Decoding Seating Preferences in Study Spaces via Explainable Machine Learning: Uncovering Micro-Scale Environment-Behavior Patterns Through the Lens of Gender and Efficiency

Abstract

Seat selection in learning spaces intuitively reflects user preferences for micro-environments and facilities. To address the lack of integrated analysis of multi-dimensional factors in existing research, this study constructs a framework merging multi-source dynamic sensing with explainable machine learning (XGBoost/SHAP/GAM) to decode the non-linear environment–behavior mechanisms underlying seat selection in study rooms. A multi-source dataset was constructed using YOLOv8 for non-intrusive extraction of seat occupancy and user attributes (gender and learning efficiency), combined with Ladybug-based luminous–thermal simulations and spatial topological measurements. The results indicate that: (1) key environmental variables exhibit distinct comfort thresholds, with an optimal illuminance of 400–600 lx and an effective attraction radius for power sockets of 1.5–3.0 m; (2) high-efficiency learners are highly sensitive to path interference, exhibiting a prominent “defensive” seat selection strategy; (3) significant divergences exist between genders regarding spatial depth preferences; and (4) compensatory and synergistic effects exist among multi-dimensional factors, where peak occupancy or superior lighting enhances user tolerance for the absence of sockets. This study quantifies the non-linear interactions between micro-physical environments and spatial behavior, providing a direct data-driven basis for refined facility deployment, dynamic luminous–thermal interventions, and scientific dynamic–static zoning in future learning spaces.

1. Introduction

In learning spaces, individuals’ seating behaviors are not driven by random habits or mere convenience; instead, they represent a comprehensive spatial response to environmental stimuli, psychological needs, and physical comfort [1,2]. Fundamentally, seat selection is the most intuitive expression of a user’s preference for the micro-scale physical environment and facility layout, which directly impacts individual learning efficiency, cognitive load, and psychological well-being [3,4]. In recent years, the imbalance between the supply and demand of physical space resources has become increasingly prominent, driven by the continuous expansion of student enrollments and the growing diversity of learning modes [5]. Therefore, scientifically decoding students’ micro-spatial selection mechanisms in complex indoor environments has become a critical challenge. Understanding these mechanisms is essential for guiding refined and human-centric space design and operation management for both educational institutions and architectural designers [6].

Regarding the interaction between the indoor micro-environment and spatial behavior, existing research primarily follows two main trajectories: physical environment perception and spatial facility layout. In terms of the physical environment, numerous studies based on environmental psychology indicate that lighting and thermal conditions are core factors affecting human comfort and task performance [7,8]. Regarding spatial layout, empirical studies based on topological geometry confirm that facility accessibility (such as proximity to power sockets or windows with natural daylight) significantly enhances seat attractiveness [9,10]. However, current research remains largely confined to isolated examinations of single variables (e.g., evaluating only lighting or analyzing only noise) or focuses on macroscopic overall layouts. These studies have not yet systematically revealed the joint mechanisms between luminous–thermal fields and multi-dimensional indoor facilities at the micro-scale [11,12]. In reality, behavioral decisions in physical spaces are often the result of trade-offs under multiple environmental constraints, meaning the linear superposition of single factors struggles to capture the full complexity of environment–behavior interactions [13].

Methodologically, traditional studies face dual bottlenecks in data granularity and analytical models. On one hand, data collection relies heavily on passive infrared sensors or subjective questionnaire surveys. These methods make it difficult to accurately obtain fine-grained occupant attributes (e.g., gender) and behavioral states (e.g., learning efficiency) across large samples [14,15]. On the other hand, existing literature frequently employs Pearson correlation analysis or multiple linear regression models. These linear assumptions fundamentally ignore the physical threshold effects of environmental variables on human comfort (such as glare aversion caused by excessive illuminance), and they struggle to effectively characterize the coupling and antagonistic effects among multi-dimensional features [16]. The aforementioned limitations lead to two critical gaps in current research: (1) a lack of high-resolution, multi-modal spatiotemporal behavioral datasets to support group heterogeneity analysis; and (2) an urgent need to introduce advanced models with strong non-linear fitting and interaction analysis capabilities to break through the “linear black box” in micro-environmental behavior research.

To address the aforementioned gaps, an innovative analytical framework integrating multi-source dynamic sensing and explainable machine learning is proposed to comprehensively decode the complex non-linear environment–behavior mechanisms underlying seat selection in study spaces. Specifically, it seeks to answer the following core questions: What are the relative importances and non-linear physical thresholds of various luminous–thermal and spatial facility features in driving seat selection? What spatial interaction and compensation effects exist among multi-dimensional environmental factors? How do spatial preference strategies exhibit heterogeneity across groups with different genders and learning efficiencies?

Methodologically, this research pioneers the integration of computer vision (high-precision extraction of real-time occupancy, gender, and efficiency attributes using YOLOv8), building performance simulation (high-resolution indoor luminous–thermal field reconstruction using Ladybug (v1.5.0)), and spatial topology computation [17,18,19,20]. Based on a newly constructed multi-source spatiotemporal dataset, the eXtreme Gradient Boosting (XGBoost) algorithm is employed for preference prediction. By deeply coupling the SHapley Additive exPlanations (SHAP) framework with Generalized Additive Models (GAM), a panoramic analysis—ranging from global attribution to local non-linear thresholds, and from main effects to two-dimensional interaction effects—is achieved [21,22,23].

The primary contributions of this research are threefold: (1) Theoretical deepening: The non-linear effects of environmental factors on seat selection are systematically quantified at the micro-scale, and the comfort thresholds of key physical variables are clearly identified. This provides empirical support for the “Adaptive Comfort Theory” and “Cognitive Load Theory” based on real-world behavioral data. (2) Behavioral decoding: Using GAM, it is revealed for the first time at the facility level that high-efficiency learning groups adopt a prominent “defensive” seating strategy, characterized by a high avoidance of disturbances and a strong reliance on key resources. Additionally, significant divergences in spatial depth preferences between male and female students are identified, expanding the understanding of micro-environmental behavioral heterogeneity. (3) Practical empowerment: Overcoming the limitations of traditional “single-facility determinism,” the compensation and synergy mechanisms between luminous–thermal conditions and spatial resources are proposed and validated. This provides direct, actionable quantitative guidelines for facility layout, dynamic–static zoning, and environmental control in learning spaces, advancing spatial design from empirical judgment toward data-driven approaches.

The remainder of this paper is organized as follows: Section 2 details the methodological framework for multi-source data collection and explainable machine learning; Section 3 systematically presents the quantitative analysis results regarding non-linear mechanisms, group heterogeneity, and interaction effects; Section 4 provides an in-depth discussion of the findings in the context of classical theories and proposes practical design strategies; Section 5 summarizes the conclusions and outlines future research directions.

2. Methodology

As illustrated in Figure 1, the proposed analytical framework [24,25] consists of four logically progressive core modules: (1) quantification of environmental and spatial features; (2) extraction of behavioral target data via machine vision; (3) multi-dimensional spatiotemporal data fusion and machine learning modeling; and (4) explainability analysis and mechanism decoding. This framework enables a full-chain analysis, ranging from micro-scale physical data collection to the visualization of deep spatial decision-making logic.

2.1. Study Area and Data Collection

2.1.1. Description of the Target Study Room

A representative study room within an academic building at Southeast University was selected as the empirical case for this study. As shown in Figure 2, the study room is located in the core academic zone of the campus (Figure 2a,b), and its host structure is a typical multi-story academic building (Figure 2c). The indoor floor plan is a regular rectangle (Figure 2d) with a total floor area of approximately 158.7 m², accommodating 151 fixed study seats arranged in a conventional matrix layout. The space incorporates various key physical and facility elements that influence seating preferences, specifically including main entrances, exterior windows for natural daylighting, internal power sockets, a front podium, and trash bins, with their spatial distribution illustrated in Figure 2d. The uneven spatial distribution of these facilities and resources provides an ideal physical setting for quantifying the resource accessibility and environmental interference levels associated with different seats.

2.1.2. Behavioral Data Collection and Privacy Protection

The collection of continuous spatiotemporal behavioral data relied on a high-definition network camera (model: KEDACOM IPC2833, KEDACOM, Suzhou, China) deployed above the front podium of the study room (at an installation height of approximately 2.4 m), whose wide-angle field of view effectively covered the vast majority of the seating area. The continuous monitoring period was from 30 November to 30 December 2025, with observation hours encompassing the entire daily opening time of the study room (08:00–20:00).

To strictly comply with research ethics and ensure the absolute privacy of space occupants, a non-intrusive “real-time processing and immediate deletion” data flow mechanism was adopted: all real-time recorded video streams were exclusively fed into the YOLOv8 object detection model on local computing nodes to extract the instantaneous bounding box coordinates and gender attribute classifications of the occupants. Once the extraction task was completed, the original image frames were immediately overwritten and physically destroyed. Throughout the entire research lifecycle, the system neither stored nor uploaded any raw visual images containing identifiable facial or personal identity features.

2.2. Seat Usage Data Extraction and Processing

2.2.1. Deep Learning-Based Object Detection and Tracking

The advanced YOLOv8 object detection framework was employed, utilizing pre-trained weights from the COCO dataset combined with fine-tuning on a customized scene dataset comprising 3850 images, to overcome the challenges of high-density occlusion and complex lighting unique to study rooms. The optimized model demonstrated excellent performance on the test set (achieving a mAP@0.5 of 95.4%). As shown in Figure 3, the model exhibited extremely high detection robustness: whether in empty (morning/evening), partially occupied, or extremely crowded (afternoon peak seating) states, the system accurately captured the real-time bounding box coordinates of seated occupants indoors, thereby laying a solid data foundation for generating the full-time-series seat occupancy matrix.

2.2.2. Spatial Mapping and Seat Status Matching

To achieve the precise mapping of dynamic occupants to the static physical space, fixed polygon masks were defined for each actual physical seat on the two-dimensional image plane. By extracting the geometric center points of the person bounding boxes output by the YOLO model and applying the Point-in-Polygon algorithm, the coordinate sequences of individuals were strictly matched to specific static seats. When an individual’s center point consistently fell within the mask region of a particular seat, that seat was marked as “Occupied” (status set to 1) at that specific time step; otherwise, it was marked as “Vacant” (status set to 0). Accordingly, a high-precision time series occupancy matrix covering the entire sample space was generated.

2.2.3. Classification of Occupant Attributes and Behavioral States

Beyond the basic occupancy status, the individual attributes and behavioral characteristics of the users were further extracted to support a deeper analysis of spatial preference heterogeneity. First, based on the classification layer output of the YOLO model, a strict confidence threshold (>0.7) was established to extract the perceived gender attribute (male/female) of the occupants. This classification relies entirely on visual macroscopic physical features captured by the overhead camera, such as hair length, body contours, and typical clothing styles. While acknowledging the inherent limitation that visual assessment captures outward physical gender expression rather than self-identified gender, this non-intrusive approach was employed as a practical proxy for analyzing group-level spatial distribution patterns. Verified via random manual sampling, the accuracy of this physical-based gender classification was approximately 91%. Second, drawing upon relevant environmental psychology literature regarding behavioral observation and the operationalization of sustained task engagement through dwell time metrics [26,27], “learning behavioral patterns” in this study refer to the observable temporal and physical engagement of an individual with their study task. Prior work has demonstrated that continuous stillness and prolonged occupancy in study spaces serve as valid behavioral proxies for cognitive engagement and productivity, with seat selection patterns—such as avoidance of aisle seats and preference for sheltered, resource-proximate positions—directly reflecting users’ underlying concentration needs [28,29]. Consequently, a “high-efficiency state” is operationally defined through observable behavioral metrics rather than internal cognitive output. By tracking the dwell time and object detection status of individuals in continuous video frames, an occupant was labeled as being in a “high-efficiency” state if they maintained a stable, focused seated posture for a continuous duration exceeding 45 min, without prolonged absences or obvious off-task behaviors (such as sleeping on the desk or continuous mobile phone usage. These fine-grained attribute data provided a solid empirical foundation for revealing the differences in spatial decision-making across groups with varying genders and efficiencies.

2.3. Quantification of Environmental and Spatial Features

2.3.1. Simulation and Field Measurement Validation of the Luminous–Thermal Environment

To obtain high-precision micro-scale physical environmental data, a 1:1 3D geometric model accurately corresponding to the actual physical space was constructed using the Rhino platform (v7.0, Robert McNeel & Associates, Seattle, WA, USA), and the Ladybug Tools suite was utilized for the quantitative simulation of the physical environment. EnergyPlus Weather (v9.2.0, U.S. Department of Energy, Washington, D.C., USA) (EPW) data for Nanjing were adopted as the external boundary conditions for the simulation, and the thermal and optical properties of all indoor surfaces were strictly configured based on on-site survey results and relevant building physics standard libraries. Ultimately, the system output the illuminance distribution (Lux) at the standard working plane height (approximately 0.75 m) and the air temperature field (°C) at the seated breathing zone height (1.1 m) during the opening hours of the study room.

Field Measurement Validation: To ensure the reliability of the data output by the luminous–thermal simulation engine, simultaneous on-site measurements were conducted during typical weather days using high-precision portable illuminance meters and temperature/humidity data loggers at 15 representative seat locations selected via spatial stratified random sampling. By cross-comparing and calibrating the collected field data with the model output results at corresponding times, the calculated root mean square errors (RMSE) for air temperature and desktop illuminance were strictly controlled within the internationally accepted ranges of 5.2% and 8.7%, respectively. Upon successful validation, a spatial interpolation algorithm was employed to accurately extract and assign a static luminous–thermal environmental feature vector to the center point of each valid seat.

2.3.2. Measurement of Facility Elements and Spatial Topological Distances

Based on an accurately calibrated 2D CAD floor plan, the coordinates of the geometric center points of key facility elements—including main entrances, exterior windows, interior windows, podiums, power sockets, and trash bins—were strictly extracted and defined. Based on this coordinate system, a Python (v3.9, Python Software Foundation, Wilmington, DE, USA) spatial analysis script was developed to iteratively calculate the shortest Euclidean distance from the center point of each valid seat in the room to the aforementioned facility network nodes. All distance values were precisely quantified in meters (m), successfully transforming the qualitative architectural floor layout of the study room into high-dimensional, structured spatial feature vectors that can be directly read and analyzed by machine learning models.

2.4. Dataset Construction

To meet the training requirements of the machine learning models, the dynamic behavioral data extracted via computer vision, the microclimate data simulated by Ladybug, and the topological distance data calculated using Python were precisely aligned and fused across spatiotemporal dimensions. The final constructed dataset comprises 7852 valid spatiotemporal samples.

In this dataset, each sample represents a micro-state snapshot of “a specific physical seat at a specific observation time point.” The input features (independent variables) were systematically categorized into three main dimensions: (1) physical environmental features (including air temperature temp and working plane illuminance illum); (2) spatial and facility features (encompassing distances to the entrance/exit dist_door, window dist_window, socket dist_socket, podium dist_podium, trash bin dist_trash, and aisle dist_aisle, as well as seat height seat_height); and (3) behavioral and temporal features (including path impact degree path_impact and time point time).

The corresponding target variables (dependent variables) move beyond the limitations of binary classification to finely characterize seat occupancy tendencies at a multi-dimensional granularity. Specifically, these include total occupancy count (total_count), occupancy counts by gender (male_count, female_count), and high-efficiency occupancy counts (eff_total, eff_male, eff_female). To ensure an objective evaluation of the machine learning models’ generalization capabilities, the overall dataset was randomly split into a training set and a testing set at a ratio of 8:2. The detailed definitions and descriptive statistical characteristics (e.g., mean, standard deviation, and extreme values) of the core variables in the dataset are detailed in

Table 1. Descriptive Statistics of Variables.

Variable Type	Abbreviation	Variable Name & Description	Unit	Mean	SD	Min	Max
Dependent Variable	total_count	Total occupancy count	Count	4.25	2.72	0.00	11.00
	male_count	Male occupancy count	Count	2.81	2.43	0.00	10.00
	female_count	Female occupancy count	Count	1.44	1.91	0.00	10.00
	eff_total	Total high-efficiency occupancy	Count	1.71	2.02	0.00	10.00
	eff_male	High-efficiency male occupancy	Count	1.06	1.56	0.00	10.00
	eff_female	High-efficiency female occupancy	Count	0.64	1.31	0.00	9.00
—Physical Environment	temp	Air temperature	°C	18.11	0.82	15.79	20.10
	illum	Desk-level illuminance	Lux	656.95	184.00	350.00	1182.97
—Spatial Features	dist_door	Distance to nearest door	m	6.42	2.02	1.69	10.57
	dist_window	Distance to nearest window	m	4.71	2.13	1.13	8.31
	dist_socket	Distance to nearest socket	m	2.47	1.24	0.82	4.32
	dist_podium	Distance to front podium	m	8.88	3.37	2.71	14.53
	dist_trash	Distance to nearest trash bin	m	8.40	3.27	2.20	14.56
	dist_aisle	Distance to nearest aisle	m	1.32	0.36	0.71	1.93
	seat_height	Physical height of the seat	m	0.90	0.21	0.65	1.26
—Behavior & Time	path_impact	Walking path disturbance level	m	0.32	0.23	0.00	1.00
	time	Time of data recording	Hour	14.00	3.74	8.00	20.00

.

2.5. Explainable Machine Learning Framework

2.5.1. Baseline and Non-Linear Predictive Models

A standard Logistic Regression was first constructed as a linear baseline model to evaluate the linear separability and initial correlations of the environment–behavior features. Given that environmental factors frequently exhibit complex threshold effects, the eXtreme Gradient Boosting (XGBoost) algorithm was introduced as the core non-linear predictive framework [30]. By ensembling multiple decision trees, XGBoost can efficiently handle high-dimensional features and adaptively capture the complex non-linear and interactive relationships among variables. During the model training phase, Bayesian Optimization combined with K-fold Cross-validation was employed for the global optimization of key hyperparameters (e.g., learning_rate, max_depth, n_estimators) to maximize the model’s generalization ability and prevent overfitting.

2.5.2. SHAP Explainability Engine

To overcome the “black box” nature of complex ensemble tree models, the game theory-based SHAP (Shapley Additive exPlanations) framework was introduced as the core explainability engine [31,32]. The SHAP algorithm can calculate a unified marginal contribution with clear physical meaning for each feature of every spatiotemporal sample. SHAP values were utilized to conduct a three-level analysis: (1) extracting global Feature Importance; (2) plotting SHAP Dependence Plots to precisely characterize the main effect curves and physical comfort thresholds of environmental variables; and (3) calculating bivariate SHAP Interaction Values to quantify the spatial synergistic compensation or antagonistic effects between luminous–thermal conditions and spatial facilities.

2.5.3. GAM-Based Modeling of Group Preference Heterogeneity

To further investigate the nuanced differences in spatial preferences among groups with different genders (male/female) and learning states (high-efficiency/normal), a Generalized Additive Model (GAM) was additionally introduced alongside XGBoost. While maintaining high interpretability, GAM possesses a remarkably strong capacity for local smooth fitting [33,34,35]. By fitting GAM partial dependence curves separately for different group subsets, the spatial psychological heterogeneity of varying populations in resource-seeking (e.g., socket demand) and interference avoidance (e.g., avoiding entrance points) strategies was visually and quantitatively compared.

3. Results and Analysis

3.1. Baseline Characteristics of Physical Environment and Spatial Facilities

The spatiotemporal distribution of indoor temperature exhibited significant diurnal fluctuations and spatial heterogeneity. As illustrated in Figure 4, the thermal environment within the study room underwent a distinct “cool–hot–mild” dynamic evolution during its opening hours (08:00–20:00). During the early morning (08:00–10:00), the overall indoor temperature was relatively low, generally ranging from 15.8 °C to 16.6 °C. With increased solar radiation and the continuous accumulation of occupants, the afternoon (13:00–16:00) marked the daily thermal peak, during which temperatures in localized hotspots climbed above 19.0 °C to 20.1 °C. This period also exhibited a non-uniform spatial distribution, with temperatures in the middle and right zones being significantly higher than those in the left zone. By the evening and nighttime (18:00–20:00), the temperature gradually decreased and became more homogeneous, stabilizing at approximately 17.0 °C to 18.2 °C. These dynamic changes in the thermal field constitute the baseline physical environment for seat selection. Consequently, users entering the study room at different times must actively seek out seating zones that align with their personal thermal comfort thresholds.

Under the combined effects of natural daylight and artificial lighting, the spatial distribution of desktop illuminance exhibited pronounced regional disparities and luminous climate gradients. As illustrated by the spatiotemporal heatmap of illuminance in Figure 5, the luminous environment was profoundly influenced by external natural daylight. During the high-illuminance daytime period (12:00–14:00), the peripheral zones adjacent to the exterior windows experienced excessively high illuminance, with peak values surging beyond 1000 lx to 1182 lx, posing a potential risk of visual glare. Conversely, the inner core zones farther from the windows maintained a relatively stable and mild luminous environment, generally stabilizing within the conventional working illuminance range of 400 lx to 600 lx. Transitioning into the nighttime (18:00–20:00), under the exclusive dominance of artificial lighting, the overall indoor illuminance decreased substantially and became more uniform, generally fluctuating between 350 lx and 550 lx. Such dramatic spatiotemporal contrasts in the luminous environment compelled users to trade off between seeking sufficient daylight and avoiding visual fatigue, acting as a crucial physical trigger driving localized seating preferences.

The spatial topology of the indoor layout and facility elements constructs a multidimensional network of accessibility and interference, defining the objective functional attributes of each seat. Figure 6 details the static distribution patterns of several key spatial distances and behavioral feature variables. Regarding distance characteristics, the distribution of power sockets (dist_socket) exhibits pronounced band-like spatial disparities; the nearest distance is merely 0.82 m, whereas the farthest exceeds 4.32 m, intuitively reflecting the uneven accessibility of power resources. Simultaneously, areas adjacent to the entrances/exits (dist_door) and main aisles demonstrate an exceptionally high degree of path interference (path_impact), with impact values approaching the extreme of 1.0 in core circulation zones. Furthermore, the distance distributions to the podium (dist_podium, ranging from 2.71 m to 14.53 m), windows (dist_window), and trash bins (dist_trash) precisely delineate the geometric relationships between seats and various resource points or sources of disturbance. These static elements essentially represent a trade-off between “convenience” and “interference resistance” in spatial utilization, providing robust structural constraints for the subsequent environment–behavior decision-making models.

3.2. Spatiotemporal Distribution Patterns of Seat Occupancy Behavior

The actual occupancy frequencies of study room seats and the distribution of high-efficiency learning states exhibited significant spatial agglomeration characteristics and group heterogeneity. As illustrated by the behavioral positioning heatmap in Figure 7, the total occupancy frequency (Figure 7c) demonstrated a highly uneven spatial distribution: high-frequency hotspots (red zones, with frequencies reaching 80 to over 96 times) were highly concentrated in specific middle–rear rows and certain corner areas, whereas interference zones adjacent to high pedestrian flow were notably underutilized. Regarding group differences, male students (Figure 7a) displayed a relatively dispersed seating distribution while forming a broad, dense cluster in the middle–left area; conversely, female students (Figure 7b) exhibited a more convergent tendency toward localized hotspots, reflecting potential gender disparities in spatial territoriality or micro-environmental preferences. More crucially, the distribution maps of high-efficiency learning states (Figure 7d–f) were not merely proportional downscalings of the total occupancy; rather, the hotspots of high-efficiency individuals (red zones, with peak frequencies exceeding 60) were extremely concentrated in specific interior seats. This intuitively demonstrates that occupants exhibited strong active seat selection behaviors within the physical space, driven not only by the basic need for seating but also by the desire to align with specific cognitive goals and high-efficiency states.

The temporal dynamic evolution of seat occupancy followed a “tidal” rhythm dominated by students’ daily routines, accompanied by a stable underlying gender structure and efficiency output proportion. As illustrated in Figure 8, the time series of total occupancy (Figure 8a) exhibited a dramatic wave-like trajectory, featuring two distinct precipitous valleys at 12:00 and 18:00 (where cumulative occupancy plummeted below 100 instances), which highly coincided with conventional lunch and dinner breaks. Conversely, the daily peaks were concentrated at 10:00, 14:00–16:00, and after 19:00 (with the nighttime peak exceeding 800 instances). Regarding the overall demographic composition (Figure 8b), male students dominated the sample base, accounting for 66.2%, which was approximately twice the proportion of female students (33.8%). In terms of learning states, occupants in the “high-efficiency” state accounted for 40.2% of the total, while the remaining 59.8% were in a normal state. Notably, the line representing the number of high-efficiency individuals exhibited a significant surge at 19:00. This macroscopic spatiotemporal rhythm and demographic profile thoroughly outlined the fundamentals of the data, establishing a factual baseline for the subsequent in-depth analysis of how multidimensional environmental factors intervened at different time points to quantitatively affect individual behavior.

3.3. Initial Correlation Between Environment and Seat Selection

The relationship between environmental variables and spatial seat selection behavior exhibits highly non-linear characteristics, making it difficult for traditional linear statistical methods to effectively capture the underlying driving mechanisms between them. As illustrated by the Pearson correlation heatmap in Figure 9, the initial linear correlations between various physical environmental features and seat occupancy variables across different user groups were comprehensively quantified. From the feature–target cross-matrix on the lower right side of the figure, it can be clearly observed that the absolute values of the Pearson correlation coefficients (r) between almost all environmental variables (whether luminous–thermal physical indicators or spatial distance elements) and all seat occupancy indicators (including total count, gender categories, and high-efficiency count) were at extremely low levels. Specifically, the correlation coefficient between the “distance to the entrance/exit (dist_door)”, which was presumed to significantly influence seat selection, and the “total count (total_count)” was merely 0.14, while the correlation between “path impact (path_impact)” and the total count was −0.13. Moreover, the linear correlations between core physical environmental indicators, such as temperature (temp) and illuminance (illum), and the total count approached zero, at −0.03 and −0.07, respectively. Furthermore, even for the high-efficiency population (eff_total) with clear learning objectives, the linear responses of various feature variables remained exceedingly weak, with the absolute values of correlation coefficients generally below 0.25 (e.g., the most impactful variable, distance to the socket dist_socket, was only −0.24). These results intuitively and conclusively demonstrate that micro-scale spatial usage behaviors in study rooms are not driven by single environmental factors in simple linear proportions. The effects of environmental elements (such as light, heat, and distance) on human comfort and psychological decision-making often involve specific thresholds and complex cross-coupling effects (e.g., higher illuminance is not always better, and the attenuation of a socket’s attractiveness is not strictly a linear decrease). Therefore, this baseline reality of comprehensively weak linear correlations provides robust logical support and absolute necessity at the methodological level for this study to subsequently break away from linear assumptions and introduce advanced machine learning frameworks (such as XGBoost and GAM) equipped with powerful non-linear fitting and feature interaction analysis capabilities.

3.4. Global Driving Mechanisms of Environmental–Spatial Features

The driving weights of environmental–spatial features on study room seat selection behavior exhibit significant hierarchical differentiation, with temporal dynamics, thermal physiological comfort, and microcirculation interference constituting the core mechanisms dominating user decisions. As illustrated in Figure 10, the global Feature Importance bar chart and the SHAP Summary plot generated by the XGBoost model intuitively and quantitatively deconstruct the relative contributions and effect directions of various variables on total seat occupancy (total_count). Regarding the ranking of feature importance weights, temporal rhythm (time) unsurprisingly occupies an absolute dominant position, with an importance score reaching approximately 0.73, acting as the strongest external driving force propelling the evolution of the system state. Excluding the temporal dimension, “temperature” (temp) within the luminous–thermal environment and “path impact” (path_impact) representing behavioral interference exhibit extremely high decision weights, scoring approximately 0.54 and 0.41, respectively, thus serving as the two core physical barriers constraining seat attractiveness. In contrast, the group of traditional spatial distance elements (e.g., distance to window dist_window, distance to door dist_door, and distance to aisle dist_aisle) constitutes the second tier of importance, with weights concentrated between 0.23 and 0.24. Meanwhile, the global statistical importance of functional indicators such as desktop illuminance (illum, approx. 0.20) and distance to socket (dist_socket, approx. 0.15) ranks relatively lower.

Incorporating the SHAP summary plot on the right allows for further analysis of the effect directions and non-linear distributions of these features: for instance, the high-value range of path_impact (red scatter points) is densely distributed in the negative SHAP value zone, confirming the strong repulsive effect of dynamic pedestrian flow on seat selection; conversely, the low-value range of dist_socket (blue scatter points) exhibits a distinct positive SHAP gain (attractiveness) locally. This importance distribution pattern profoundly indicates that students’ spatial selection behaviors follow an implicit logic akin to “Maslow’s hierarchy of needs”: within the established framework of daily routines, students prioritize meeting the baseline for thermal comfort and strenuously avoid external interference to maintain concentration; only after securing these two “safety bottom lines” do they proceed to make trade-offs and compromises regarding secondary resource facilities, such as window views and power sockets. This provides a direct algorithmic basis for dismantling “single-facility determinism” and implementing synergistic designs of environments and spaces.

3.5. Non-Linear Main Effects and Comfort Thresholds of Key Variables

The main effects of environmental–spatial features on study room seat selection behavior are not simply linear monotonic increases or decreases; rather, they exhibit highly regular non-linear partial dependence characteristics and clear physical comfort thresholds. As illustrated by the SHAP dependence plots in Figure 11, fundamental physical environmental factors such as luminous and thermal conditions demonstrate highly significant “comfort zones” and “abrupt critical shifts.” Specifically, observing the partial dependence curve for desktop illuminance (illum, Figure 11g) clearly reveals the users’ non-linear adaptation boundary to the luminous environment: when the desktop illuminance falls within the conventional range of 400 lx to 600 lx, the SHAP values remain in the positive gain zone, reflecting the strong attractiveness of this range as the “optimal illuminance threshold” for learning behaviors. However, when the illuminance crosses the implicit boundary of 600 lx and continues to rise (especially after exceeding 800 lx), the SHAP values representing the model’s predicted impact exhibit a precipitous plunge and fall deeply into the negative zone. From the perspective of physical behavior, this numerical pattern accurately corroborates that excessive direct natural sunlight or unreasonable artificial lighting can easily cause visual glare and fatigue, thereby generating a severe negative repulsion effect. Similarly, the curve trend for air temperature (temp, Figure 11h) explicitly outlines the lower limit of human thermal physiology. When the temperature drops below 17.5 °C, the SHAP values predominantly reside in the negative range, indicating that a cold micro-environment has a significant repellent effect; conversely, when the temperature crosses 17.5 °C and climbs to the range above 18.5 °C, the marginal utility rapidly surges and turns positive, demonstrating that a comfortably warm micro-environment can greatly enhance the selection probability of a seat.

The accessibility of facility resources drives seat preferences through a typical “effective service radius” and a non-smooth distance decay pattern. Analyzing the non-linear response curve for the distance to a power socket (dist_socket, Figure 11d) reveals that the attractiveness of power resources does not decrease uniformly as the distance increases. The curve features a critical structural inflection point at a distance of approximately 1.5 m from the socket (indicated by the dashed “Threshold” line in the figure). Within this distance, seats generally maintain high marginal attractiveness. However, when the distance crosses a specific convenience threshold (e.g., the area approaching 3.5 m in the figure), the SHAP value experiences a sharp decline, plummeting into a deep trough of negative feedback. This phenomenon profoundly demonstrates that learners in the modern digital age have a “rigid baseline demand” for power support. For seats located beyond the effective connection length of a standard power cord, their spatial practical value suffers a “precipitous” shrinkage; this threshold directly quantifies the optimal coverage radius for socket facility configurations in study rooms.

Spatial topological layout plays a dual role in regulating “interference management” and “psychological security” in seat selection, forming specific optimal distance peaks along the spatial depth. The SHAP curves for the distance to the door (dist_door, Figure 11b) and the distance to the podium (dist_podium, Figure 11c) vividly depict users’ psychological avoidance of circulation interference. Regarding the distance to the entrance/exit, the curve exhibits a classic “inverted U-shape” characteristic: when the distance is too close (<4.0 m), the noise and light disturbances caused by frequent pedestrian flow and door operations result in negative SHAP values; conversely, when the distance is between 6.0 m and 8.0 m, the curve reaches a positive peak, creating a “golden buffer zone” that maintains appropriate tranquility without causing a sense of oppression due to excessive remoteness; once the distance becomes too far (>10.0 m), the attractiveness decays slightly again due to a feeling of isolation in spatial dead corners. On the other hand, the distance to the podium (dist_podium) shows a relatively stable positive attraction in the front-row areas (<6.0 m) and subsequently plummets into the negative zone after exceeding 8.0 m, reflecting the comparative advantages of the front and middle rows in terms of an open field of view or a focused learning atmosphere. These non-linear inflection points and extreme value zones extracted via the algorithm provide precise quantitative parameter support for the refined layout of indoor facilities and the design of dynamic–static zoning at the microscale.

3.6. Heterogeneity of Spatial Preferences: Efficiency and Gender Perspectives

Spatial behavioral preferences exhibited significant strategic divergence under varying learning efficiency states, with high-efficiency learners demonstrating more stringent and rational environmental selection standards in terms of resource acquisition and disturbance avoidance. As illustrated in Figure 12 by the comparison of non-linear partial dependence between total occupancy and high-efficiency occupancy based on the GAM model, the high-efficiency group (green curve) exhibited fluctuation patterns similar to the overall trend (black curve) across multiple key features but presented more intense positive and negative feedback in extreme value ranges. In the dimension of resource acquisition, observing the distance to power sockets (dist_socket, Figure 12d), the high-efficiency group formed an exceptionally concentrated and towering positive preference peak within the optimal service radius of 2.0 m to 3.0 m, which then rapidly plunged into negative values after exceeding 3.5 m. This indicates that stable power support is a rigid prerequisite for maintaining a deep learning state. In the dimension of disturbance avoidance, regarding path impact (path_impact, Figure 12e), when the interference index crossed the high-intensity threshold of 0.8, the partial dependence value of the high-efficiency group exhibited a deeper “precipitous” drop than the total population, reflecting a strong evasion of areas with high-frequency pedestrian flow. Furthermore, from the perspective of temporal rhythm (time, Figure 12i), the peak of high-efficiency learning behavior during the evening period (19:00–20:00) far exceeded that of the daytime, demonstrating an extremely strong nocturnal concentration of focus. This series of numerical comparisons profoundly reveals that high-efficiency learning behavior is not a random positioning within the physical space, but rather a highly proactive “defensive” spatial strategy: users minimize the cognitive load brought by external physical stimuli by precisely targeting high-resource, low-interference, and high-quality micro-environments, thereby ensuring the continuity of deep thinking and high-efficiency output.

In the gender dimension, male and female students exhibited highly heterogeneous spatial psychological characteristics and functional demand differences in seat selection. As shown in Figure 13, the GAM model precisely quantified the preference divergences between male (blue curve) and female (pink curve) students across multiple environmental–spatial elements. The most significant difference lies in the perception of spatial depth (dist_podium, Figure 13c): female students’ preference curve reached its highest peak in the front–middle row area, 3.0 m to 6.0 m from the podium, and subsequently decayed substantially into the negative zone as the distance increased; conversely, male students’ preference curve displayed a diametrically opposite trend, climbing monotonically with increasing distance and reaching a positive extreme in the rear peripheral areas beyond 10.0 m. Additionally, regarding the reliance on power sockets (dist_socket, Figure 13d), the positive peak for males was significantly higher than that for females, reflecting that they might have a higher frequency of electronic device usage and greater demand for power; meanwhile, in terms of temperature adaptability (temp, Figure 13h), the female curve fluctuated more gently overall and maintained better positive feedback in the higher temperature range (above 17.5 °C), reflecting subtle differences between the sexes in thermal physiological comfort thresholds. These quantitative facts not only corroborate the theories of “territoriality” and “personal space” in environmental psychology at the micro-scale—with females tending to choose front–core areas with clear views and strong psychological security, while males preferring rear–peripheral areas with high concealment, backing against walls or corners—but also provide robust data support for the gender-responsive refined design of internal study room resources (e.g., socket density, thermal control zoning).

3.7. Spatial Coupling and Interactive Synergistic Effects

Seat selection in real physical spaces is not independently governed by a single environmental variable; rather, it exhibits complex spatial coupling and interactive synergistic effects among various facilities and physical dimensions, profoundly revealing users’ trade-off and compromise strategies under multiple constraints. As illustrated by the two-dimensional bivariate interaction heatmaps in Figure 14, the joint mechanisms between multiple key pairs of variables are intuitively and quantitatively mapped via a color gradient from dark red (high preference) to dark blue (strong avoidance). First, a typical “crowding compromise” phenomenon can be observed in the interaction plot of “socket distance vs. time” (dist_socket vs. time): during peak occupancy hours such as 14:00 and 19:00 (indicated by distinct dark red horizontal bands in the figure), the occupancy probability remains high even if seats are far from a socket (e.g., >3.5 m). However, during off-peak hours, the disadvantage of a long socket distance leads to a sharp collapse in seat attractiveness (with the dark blue area expanding significantly), vividly depicting the concessionary behavior of users forced to abandon rigid power demands when resources are constrained.

Second, analyzing the coupling relationship between “illuminance and socket distance” (illum vs. dist_socket) reveals that close proximity to a socket (<2.5 m) establishes a baseline of high preference. If the desktop illuminance simultaneously falls within the optimal comfort range of 400 lx to 600 lx, the red saturation in the heatmap reaches its peak, demonstrating a synergistic amplification effect resulting from the combination of facility convenience and a high-quality physical environment. Conversely, if the socket distance exceeds the effective threshold (>4.0 m, shown as the dark blue vertical band on the right), even superior lighting conditions cannot reverse this disadvantage, indicating that the absence of core resources exerts a strong “veto” effect. Furthermore, in the interaction plot of “window distance vs. socket distance” (dist_window vs. dist_socket), when seats are far from sockets (the upper half of the chart), only those extremely close to windows (<2.0 m, the leftmost warm-colored band) manage to regain some attractiveness. This reveals a spatial compensation mechanism where natural daylight and an open field of view offset the lack of power access. These two-dimensional interaction phenomena not only dismantle the linear assumption of “independent variable effects” prevalent in traditional research but also decode the hierarchy of users’ psychological decision-making at a deeper level. They demonstrate that in future refined spatial designs, facility managers can effectively balance overall seat attractiveness through “compensatory” strategies—such as optimizing the luminous environment or window views in socket-free zones—thereby maximizing the spatial utilization efficiency of study rooms.

4. Discussion

4.1. Micro-Environmental–Behavioral Decision Framework Under Multidimensional Constraints

The results of this study profoundly reveal that study room seat selection behavior at the micro-scale is not dominated by a single environmental stimulus, but rather driven by a multidimensional decision-making framework centered on “Environmental Controllability” and “Interference Management” [36]. Global analysis based on explainable machine learning (XGBoost and SHAP) indicates that users’ seat selection behaviors in physical spaces follow an implicit filtering mechanism akin to “Maslow’s hierarchy of needs”: First, basic thermal physiological comfort (e.g., avoiding cold zones below 17.5 °C) and the evasion of micro-circulation interference (e.g., staying away from areas with high-frequency pedestrian flow) constitute an uncompromising “safety baseline.” Second, only after satisfying this baseline do users further pursue secondary resource elements (e.g., power support from sockets, adequate natural daylighting). More crucially, the two-dimensional interaction analysis overturns the traditional assumption of “independent effects of each element,” confirming the existence of complex compensation and compromise strategies in real-world spaces. For instance, when a seat is in a disadvantaged position far from a socket (>3.5 m), its spatial attractiveness is “compensated” or accepted through compromise only if it is close to an exterior window or during specific peak hours. This comprehensive decision-making framework accurately characterizes the intensity and boundaries of individual preferences under complex environmental stimuli, demonstrating that the seat occupancy rate is not merely a function of spatial facility density, but rather a behavioral response resulting from the synergistic interaction of multidimensional environmental factors [37]. Critically, this highlights the necessity and superiority of the applied analytical framework. Traditional linear models (e.g., Pearson correlation or multiple linear regression) commonly used in previous studies would have entirely masked these “threshold effects” and “compensatory strategies” due to their assumption of monotonicity. By justifying the selection of the XGBoost ensemble model coupled with the SHAP/GAM explainability frameworks, this study successfully transitions from merely confirming “what” environmental factors matter to precisely decoding “how” they interact non-linearly, thereby providing a more rigorous methodological foundation for environmental psychology.

4.2. Quantitative Validation and Micro-Scale Extension of Classical Theories

Through high-precision spatiotemporal behavioral data and non-linear models, this study provides direct quantitative empirical evidence at the micro-scale for several classical theories in environmental psychology and cognitive psychology [38,39]. First, this study highly aligns with the “Adaptive Comfort Theory.” Traditional theory posits that individuals adapt to their environment through proactive behavioral adjustments, and the SHAP dependence plots in this study clearly demonstrate how students optimize their micro-environmental comfort by actively occupying the illuminance range of 400–600 lx and the temperature range of 17.5–18.5 °C. Second, the extreme sensitivity and strong avoidance tendency toward “path interference” exhibited by high-efficiency learning groups in the non-linear partial dependence plots (GAM) provide solid evidence for the spatial application of the “Cognitive Load Theory” [40,41]. By making extremely stringent environmental choices (e.g., locking onto deep, low-interference zones and the optimal socket service radius), high-efficiency groups effectively shield themselves from irrelevant external stimuli, thereby preserving valuable cognitive resources for deep learning tasks. Finally, the significant divergence discovered in the gender difference analysis—where male students prefer the deep rear areas while female students prefer the front–middle core areas—precisely corroborates the theories of “Territoriality” and “Personal Space” at the micro-facility level, indicating inherent psychological heterogeneity between genders in establishing spatial security and environmental defense mechanisms [42,43].

4.3. Strategies for Refined Design and Management of Learning Spaces

Based on the non-linear thresholds and group preference characteristics quantitatively extracted in this study, the following data-driven, direct recommendations are proposed for the refined design and operational management of future learning spaces: (i) Precise deployment of power facilities: The attenuation of socket attractiveness exhibits a clear “precipitous” critical point. Designers should abandon the strategy of blind, uniform distribution and ensure that within high-value study zones, the absolute physical distance from any seat to a power socket is controlled within the optimal service radius of 1.5 m to 3.0 m; areas exceeding 3.5 m will face a critically high risk of resource idleness. (ii) Dynamic intervention in luminous–thermal environments: Addressing the phenomenon where attractiveness plummets once illuminance exceeds 600 lx, window-adjacent areas should not rely solely on high light transmittance but must be equipped with dynamic shading facilities (e.g., blinds) to avert direct glare [44,45]. Simultaneously, given females’ better adaptability to slightly warmer environments, localized microclimate zoning can be implemented in HVAC design [46]. (iii) Dynamic–static zoning based on interference gradients: The area within a 4.0 m radius of the entrance/exit serves as a high-repulsion zone, while a “golden buffer zone” exists at a distance of 6.0 m to 8.0 m. It is recommended to eliminate fixed study seats proximal to entrances/exits, repurposing them into transitional lounge areas or retrieval zones. Concurrently, high-efficiency seat clusters catering to immersive learning needs should be centrally deployed in the “deep quiet zones” at the middle–rear of the space, away from main aisles, thereby maximizing spatial service efficiency. Ultimately, these data-driven insights empower architects and facility managers to transition from intuition-based layouts to precise, evidence-based spatial configurations, directly optimizing the operational efficiency and human-centric quality of contemporary educational buildings.

4.4. Limitations and Future Prospects

Although this study proposes an innovative research paradigm integrating multi-source sensing and explainable machine learning, certain limitations remain that explicitly pave the way for future investigations. First, regarding the temporal and seasonal scope, the current spatiotemporal dataset was collected during the winter transition period (November to December). As human thermal comfort thresholds, clothing insulation levels, and natural daylighting angles shift significantly across seasons, future studies must extend the analysis to include longitudinal data from the summer season to verify the seasonal adaptability of the findings. Second, regarding spatial typology, this research specifically recognized the specificity of its approach by focusing on a traditional, “formal” study room with a fixed matrix layout. Acknowledging that contemporary learning environments are increasingly embracing informality, flexibility, and collaborative peer interactions, future frameworks must be adapted to evaluate flexible seating, movable furniture, and the complex social dynamics within these informal spaces. Third, the geographical and cultural context of a single university in China may limit the generalizability of the findings, necessitating future cross-cultural comparative studies.

Future research can be expanded across the following dimensions: First, introducing multi-modal Internet of Things (IoT) sensor data to construct a comprehensive indoor environmental evaluation system encompassing light, thermal, acoustic, and air quality factors. Second, leveraging computer vision technology to delve into the social distances and “Group Dynamics” effects among users, thereby exploring the influence of sociological factors—such as seat-reserving behaviors, peer interactions in flexible spaces, and companion learning—on spatial selection. Third, conducting cross-seasonal (e.g., summer versus winter) longitudinal time series tracking to investigate the potential effects of long-term spatial experiences on individual learning performance. With the continuous iteration of artificial intelligence sensing and explainable technologies, refined behavioral predictions within the built environment will provide profound insights for creating healthier, more efficient, and more human-centric architectural spaces.

5. Conclusions

This study successfully quantified the non-linear driving mechanisms of the micro-scale physical environment and spatial facilities on study room seat selection by integrating computer vision, luminous–thermal simulation, and explainable machine learning (XGBoost/SHAP/GAM). The core conclusions are as follows:

(1)

Identification of non-linear comfort thresholds for key environmental factors: This study precisely defined the service boundaries of micro-scale spatial resources. For instance, the optimal desktop illuminance benefit range is 400–600 lx; the effective attraction radius of power sockets is concentrated between 1.5 m and 3.0 m, with the seat value experiencing a “precipitous” decline beyond 3.5 m; and a “golden buffer zone” that balances tranquility and convenience is formed at a distance of 6.0 m to 8.0 m from the entrance/exit.

(2)

Revelation of group heterogeneity in spatial behavioral preferences: High-efficiency learning groups exhibited a stringent “defensive” seat selection strategy, being extremely sensitive to path interference and highly dependent on core resources. A significant divergence in spatial depth perception was observed between male and female students (females preferred front–middle rows, while males preferred rear peripheral zones), corroborating gender differences in territorial perception at the micro-scale.

(3)

Verification of interactive compensatory mechanisms among multi-dimensional environmental factors: Seat selection is a multi-dimensional trade-off under complex constraints. During peak occupancy periods, users demonstrated evident “crowding compromise” behaviors. Simultaneously, high-quality natural daylighting and an open field of view could spatially and effectively “compensate” for the decline in attractiveness caused by the lack of power sockets.

Theoretically, this study constructed a micro-environmental–behavioral decision-making framework, providing quantitative evidence for classical theories such as “adaptive comfort” and “cognitive load.” Methodologically, it demonstrated an innovative research paradigm combining multi-source non-intrusive sensing with explainable artificial intelligence (XAI). The physical thresholds extracted in this research will provide direct data-driven support for precise facility deployment, dynamic luminous–thermal interventions, and scientific dynamic–static zoning in future learning spaces, ultimately facilitating the creation of healthier, more efficient, and human-centric built environments.