A Multimodal Explainable Deep Learning Approach for E-Commerce Recommendation and Behaviour Prediction

This paper introduces an interpretable deep learning model, which integrates the personalized recommendation and customer behaviour prediction of e- commerce platform together. The model combines multimodal data sources, such as clickstreams, transaction histories, product metadata, and textual reviews using a shared backbone consisting of CNN/RNN/Transformer encoders. A joint optimization is developed to address the trade off between recommendation accuracy and behavioural prediction, and an explainability module combines innate attention mechanism with post hoc attribution technique (SHAP, LIME). Experiments on Amazon, MovieLens and Taobao datasets show that the framework achieves strong performance (Precision @ 10 = 0.64, Recall10 = 0.58, NDCG${@} 10=070$, RMSE$=082$) while generating high fidelity explanations (fidelity$\approx 089$). Ablation studies show minimal gap between interpretability and performance, which demonstrates the feasibility of incorporating transparency into recommendation workflows. Case studies also verify that the generated user friendly explanations increase user trust and business utility, making the framework ideal for scalable, real world applications.