Explainable machine learning for sustainable education: Predicting college students' reliance on generative artificial intelligence

As generative artificial intelligence (GenAI) becomes increasingly embedded in education, students are using it to support knowledge acquisition, explore solutions, and stimulate creative thinking. At the same time, its convenience has raised concerns about overreliance. To assess and manage such dependence so that GenAI can be used sustainably in education, this study is based on questionnaire data collected from students at multiple universities in China. We first applied principal component analysis and numerical binning to the target variable, and then trained six machine learning models to predict levels of GenAI dependence. Performance comparisons showed that the Random Forest (RF) model performed best, with an F1-score of 0.836. We further interpreted the results of the RF model using SHAP and PDP methods. Classroom speaking pressure emerged as the key predictor of GenAI dependence, accounting for 22.9% of the explanatory power. Higher speaking pressure is associated with higher dependence, and this effect is markedly stronger in the high-dependency group than in the low-dependency group. Furthermore, this study conducted ablation studies and multi-dimensional sensitivity analyses, fully validating the robustness and reliability of the preprocessing strategies. This study not only proposes an effective, explainable approach to predicting GenAI dependence, but also offers empirical evidence to guide students toward more responsible use of GenAI. By identifying critical risk factors and informing targeted interventions, the findings help balance technological support with autonomous learning in AI-enhanced education, thereby advancing inclusive and sustainable educational practices in line with the "quality education" agenda of United Nations Sustainable Development Goal 4 (SDG 4).