To address the challenge of interpreting complex spatiotemporal patterns in multimodal English learning behaviors, this study proposes a novel computational framework integrating dynamic graph construction with adaptive interpretable machine learning. First, multi-source data fusion technology is used to integrate logs, data, and evaluation scores to construct a time-series learning behavior graph. Second, the XGBoost LSTM (long short-term memory) hybrid model is used to predict learning effectiveness, and the hierarchical sampling SHAP interpretation algorithm is proposed simultaneously. Third, a three-level warning mechanism is constructed using dynamic threshold sliding windows. Key behavioral characteristics are identified through SHAP values and a risk probability model is created using logistic regression. Validated across educational and industrial datasets, the framework enhances interpretability in predictive maintenance scenario, demonstrating cross-domain applicability in IoT sensor networks and healthcare monitoring systems.