Search Results (19,342)

In the context of educational digital transformation, automated grading of paper-based examinations faces two challenges: data privacy concerns and data heterogeneity across educational institutions. Existing Optical Mark Recognition systems are predominantly designed under centralized learning paradigms, requiring examination data to be collected and processed at a central server, which may expose sensitive student information. To address this limitation, this paper proposes a federated OMR framework for automated examination paper digitization and assessment in distributed environments. The proposed system consists of five stages, employing YOLO26 for answer-region localization, ORB-based image registration for geometric alignment, and EfficientNet-B0 for answer-state classification, automated grading, and result aggregation. Federated Averaging is utilized to train a global model across five clients representing heterogeneous data domains with different answer-sheet layouts, ink characteristics, and marking styles, without sharing raw data. Experimental results demonstrate that the federated model achieves an ROI-level accuracy of $99.42\%$ and a Macro F1-score of $95.82\%$. Furthermore, the proposed system attains a Sheet-level accuracy of $84.62\%$, outperforming local training approaches while achieving performance comparable to centralized learning. These findings suggest the feasibility of Federated Learning for developing privacy-preserving and highly generalizable OMR systems for large-scale educational assessment. Full article

(1) Background: Information diffusion models developed for graph-based platforms such as Reddit and broadcast architectures such as Telegram identify temporal features—particularly the timing of peak spread—as dominant predictors of coverage. Whether these predictors generalise to platforms where content is distributed through algorithmic recommendation rather than social-graph contagion remains an open question. (2) Methods: We analyse the YouNiverse dataset, comprising 133,364 English-language YouTube channels observed weekly from January 2015 to September 2019 (18.9 million observations). We derive channel-level diffusion features—including time-to-peak, post-peak decay rate, diffusion volatility, and upload frequency—and train three machine learning models (Linear Regression, Random Forest, and LightGBM) on two tasks: predicting peak weekly view growth (regression) and identifying viral channels (classification). A single-feature naive baseline (subscriber count alone) establishes the marginal contribution of the broader feature set beyond subscriber count alone, and a temporal split experiment (training on channels peaking before 2018, testing on 2018–2019) assesses cross-temporal stability. Because subscriber count and subscriber rank are measured at the October 2019 crawl, this is a retrospective characterisation rather than a strict real-time forecasting design. (3) Results: LightGBM achieves $R^2=0.776$ (5-fold CV: $0.778\pm 0.003$) compared with $R^2=0.548$ for the naive baseline, a net gain of $+0.228R^2$. Because subscriber rank and subscriber count are near-perfectly collinear, we interpret them jointly as a channel-size dimension (42.2% of total mean absolute SHAP attribution), rather than as independent effects. Time-to-peak ranks fourteenth (1.1%), in contrast to its dominant role on Reddit ($r=0.995$, rank #1). For virality classification, LightGBM achieves ROC-AUC $=0.967$. Under the temporal split, Random Forest ($R^2=0.703$) outperforms LightGBM ($R^2=0.683$), showing greater cross-temporal stability within this retrospective split. (4) Conclusions: Within the 2015–2019 data, the results are consistent with algorithmic recommendation weakening the relationship between temporal diffusion dynamics and coverage magnitude at the channel level. Time-to-peak is weakly informative in this setting, while generalisation to the current recommendation system requires validation on newer data. Full article

Deep-sea hydrothermal vents and cold seeps sustain highly productive animal communities through chemosynthetic symbioses, among which bathymodioline mussels are prominent examples. Bathymodioline gill bacteriocytes accommodate intracellular chemosynthetic symbionts, including sulfur- and/or methane-oxidizing bacteria depending on the host species, while remaining sheltered in an epithelium continuously exposed to environmental microorganisms, creating a fundamental immunological problem: how can an innate defense system remain effective without eliminating the microbial partners on which host nutrition depends? This review examines this problem through Bathymodiolus azoricus, integrating two decades of work on its cellular immunity, gill transcriptome, microbial challenge responses and symbiosis biology with recent mechanistic studies from related bathymodiolines. Central to the present synthesis are previously reported B. azoricus observations showing that gill tissue can mount local transcriptional responses to bacterial challenge, while hemolymph serum differentially modulates immune-gene expression following exposure to symbiont preparations or non-symbiotic Vibrio. Immune-gene expression also varies along the anterior–posterior gill axis, with lower expression in the posterior budding zone than in mature anterior filaments. We interpret this zonation primarily as a feature of tissue maturation rather than demonstrated active immune suppression, consistent with evidence that newly formed filaments are initially aposymbiotic and become colonized only after formation. Together, these observations evoke a host–symbiont–pathogen triad in which local gill-tissue responses, systemic humoral modulation and gill development constitute interacting levels of immune organization and compartmentalization. As a working hypothesis, we propose that this triad is reconciled principally through spatial and developmental compartmentalization of immune competence rather than through generalized immune suppression, predicting that immune-gene expression should track gill maturation state rather than symbiont occupancy per se. We consider this tissue-level model alongside comparative evidence for putative symbiont-uptake mechanisms, post-engulfment microbial discrimination, lysosomal regulation, symbiont digestion and bacteriocyte turnover, including the mTORC1-dependent phagosome-digestion checkpoint demonstrated in Bathymodiolus japonicus. Rather than assuming that these mechanisms are conserved across species, we distinguish explicitly between findings established in B. azoricus, evidence from other bathymodiolines and canonical pathways used as mechanistic context. We conclude by identifying unresolved components of B. azoricus immunity, including the prophenoloxidase system, the broader antimicrobial-peptide repertoire and the relationship between cellular checkpoints and tissue-level gill zonation, and consider the prospective biotechnological relevance of mechanisms that tolerate persistent microbial symbiosis without loss of immune vigilance. Full article

Background and Objectives: The rehabilitation of the atrophic posterior maxilla with pterygoid implant presents a significant challenge due to complex regional anatomy and limited visual access. This study aimed to compare the accuracy of a dynamic computer-assisted implant surgery system (dCAIS) for pterygoid implant placement in fully and partially edentulous maxillae using radiographic marker registration (RMR) and markerless tracing registration (MTR), respectively. Materials and Methods: Forty pterygoid implants were retrospectively evaluated in 40 patients, divided into two groups: fully edentulous (n = 20) and partially edentulous (n = 20). Implant placement was performed using a dynamic navigation system (dCAIS), utilizing either bone-anchored mini-screws or tooth-surface tracing as reference points. Accuracy was assessed by superimposing preoperative plans with postoperative cone beam computer tomography (CBCT) scans. Coronal, apical, and angular deviations were measured and analyzed using the Data Science Workbench (version 14). The level of statistical significance was set at α = 0.05 (two-tailed). Results: The mean deviations for fully and partially edentulous maxillae measured 1.24 ± 0.27 mm and 1.40 ± 0.27 mm at the coronal level (p = 0.065), 1.30 ± 0.49 mm and 1.30 ± 0.54 at the apical level (p = 0.995), and 0.72 ± 0.38° and 0.82 ± 0.43° in angular deviation (p = 0.560), respectively. No statistically significant differences were found between the two groups. Notably, mean angular deviation was remarkably low (<1°) in both groups. Conclusions: No statistically significant between-group differences in accuracy were detected in coronal, apical, and angular deviations between the markerless tracing registration method in partially edentulous patients and the bone-anchored radiographic marker registration method in fully edentulous patients. Full article

Over the last decade, the proliferation of electric vehicles (EVs) has highlighted the importance of robust battery management systems (BMSs) to mitigate cell imbalance driven by manufacturing tolerances, thermal gradients, and non-uniform aging. To address these limitations, this study presents a MATLAB R2023b/Simulink-based comparative performance analysis of passive and active cell balancing topologies for lithium-ion battery packs. Using an equivalent circuit model based on the ORION 18650/26 cell, twelve distinct configurations encompassing passive switched-resistor alongside active inductor, capacitor, transformer, and converter topologies were evaluated. To isolate intrinsic charge-transfer dynamics from multi-cell network latency, all topologies were benchmarked in a standardized adjacent two-cell baseline under a strict convergence threshold (ΔOCV ≤ 1 mV). The simulation results demonstrate that parallel two-inductor and buck–boost topologies achieve the fastest equalization speed (≈1.47–2.53 s), whereas switched-capacitor configurations yield the lowest total energy dissipation (≈0.0011 Wh–0.0013 Wh). Furthermore, to evaluate string-level scalability and multi-hop energy transfer dynamics, the high-performing buck–boost topology was extended and benchmarked in a four-cell series (4S) configuration. The simulation results demonstrate that while the adjacent two-cell baseline achieves fast equalization (≈1.47–2.53 s), the 4S string reaches multi-cell convergence within 12.47–13.94 s, providing quantitative insights into multi-hop routing latency. Overall, this work provides an unconfounded quantitative baseline to support BMS engineers in selecting optimal balancing topologies tailored to specific EV performance, space, and economic constraints. Full article

Concentrated sourcing is a structural property of the world economy rather than an occasional accident: across 168 national economies and 1118 four-digit product markets reconstructed from harmonized cross-border flow records, 27.7% of macro-level economy–product supply systems draw more than half of their imports from a single origin and 19.3% are critically dependent. Treating each such market as a system rather than as a set of bilateral links changes what can be asked of it, and this paper specifies the economy–product supply system in systems-engineering terms—boundary, elements, internal relations, external environment, state and state transition—and represents it as a time-evolving hyperedge over source countries. Three coupled questions follow, answered jointly by HyperSRM: which dependency state a system will occupy next year, whether it will diversify, reconcentrate, hold, or merely substitute one origin for another, and which origins are most consistent with the observed conditions preceding a material entry. Shared country and product embeddings support two temporal set-encoding branches, a candidate-conditioned branch for origin ranking and a candidate-free branch for state and mode forecasting, a sign-constrained gravity–capability–connectivity prior supplies an observational plausibility score with end use and maritime reachability as its context, and risk weights derived from the state head direct effort toward the most exposed systems. Developed on CEPII BACI, rebuilt independently on Eurostat Comext and audited against U.S. Census data at the level of the labels themselves, the framework returns calibrated state probabilities and a ten-origin shortlist that captures 58.1% of the following year’s risk-weighted material-entry mass and is accompanied by explicit out-of-pool diagnostics. These outputs describe the import-sourcing layer of resilience and are intended for analytical triage rather than a complete assessment of supply resilience. Three system-level regularities carry beyond the model: the arrival of a new origin is a weak proxy for diversification, critical dependency is close to absorbing for specified intermediate inputs but not for final goods, and the 2020–2021 contraction rearranged source sets without widening them—so resilience monitoring built on source counts misreads the direction of change. Full article

Early in ovo sex identification is important for improving hatchery resource utilization; however, sex-related optical differences are weak during early embryonic development and exhibit clear stage dependence. In this study, longitudinal visible–near-infrared transmission spectra were acquired from 1600 fertilized pink-shelled eggs of Bashang Long-tailed chickens across four independent incubation batches during incubation days 1–7. Before model development, 200 eggs were reserved as a fixed internal hold-out test set, while the remaining 1400 eggs were used for exploratory incubation-day and feature screening, five-fold cross-validation, and model comparison. Among all 21 dual-day combinations, D2 + D5 achieved the highest exploratory validation AUC of 0.951. The selected scheme required separate measurements on D2 and D5, with the final sex classification completed on D5. Further incorporation of the signed difference spectrum ΔS and relative change rate R increased the AUC to 0.965, demonstrating the incremental discriminative value of longitudinal developmental-change information. After the input scheme had been fixed, DSSF-Net achieved the numerically highest performance among the evaluated models under unified egg-level stratified five-fold cross-validation, with an accuracy of 0.915 ± 0.015, an AUC of 0.974 ± 0.012, and an F1-score of 0.915 ± 0.014. The final model achieved an accuracy of 0.915 and an empirical AUC of 0.966 (95% CI: 0.940–0.987) on the fixed internal hold-out test set. Complete four-round leave-one-batch-out validation yielded mean accuracy and AUC values of 0.846 ± 0.014 and 0.926 ± 0.012, respectively, indicating that the longitudinal spectral information from D2 + D5 maintained relatively stable discriminative performance across incubation batches within the investigated population. These results demonstrate the feasibility of longitudinal spectral modeling for early sex identification in fertilized pink-shelled eggs of Bashang Long-tailed chickens. Systematic offline validation was completed under laboratory conditions, including evaluation on the fixed internal hold-out test set and four-round leave-one-batch-out validation, supporting the effectiveness and batch-level stability of D2 + D5 longitudinal developmental spectra. Post hoc wavelength-level interpretation further identified 594–620 nm and 660–675 nm as two candidate spectral regions with relatively high predictive value. Collectively, these findings provide a reliable longitudinal spectral modeling framework for early non-destructive in ovo sex identification and establish an experimental and methodological foundation for wavelength selection, device development, and subsequent engineering validation of multispectral detection systems under hatchery conditions. Full article

The increasing adoption of Retrieval-Augmented Generation (RAG) in the oil and gas industry has created a growing need for systematic evaluation of domain-specific knowledge bases, particularly with respect to retrieval failures, numerical and entity inconsistencies, knowledge timeliness, and unsupported generation. This study presents a domain-adapted evaluation and optimization framework for industrial RAG knowledge bases based on an LLM-as-a-Judge paradigm. The framework organizes evaluation into four dimensions—Data, Retrieval, Generation, and Utility (DAAE)—and combines deterministic metrics with LLM-based semantic assessment. A two-tier evaluation procedure combines retrieval-based screening with fine-grained LLM judging while retaining retrieval failures in end-to-end evaluation statistics. Rather than introducing new retrieval or generation algorithms, the framework integrates established RAG techniques with evaluation criteria motivated by oil-and-gas knowledge characteristics, including domain-entity and numerical consistency, temporal validity, chunk-level semantic integrity, and controlled abstention. Evaluation results are mapped to corresponding optimization actions across the data, retrieval, and generation layers, including semantic-aware chunking, metadata augmentation, hybrid sparse–dense retrieval, Cross-Encoder reranking, and structured evidence-grounded prompting. The framework was evaluated using the Intelligent Knowledge Base for Natural Gas Economic Research and an expert-annotated benchmark comprising 150 domain questions. In the industrial before–after comparison, Retrieval Hit@5 increased from 58.0% (87/150) to 89.3% (134/150), Context Precision increased from 0.64 to 0.88, and Faithfulness increased from 0.65 to 0.94. These values are reported as system-level point estimates rather than as component-wise causal effects. The results demonstrate the practical value of evaluation-guided optimization for improving the reliability of domain-specific RAG systems in natural-gas economic research and provide an industrial case for systematic RAG assessment and iterative optimization in the energy sector. Full article

We develop a PDE-constrained optimization framework for calibrating a regime-switching Heston–Merton model to S&P 500 index option prices. The model features two latent Markov regimes modulating stochastic volatility parameters and compound Poisson jumps, capturing the stylized fact that market volatility clusters differently during normal and crisis periods. Using real data from the Federal Reserve Economic Data (FRED) database covering July 2016 to July 2026 (2609 business days), we identify crisis regimes via VIX thresholds and estimate transition probabilities. Our empirical analysis reveals that crisis regimes exhibit 3.78 times higher long-run variance, 1.60 times higher vol-of-vol, and 113 times higher jump intensity compared to normal regimes. We derive the full adjoint system for the forward PIDE, including the previously undocumented jump operator adjoint and Markov-switching generator adjoint, and demonstrate that the adjoint method reduces per-iteration PDE solves from order-P to 2 regardless of parameter dimensionality. A panel calibration exercise demonstrates superior in-sample fit (${\mathrm{RMSE}}_{\mathrm{IV}}=1.24$ vol points) versus the nested Heston (2.87), Bates (2.31), and Black–Scholes (19.46) models. Out-of-sample Diebold–Mariano tests confirm statistically significant forecasting gains at the 1% level. The Feller condition is satisfied in both regimes. Full article

Time-fractional generalizations of the Lindblad master equation describe open quantum systems whose coupling to the environment exhibits power-law memory. We develop the optimal-control theory of such systems and analyse the consistency of the adjoint calculus on which every gradient-based pulse-design method relies. Casting the density operator in a fractional Bochner–Sobolev space of Hilbert–Schmidt operator valued functions, we establish well-posedness through Mittag–Leffler resolvent families, prove that the completely positive trace-preserving (CPTP) structure is preserved for the controlled, time-dependent generator without recourse to subordination, and obtain existence and uniqueness of optimal controls by the direct method. The adjoint is governed by the right Riemann–Liouville derivative with a fractional-integral terminal condition, a structure established for Caputo dynamics with a Mayer cost by Bergounioux and Bourdin, who also showed that a pointwise terminal costate cannot exist. Our central result concerns the discrete counterpart of that fact, where existence is never lost: imposed on the right-Caputo adjoint of a convergent scheme, the pointwise condition yields a bounded costate and a well-defined reduced gradient carrying an error that is mesh-independent. We further establish a sharp rate in the classical limit: the defect vanishes exactly linearly, $\parallel \Delta \left(\beta \right)\parallel =C(1-\beta )+O\left({(1-\beta )}^2\right)$, with C given in closed form through a digamma series. Two consequences follow: monotonicity of the defect in the memory order is proved near $\beta =1$, and the memory order is locally identifiable from gradient data alone. A corrected adjoint restores consistency with proven convergence rates. Numerical experiments on two-level, three-level and two-qubit open systems (Liouville dimension up to 16) confirm the mesh-independence, reproduce C to three significant digits, and recover the full rate on graded meshes. Full article

Background: Exercise-induced left ventricular hypertrophy (LVH) has been extensively investigated in adolescent athletes using the conventional two-tiered classification (2TC). However, the four-tiered classification (4TC) allows further differentiation of LVH patterns by incorporating three-dimensional information on left ventricular (LV) geometry. This study aimed to compare LVH assessment using the two classification systems and characterize the reclassification patterns. Methods: A total of 121 adolescent athletes (mean age: 15.1 ± 1.6 years) and 114 adult athletes (mean age: 22.9 ± 3.7 years), all competing at the national level, underwent comprehensive echocardiographic and anthropometric evaluations. Results: The application of 4TC resulted in redistribution across LV geometric categories compared with conventional 2TC. Twenty-three (19.0%) adolescent and 25 (22%) adult athletes were reclassified, including seven (5.8%) and 17 (15%), respectively, who shifted from normal geometry under the 2TC to different LVH categories under the 4TC. Reclassification was observed across sex and sporting discipline subgroups, although the patterns varied. In runners and triathletes, the combined prevalence of eccentric non-dilated and eccentric dilated LVH was 22.7% using the 4TC versus 4.5% eccentric LVH under the 2TC classification. LVM was strongly associated with the combined contribution of cumulative training duration, lean body mass, and body surface area (r = 0.785, p < 0.001), whereas training duration alone was not significantly associated with the LVH parameters. Conclusions: The 2TC and 4TC approaches result in different assessments of LV geometry in highly trained athletes, with the 4TC providing additional characterization of LVH patterns by considering the presence or absence of LV dilatation. Reclassification patterns varied according to sex and sporting discipline, highlighting the importance of sport-related characteristics when interpreting exercise-induced LV remodeling in athletes. Full article

The Hyperspectral Infrared Atmospheric Sounder II (HIRAS-II) onboard the Fengyun-3 satellites is a Fourier-transform infrared spectrometer that requires high radiometric calibration accuracy, making the characterization and correction of polarization effects essential. Although the gold-coated scan mirror introduces only weak polarization, its rotation changes the polarization orientation relative to the fixed polarization-sensitive axis of the downstream optics, producing scan-angle-dependent radiometric errors. To characterize and correct this effect, we designed and constructed a dedicated polarization test apparatus and used it to conduct a prelaunch thermal-vacuum (TVAC) polarization test. During the test, HIRAS-II observed the same stable 290 K area-source blackbody over a densely sampled scan-angle range, with the internal calibration target and cold shield serving as the warm and cold references, respectively. Guided by a polarization-induced radiometric error model formulated within the two-point calibration framework, we developed a decoupled two-step least-squares method to retrieve the polarization parameters from the resulting measurements. The method first estimates the equivalent polarization-axis angle of the downstream optical system from the phase of the band-averaged angular modulation and then retrieves the effective combined polarization parameter separately for each field of view (FOV) and spectral channel. The retrieved parameters were subsequently used to calculate the scan-angle-dependent polarization correction term and correct the calibrated spectra. After correction, the FOV-averaged standard deviation over the scan angle decreased from 0.023 to 0.007 K, from 0.024 to 0.007 K, and from 0.045 to 0.014 K in the long-wave (LW), mid-wave 1 (MW1), and mid-wave 2 (MW2) bands, respectively. The corresponding maximum reductions in brightness temperature deviation were 0.093, 0.064, and 0.174 K. The model, experimental approach, and retrieved prelaunch parameters establish a basis for the on-orbit evaluation and correction of scan-angle-dependent polarization-induced radiometric errors. Reducing these errors improves the radiometric calibration accuracy of HIRAS-II and helps provide more reliable Level-1 radiance data for atmospheric profile retrievals and data assimilation in global numerical weather prediction (NWP) systems. Full article

This paper presents an approach for optimal operation of self-healing networked microgrids (NMGs) under both normal operation and emergency conditions using the pufferfish optimization algorithm (POA). The proposed methodology is based on an energy management system (EMS) with two levels and independent functions. The lower-level is designed for normal operation, where the local controller of each microgrid (MG) performs the optimal dispatch of power from the dispatchable sources. During an emergency case in any MG, the higher-level EMS is activated, and the global controller is brought into operation. Physically, the NMGs are connected by tie-lines, while cyber links are established to exchange information and control signals for coordinated operation. Each MG operates to supply its local demand during normal operation conditions, resulting in no electrical power exchange between MGs. İn case of generation deficiency or a fault leading to generation outage, electrical power can be exchanged through the existing interconnections, enabling the affected microgrid to receive support from neighboring MGs. The main objective of POA is to minimize the total operating cost, in which the economic impact of network power losses is incorporated into the single objective function. Simulation studies were conducted using MATLAB and DIgSILENT software over one day. The performance of POA was compared with Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Grey Wolf Optimizer (GWO) under the same computational settings. Statistical and convergence analyses show that POA achieves the lowest mean operating cost across all studied cases, with low run-to-run variability and favorable convergence behavior. The results demonstrate the effectiveness of the proposed approach in improving the economic operation of NMGs under both normal and emergency conditions. Full article

Airborne dust plumes are difficult to localize in underground mine-face video because the scene is dark, illumination moves with machinery, and dust is confused with lamp bloom, reflective steel, and water-spray aerosol. We present DustVeil, a label-free two-stage system for image-space plume localization. Its software teacher combines background-referenced veiling (C1), local texture decay (C2), and absolute dark-channel response (C3) with a probabilistic soft-OR, then applies glare and chroma gates. The teacher returns a dimensionless response map in [0, 1], a binary plume mask and the corresponding image-area ratio; it does not estimate dust concentration, particle-size distribution, respirable exposure, or hazard categories. Teacher outputs from 342 frames in 114 clips/24 sessions supervise a 0.47 M parameter TinyU-Net. Evaluation uses a 144-image synthetic calibration set and a 72-frame real test set drawn from 72 clips in 18 sessions, with all roles separated at clip and session levels. Thresholds are selected only on synthetic masks and frozen before real scoring. After replacing per-image score normalization with fixed baseline-normal calibration and using reference implementations of the anomaly methods, DustVeil obtains IoU/F1 of 0.366/0.500 and the lowest clean-frame false-positive area (3.7% versus 13.9–59.9%). A separate water-spray set quantifies visual specificity. TinyU-Net runs at 610 FPS for network-only inference and 233 FPS aggregate in the measured six-stream decode-to-mask pipeline; optical flow is excluded from these figures. The validated scope is six fixed visible-light RGB cameras with camera-specific unlabelled calibration at one site, rather than concentration monitoring or camera-disjoint deployment. Full article

Digital access occupies a prominent place in educational policy debates, yet its predictive contribution to mathematics achievement, relative to psychological and socioeconomic factors, remains insufficiently quantified. Existing evidence often relies on linear models, examines predictor blocks separately, or considers technological resources without placing them alongside psychological constructs within a common framework. This study compares the predictive importance of psychological, socioeconomic, demographic and ICT-access indicators for mathematics achievement in PISA 2022. The analysis used data from 141,563 students across 19 education systems and included seven predictors: mathematics self-efficacy, mathematics anxiety, sense of school belonging, economic, social and cultural status (ESCS), gender, ICT resources at home and ICT resources at school. Weighted gradient boosting models were fitted separately to each of the ten plausible mathematics values and interpreted using TreeSHAP; a weighted random forest with permutation importance was used as a robustness check. The full model explained 39.2% of the weighted test-set variance in the plausible-value outcomes (R² = 0.3922, ${SE}_{total}=0.0066$, and RMSE = 77.95 score points). Mathematics self-efficacy ranked first under both criteria (42.3% of SHAP importance; 57.8% of permutation importance), ahead of socioeconomic status (30.9%; 32.0%), while the ICT-access block contributed 6.9% and 2.1%, respectively and added 0.0122 to test-set R². The two importance rankings were identical (Spearman’s ρ = 1.000). The SHAP ranking was unchanged across all 800 plausible-value × replicate-weight runs, matched XGBoost permutation importance, and was reproduced under a school-grouped train–test split. The model also detected a non-monotonic association between school belonging and predicted achievement, together with a MATHEFF × ESCS interaction pattern, supported by a direct-outcome interaction model, in which the modelled association between self-efficacy and achievement was stronger at higher ESCS levels. Because PISA 2022 is cross-sectional and plausible values are designed for population-level inference, these findings should be interpreted as predictive and associational rather than causal. The results suggest that, in systems where ICT access is already widespread, reported access to technological resources contributes comparatively little to prediction once psychological and socioeconomic indicators are considered, although sensitivity analyses indicate that home ICT access is partly affected by missingness patterns. Full article