An ensemble multi-modal cross-attention model for e-commerce customer satisfaction prediction

With the rapid growth of e-commerce, accurately predicting customer satisfaction has become critical for optimizing marketing strategies, reducing customer churn, and improving repurchase intention. Existing satisfaction prediction models predominantly rely on single-modality structured features and overlook the intrinsic interactions among multi-source heterogeneous signals—textual reviews, logistics time series, price and payment records, and geographical profiles—limiting their predictive performance; empirical results on the Olist dataset indicate that no individual modality matches the predictive performance of the full multimodal configuration. To address this, we propose the Multi-modal Cross-attention Network (MM-CrossNet), which treats these four signal types as serialized tokens and models their deep nonlinear interactions through a Transformer-style multi-head cross-attention mechanism, unlike prior approaches that rely on simple concatenation or independent encoders. A heterogeneous ensemble of LightGBM, XGBoost, and a deep MLP generates the final prediction. The framework targets post-delivery retrospective analysis, providing actionable satisfaction scores for proactive retention and marketing interventions. An empirical study on the Olist Brazilian E-commerce Public Dataset (96,358 samples, 23 features, two-year span) demonstrates competitive performance against nine mainstream baselines (including LR, DT, RF, XGBoost, LightGBM, DeepFM, and DIN), with MM-CrossNet achieving the highest Accuracy (0.8425) and F1-score (0.9059). The paired prediction difference relative to XGBoost is statistically significant according to McNemar’s test ( $$p<0.05$$ ), although the absolute performance margin is modest. Ablations show that the textual and logistics modules provide the largest contributions to predictive performance, while feature importance identifies delivery timeliness and product information quality as prominent predictive factors of customer satisfaction. This study offers a methodological reference for fine-grained e-commerce operations, with generalizability across platforms and languages left for future validation.