Integrating Deep Learning And Explainable AI for Interpretable Educational Intervention Analysis

We present a modular framework that integrates deep learning and explainable causal inference for interpretable analysis of educational interventions. Using data from the NeurIPS Causal Education Challenge, we simulate a semi-synthetic treatment scenario on the Eedi platform and apply non-Gaussian ICA for latent representation learning. The causal structure is inferred using the ICALiNGAM algorithm, and treatment effects are estimated with the DoWhy library. To support interpretability, SHAP values are computed and integrated into the inferred causal graph. The framework accurately estimates average treatment effect (ATE), average treatment effect on the treated (ATT), and conditional average treatment effect (CATE), each with values close to 1.996. It also achieves strong predictive performance with classification accuracy of 86.9 percent, area under the ROC curve of 0.952, and a root mean square error of 0.301. Compared to traditional methods such as difference-in-differences, instrumental variables, and item response theory, our method combines reliable causal estimation with model interpretability. Even though it operates on latent features, SHAP value rankings remain structurally aligned. The framework supports stratified intervention analysis and offers a transparent, data-driven approach to evaluating the impact of instructional strategies in real-world learning environments.