This study developed a novel two-stage classifier model that integ Support Vector Machine (SVM) and Recurrent Neural Network (RNN) with an attention-based dual-encoder architecture for analyzing student behavior sequence data in educational data mining. Unlike traditional RNN + SVM frameworks, our approach uniquely combines (1) a hierarchical feature fusion mechanism that merges base sequence encoding (capturing temporal dependencies via GRU) and attentiondriven encoding (highlighting performance-critical behaviors); (2) an adaptive attention module that dynamically weights behavioral sequences (e.g., prioritizing “library” access with >50% attention weight), enabling targeted focus on academicperformance-related patterns; (3) an end-to-end pipeline where RNN-extracted deep features are directly optimized for SVM classification, eliminating manual feature engineering. Through experimental validation, the model outperformed traditional methods in accuracy (86.9%) and recall (81.6%), particularly for long-sequence behavior data. Results confirm that increasing feature dimensions (optimal at 50 dimensions) enhances prediction capabilities but plateaus beyond thresholds. This framework provides a robust tool for actionable insights in educational policy-making. Future work will expand data diversity to strengthen practical applications.