Analysis of Motivating Factors for University Students' Study Abroad Decisions and Research on a Multimodal Classification Prediction Model

Abstract

: Addressing the challenges of multidimensional factors influencing university students' study abroad decisions and the extreme imbalance in sample distribution, this study constructs a predictive framework integrating feature decoupling, data balancing, and multi-algorithm comparison. The study first performs refined preprocessing on 1,645 multidimensional student feature data points, identifying core positive-correlated drivers of study abroad such as annual household income, family history of overseas residence, and CET-6 scores. To address the severe imbalance where negative samples vastly outnumber positive ones, SMOTE oversampling is introduced to reconstruct the feature space, significantly enhancing the model's accuracy in capturing positive study abroad samples. By integrating K-Means clustering with t-SNE dimensionality reduction, the study revealed the latent structural distribution within the student population. At the algorithmic level, it compared the efficacy of logistic regression, support vector machines, and random forest classifiers. Experiments demonstrated that random forest exhibits optimal generalization capabilities for handling such nonlinear complex relationships, achieving prediction accuracies of 0.8731 for intention prediction and 0.9042 for behavior prediction. This study provides a scientific quantitative basis for international talent cultivation and targeted resource allocation in higher education institutions.