FedX-Edu: Explainable Academic Performance Prediction with Hybrid LSTM-Transformer Fusion in Federated Learning

TLDR

A new paradigm for explainable AI in educational prediction systems is established by developing a hybrid LSTM-Transformer architecture that enhances interpretability through feature importance tracing and decision visualization, while preserving data privacy via distributed computation.

Abstract

This study proposes a novel academic performance prediction model integrating temporal features and attention mechanisms within a federated learning framework. By developing a hybrid LSTM-Transformer architecture, we achieve dynamic fusion of sequential patterns and global dependencies from multimodal learning behaviors. The model enhances interpretability through feature importance tracing and decision visualization, while preserving data privacy via distributed computation. Experimental results demonstrate 15.7% RMSE improvement over baseline models, with visualized knowledge graph analysis revealing actionable pedagogical insights. This work establishes a new paradigm for explainable AI in educational prediction systems.