmental knowledge (“EC1” and “EC2”), and demographic features (“Grade,” “Gender”) was relatively low, suggesting that the marginal predictive contribution of these factors at the overall sample level is limited (Figure 7).

Figure 8 presents the SHAP bee swarm summary plot for the 15 feature variables. The results show that the SHAP values for the three indicators—“Environmental Habit Formation,” “Active Practice Engagement,” and “Resource and Effort Investment”—are mainly concentrated in the positive range. This indicates that the higher the value of these three variables, the greater the likelihood of the individual demonstrating active pro-environmental behavior. “Sense of Responsibility and Crisis” and “Relational Cognition” also displayed clear positive influence, suggesting that the more sensitive an individual’s perception of the ecological environment and the more accurate their understanding of its relevance, the more their behavior is oriented toward environmental protection. Since multiple top-ranked features belong to the same thematic domain (“Practice and Investment”), there is likely correlation and shared variance among them. Accordingly, this study interprets the finding of “Practice Primacy” as a domain-level (cluster-level) dominance, rather than as independent causal effects from single indicators. Under identical modeling and evaluation procedures, the domain-level ranking (PA dominating AV and part of EC) showed consistency across different models, supporting the robustness of the domain-level mechanism.

Notably, some background variables (e.g., gender, family location) rank lower in the importance hierarchy, with smaller overall deviation in SHAP values, indicating a limited marginal impact on the final prediction outcome. Furthermore, policy-cognitive variables, such as participation in relevant courses or awareness of environmental laws, although not the primary factors, still influence individual environmental behavior tendencies to some extent. This demonstrates the potential role of educational intervention and information dissemination in shaping environmental behavior.

4.5. Heterogeneity Analysis

In the overall sample, “Environmental Habit Formation” emerged as the foremost influential factor. “Crisis and Responsibility Awareness” and “Active Engagement in Practice” held greater weight among the female group, demonstrating stronger action-oriented characteristics. For the male group, on the other hand, “Resource and Ef-fort Investment” and “Value-Based Cognition” had a greater share, indicating that men’s environmental actions are more driven by emotions and values. The participants were categorized into lower-year and upper-year groups based on their academic standing. The lower-year group consists of first-year and second-year students, while the upper-year group includes third-year and higher-year students. With increasing academic year, “ Active Engagement in Practice “ and “ Resource and Effort Investment” became more important within the model, suggesting that senior students developed a more systematic environmental awareness at the cognitive level, but additional opportunities and resources to support hands-on participation remain necessary. In contrast, lower-year cohorts were more influenced by external factors like “External Impact”, Environmental Habit Formation”, and educational interventions such as “Course Participation.” Finally, for the legal awareness groups, categorization was based on participants’ scores on the legal awareness items (EC4). A score of 1–2 was categorized as the low legal awareness group, and a score of 3–4 was categorized as the high legal awareness group. Findings from analyzing groups based on “Level of Legal Awareness” revealed that “Crisis and Responsibility Awareness” and “Value-Based Cognition” were more prominent in the “Knowledgeable Group”, while model predictive accuracy was higher for this group, indicating that Legal Awareness enhances the consistency of pro-environmental behavior.”

The heterogeneity analysis above further confirms that the drivers of individuals’ pro-environmental behaviors are multidimensional and vary across different groups. In practice, interventions should be tailored to specific individuals, circumstances, and regions in order to strengthen the ability to implement precise policies. For example, for lower-year students, a focus should be placed on increasing curricular practical training. For female students, action may be incentivized by emphasizing values and psychological identification. It is also important to enhance the rate of environmental law popularization to increase overall group-wide action (

Table 6. Comparison of the Main Feature Weights in the CatBoost Model across Different Heterogeneous Subgroups.

Feature Variables	Overall Sample	Male Group	Female Group	Upper-Year Group	Lower-Year Group	High Level of Legal Awareness Group	Low Level of Legal Awareness Group
PA2	22.22%	27.13%	22.35%	15.31%	26.50%	20.95%	17.09%
PA1	21.22%	13.36%	31.69%	22.49%	17.66%	26.10%	30.10%
PA3	20.20%	28.52%	18.78%	32.16%	23.69%	23.03%	25.89%
AV2	5.02%	3.09%	3.22%	2.18%	3.50%	4.43%	2.35%
EC3	3.49%	5.64%	1.98%	4.94%	5.37%	3.98%	2.69%

).

The heterogeneity analysis further demonstrates that the drivers of individual pro-environmental behavior are multidimensional and vary across different groups. In practice, interventions should be customized to specific individuals, contexts, and regions to enhance the precision of policymaking. For example, for lower-division students, the focus should be on increasing practice-based training within the curriculum. For female students, behavior can be motivated by emphasizing value alignment and psychological identity. Concurrently, increasing the popularization of environmental laws could enhance the level of collective action across the entire population.

5. Discussion

5.1. Core Findings: A Shift from the “Knowledge–Action Gap” to a “Practice Primacy” Driver Model

The most significant contribution of this study lies in its data-driven revelation of the internal hierarchical structure of the drivers of university students’ PEB. This finding serves as a critical supplement to traditional linear models that typically start from cognition and attitudes [81,82]. First, our model results powerfully confirm and quantify the existence of the “Attitude-Behavior Gap.” The SHAP analysis clearly shows that although attitudinal variables like the Sense of Responsibility and Crisis (AV2) have a significant positive predictive effect on PEB, their global importance is far surpassed by that of practice- and habit-related variables. This contrasts with existing research [83,84] that identifies attitude as a key antecedent of environmental intention. Our model, by directly targeting actual behavior, captures the attenuation of attitude’s explanatory power during the conversion from “intention” to “action.” This suggests that a positive attitude does not automatically translate into consistent behavior, highlighting a significant structural divide between the two.

More importantly, this study identifies what we term a “Practice Primacy” driver model. The feature importance ranking (Figure 5) unequivocally indicates that the three practice- and quasi-behavioral variables—Environmental Habit Formation (PA2), Active Practice Engagement (PA1), and Resource and Effort Investment (PA3)—are the absolute dominant factors in predicting PEB, with a cumulative contribution of over 60%. This compelling evidence suggests that behavior itself may be the most potent catalyst for shaping and reinforcing long-term pro-environmental conduct. Initial, even minor, environmental practices may, by forming habits and enhancing self-efficacy, in turn strengthen an individual’s environmental identity and attitudes, thereby creating a positive feedback loop of “Behavior → Habit → Attitude Reinforcement.” This offers a novel perspective on the “knowledge–action gap”: the key to bridging this gap may not lie solely in the traditional “Knowledge-to-Action” pathway [28,30], but more so in an often-overlooked yet potentially more effective “Action-to-Cognition” feedback pathway.

From a comprehensive ranking perspective, the Practice and Investment (PA) variables appear at the top of the model’s importance hierarchy and maintain consistent direction and magnitude in the out-of-sample test set. This indicates that the capacity and resource provision for “action-promoting-action” are the primary constraints on students’ PEB. In comparison, although Attitude and Willingness (AV) are positively correlated with behavior, there is a systematic attenuation in the “intention-to-action” conversion chain. SHAP interaction analysis reveals a synergy between “practice” and “psychological distance/functional trade-off cognition” (EC2/EC3): when the provision for practice is weak, merely enhancing cognition does little to significantly improve PEB. However, once practice has reached a median level, a moderate increase in “psychological proximity” and “clarity of functional trade-offs” yields additional marginal benefits. This result suggests that university interventions should no longer focus solely on knowledge or attitudes as the primary levers but should instead prioritize supplementing the capacity and contextual conditions that make pro-environmental actions “doable.”

5.2. Practice Primacy: From an Empirical Observation to Theoretical Grounding

The core finding of “Practice Primacy”—the overwhelming predictive power of practice-related variables—is best understood not as a new theoretical proposition, but as a robust empirical observation from our data-driven model. The following discussion aims to ground this observation in established theoretical mechanisms within behavioral science to clarify its contribution.

The findings of this study provide strong, quantitative support for Habit Formation Theory in the context of PEB [85]. The high importance of “Environmental Habit Formation” (PA2) suggests that many daily environmental actions are not products of deliberate cognitive deliberation but are automated habits triggered by contextual cues. The practice-related variables in our model represent the engine of this habit loop, indicating that initial actions are crucial for forming the routines that appear to override fleeting attitudes in predicting behavior.

“Practice Primacy” resonates with Experiential Learning Theory, which posits that meaningful knowledge is acquired through action (“learning by doing”) [86,87]. When students engage in environmental practices, they gain not just abstract information but also concrete skills and an enhanced sense of self-efficacy. Our results suggest that this experientially acquired knowledge is a more potent predictor of future behavior than declarative knowledge gained through passive means.

At a deeper level, our observation aligns with the principles of Embodied Cognition, which argues that cognition is grounded in the body’s physical interactions with its environment [88,89]. From this perspective, PEB is a physical practice, and the bodily engagement in actions like recycling creates a more durable behavioral disposition than abstract reasoning alone. “Practice Primacy” thus highlights that the body’s “procedural knowledge” (knowing how) can be a stronger predictor than the mind’s “declarative knowledge” (knowing that).

In summary, the “Practice Primacy” pattern is not an isolated phenomenon but a convergence of these established mechanisms. Initial practices serve as the catalyst for habits, the vehicle for experiential learning, and the process of embodied cognition. Therefore, the primary theoretical contribution of this study is providing powerful, data-driven evidence for the dominance of these practice-based mechanisms in bridging the knowledge–action gap, offering a crucial supplement to traditional cognition- and attitude-centric models.

5.3. Non-Linearity and Heterogeneity of the Driving Mechanisms

In addition to identifying the relative importance of various drivers, the interpretability analysis in this study further reveals the complexity of their mechanisms of action. This complexity is mainly manifested in two aspects—non-linear effects and group heterogeneity—which transcend the explanatory scope of traditional linear models.

First, the influence of several key variables exhibits significant non-linear characteristics. Taking environmental knowledge as an example, the SHAP dependence plot suggests the potential existence of a “Cognitive Saturation” phenomenon in its marginal effect on PEB. That is, as an individual’s knowledge level (e.g., legal cognition) increases from low to medium, it has a significant promoting effect on pro-environmental behavior. However, once the knowledge level surpasses a certain threshold and becomes high, merely adding more knowledge yields diminishing or even negligible returns in promoting behavior. This non-linear relationship, which is difficult to capture in previous linear models [10,51], suggests that knowledge-infusion-based educational interventions may have a utility boundary. Consequently, different intervention strategies are required for groups with varying levels of cognition.

Second, one of the core findings of this study is the significant group heterogeneity in the driving mechanisms of PEB. As shown in Table 5, by independently modeling different subgroups, we discovered that their behavioral decisions rely on distinctly different psychological “levers” [41]. The gender difference was particularly striking: female students were more influenced by affective and value-based factors, such as Responsibility and Crisis Awareness (AV2) and Value-based Cognition (EC3), exhibiting a “Value-driven” decision-making model. In contrast, male students were more influenced by action-related factors, such as Resource and Effort Investment (PA3) and Environmental Practical Involvement (PA1), demonstrating an “Action-driven” model. Similarly, differences across academic years revealed a psychological shift during students’ development: lower-level students’ behavior was more susceptible to external structural interventions like “Course Participation,” whereas upper-level students relied more on internalized, systemic cognition, such as “Relational Cognition” (EC2). These findings provide compelling evidence that no “one-size-fits-all” universal intervention strategy exists. It is therefore crucial to design precise and differentiated programs tailored to the psychological characteristics and decision-making models of different groups [41,90].

On one hand, the non-linear effects present a “threshold-saturation” structure: the marginal effects of EC2/EC3 are significant in the low-to-medium range, suggesting that the priority should be to help students cross the “minimum actionable threshold.” Once cognition enters the high range, the diminishing marginal returns indicate a lower cost-effectiveness of continuing to infuse knowledge. On the other hand, the heterogeneity analysis shows that for students who have “not taken relevant courses” or “lack participation in clubs/projects,” the marginal benefit of enhancing PA is significantly higher than that of merely raising AV. Conversely, for students who “already have a strong practical foundation,” a moderate enhancement of complementary cognitions like EC2 and EC3 is more effective in consolidating and expanding their behavior. This sequential difference, termed “capacity-first, cognition-complementary,” provides direct evidence for stratified group interventions and precise resource allocation. Furthermore, error analysis and subgroup calibration indicate that the model’s prediction errors across different academic years and living arrangements can be significantly explained by “capacity-provision variables” (e.g., participation duration, task completion rates), further confirming the dominance of “contextual/capacity constraints” within the university ecosystem.

5.4. Theoretical and Practical Implications

The findings of this study hold significant theoretical and practical implications for both environmental behavior theory and pro-environmental education practices in higher education.

At the theoretical level, this study provides a critical supplement and extension to traditional models such as the Theory of Planned Behavior (TPB) and the Value-Belief-Norm (VBN) Theory through its integrated “prediction-plus-explanation” framework [91]. Our most central contribution is the proposed “Practice Primacy” driver model and the “Action-to-Cognition” feedback pathway, which challenge the linear thinking that overemphasizes cognition and intention in traditional theories [10,51,92]. This finding suggests that behavior, habits, and situational opportunities may play a more central role in the formation and consolidation of PEB than previously acknowledged, offering a new theoretical perspective on the underlying mechanisms of the “knowledge–action gap.” Furthermore, our quantitative revelation of non-linear effects and group heterogeneity advances environmental behavior research from exploring “average causal effects” to understanding “complex and contextualized mechanisms.”

At the practical level, our findings offer three precise, data-driven implications for universities and relevant policymakers to design more effective pro-environmental intervention strategies. First, the focus of intervention should shift from “knowledge dissemination” to “practice creation.” Given the “Practice Primacy” driver model, the center of gravity for intervention strategies should move away from traditional knowledge-based campaigns toward creating low-threshold, convenient, and high-frequency opportunities for students to engage in environmental practices. For example, optimizing campus waste-sorting facilities, promoting green office systems, and organizing regular community-based environmental activities can integrate pro-environmental behaviors into students’ daily routines. By doing so, habits and attitudes can be internalized through regular action, ultimately bridging the knowledge–action gap. Second, intervention programs should transition from “one-size-fits-all” to “precision irrigation.” Our study revealed significant group heterogeneity, suggesting that interventions must be tailored based on psychological and demographic factors. For instance, male students, who are more action-driven, may respond better to campaigns that promote environmental practices through hands-on experiences. On the other hand, female students, who tend to be more value-driven, may benefit from campaigns that align with their environmental values. Lower-division students, who are more susceptible to external influences, should receive early intervention integrated into courses and orientation programs. Upper-division students, who are more internalized in their beliefs, might benefit from deeper, more intellectual engagement in environmental decision-making, such as through research opportunities and project-based learning. Third, intervention resources should be reallocated from “universal input” to “marginal optimization.” The concept of “cognitive saturation” means that once students have acquired a certain level of knowledge, additional knowledge-based interventions offer diminishing returns. Instead, interventions should focus on providing higher-order engagement opportunities such as collaborative projects, peer learning, and behavioral tracking. These efforts can better help students internalize their pro-environmental attitudes and behaviors.

Accordingly, interventions for universities can follow an actionable “three-layer, four-loop” blueprint. Layer 1 (Propelling the Base Threshold): Use “micro-practice assignments + localized feedback” to help low-to-medium cognition groups cross the activation threshold. Layer 2 (Capacity and Provision): “Reduce friction and enhance capacity” through workshops, improved facility accessibility, and energy consumption visualization. Layer 3 (Stabilization and Diffusion): Solidify norms through club/project incubation and a tiered honor system. In practical terms, metrics such as “course hours,” “project participation duration,” and “facility accessibility” should be incorporated into the performance assessments of green campuses, alongside traditional “knowledge tests/attitude surveys.” In project funding and resource allocation, priority should be given to facility and mechanism upgrades that significantly reduce implementation costs and enhance situational accessibility.

The “Practice Primacy” model emphasizes the necessity of shifting from traditional knowledge dissemination interventions to creating opportunities for engaging in pro-environmental behaviors. This practice-oriented intervention strategy helps internalize students’ habits and attitudes through frequent environmental practices in their daily lives, thereby advancing the achievement of SDG 12 (Responsible Consumption and Production) by encouraging responsible consumption behaviors and reducing environmental impact. Additionally, by addressing the knowledge–action gap, this study supports SDG 13 (Climate Action), calling for urgent climate action. The research also intersects with SDG 4 (Quality Education), highlighting the role of customized educational interventions in fostering lasting pro-environmental behaviors and contributing to global sustainable development. Furthermore, integrating sustainability concepts into early education allows schools to cultivate future citizens with environmental awareness and responsibility, encouraging student participation in environmental practices. This contributes to the creation of sustainable, inclusive, and resilient urban communities, thereby making schools fundamental pillars for achieving SDG 11 (Sustainable Cities and Communities).

5.5. Limitations and Future Research

Although this study offers innovative findings and methodological approaches, it is subject to several limitations, which in turn point to new directions for future research.

First, the reliability and validity of the Environmental Cognition Evaluation scale were low (Cronbach’s α = 0.313), which may be primarily due to issues with item ambiguity and group suitability. Moreover, this study did not conduct measurement model tests to establish convergent and discriminant validity, including tests for AVE, CR, and Fornell-Larcker/HTMT values, as the focus was on exploratory analysis and the predictive accuracy of machine learning methods rather than traditional confirmatory factor analysis. Second, regarding causal inference, this study employed a cross-sectional design. While it successfully identified key predictors, it cannot make strict causal claims. For instance, the potential reciprocal causality between behavior and habits within the “practice-reinforcement” loop is difficult to disentangle with cross-sectional data. Finally, regarding sample representativeness, the data for this study were collected from a selection of universities in Zhejiang Province, China. Therefore, caution should be exercised when generalizing the findings to university student populations in different cultural, social, and institutional contexts. In terms of behavioral measurement, the dependent variable (PEB) in this study relied on a self-report scale, which may not completely eliminate social desirability bias.

Future research could address these limitations in several ways. First, to improve the generalizability of the findings, cross-regional and cross-cultural comparative studies should be conducted to test the applicability and context-dependency of the results. Second, adopting longitudinal tracking designs or field and quasi-experimental designs would allow for more rigorous testing of the causal effects of specific interventions. To enhance the internal consistency of the measurement, future studies could revise the Environmental Cognition Evaluation scale to reduce cross-dimensional influences or validate its effectiveness across different cultural backgrounds and populations. Additionally, future research could provide a more authentic representation of students’ pro-environmental behaviors by incorporating objective behavioral data—such as energy consumption records from campus cards, usage data from smart recycling bins, or behavioral tracking via mobile applications—and cross-validating it with subjective self-report data to mitigate social desirability bias. Furthermore, to test the robustness of the findings, future work could incorporate multi-source metrics, such as cluster-level importance, permutation importance, and interaction terms, to supplement the current framework and improve the reliability of the interpretations [93].

6. Conclusions

This study aimed to apply an eXplainable Artificial Intelligence (XAI) framework to investigate the key drivers of pro-environmental behavior (PEB) among Chinese university students, with the goal of dissecting the long-standing “knowledge–action gap.” By analyzing survey data from 463 university students, this research constructed a high-precision CatBoost predictive model (with an accuracy of 93.55%) and conducted an in-depth analysis of its internal decision-making mechanisms using the SHAP framework.

The core conclusions of this study challenge traditional linear behavioral models that typically start from cognition and attitudes. We found that a “Practice Primacy” model plays a dominant role in driving PEB: the formation of environmental habits, participation in environmental practices, and the investment of related resources are the overwhelmingly dominant factors in predicting whether an individual will take pro-environmental action. In contrast, while traditional cognitive and attitudinal variables still hold predictive value [94,95], their importance ranks relatively lower. This finding reveals an overlooked “Action-to-Cognition” feedback pathway, suggesting that behavior itself may be the most effective catalyst for shaping and consolidating environmental awareness and habits. Furthermore, this study quantified the complexity and context-dependency of these driving mechanisms. We identified non-linear effects in the influence of certain variables (such as “cognitive saturation”), as well as significant group heterogeneity. For example, the behavioral decisions of female students tend to be more “value-driven,” whereas those of male students are more “action-driven.” Lower-division students are more susceptible to external educational interventions, while upper-division students rely more on internalized, systematic cognition. This study underscores the importance of moving beyond traditional knowledge dissemination to fostering opportunities for pro-environmental behavior. By addressing the knowledge–action gap, this research provides valuable insights into how behavior change, driven by practical engagement, contributes to the achievement of SDG 12 (Responsible Consumption and Production) and SDG 13 (Climate Action). The findings highlight the role of customized educational interventions in aligning with SDG 4 (Quality Education) and demonstrate how integrating sustainability concepts into education can equip future generations with the skills and attitudes needed for long-term environmental responsibility, thereby emphasizing the crucial role of educational institutions in advancing SDG 11 (Sustainable Cities and Communities).

In summary, by constructing an integrated “prediction-plus-explanation” analytical path, this study not only provides a reusable, high-precision toolkit for environmental behavior research at the methodological level but also offers a new “Practice Primacy” perspective for understanding the structural causes of the “knowledge–action gap” at the theoretical level. Collectively, these findings point to a clear practical implication: to promote pro-environmental behavior among university students, future intervention strategies must undergo a shift in focus—from traditional knowledge dissemination to a greater emphasis on creating practical opportunities, cultivating behavioral habits, and designing precise, differentiated incentive programs based on the psychological characteristics of different groups.