MSMEs need a recommendation system that simultaneously captures the evolution of user intent over time and the relationship structure between entities (users, products, sessions, categories, and security events). The problem with this research is that LSTM excels in sequence, but its performance drops on a rare timeline, a general situation in MSME logs. GNN is strong for cross-entity relationships, but it does not explicitly model temporal dynamics. The gap arises because many pipelines still separate design signals, temporal behavior, and session security, reducing explainability and long-term reliability. Our contribution proposes a calibrated and security-aware hybrid that integrates CNN, heterogeneous GNN with reverse edges, and LSTM for behavioral sequences. Multitask-trained models (BCE for purchase links and λ· BCE for session risk) with L2 regularization and post-practice calibration, chronological data sharing prevents leakage. The goal is to design and evaluate CNN-plus CNN-enhanced GNN-LSTM hybrids to improve the accuracy of recommendations and reduce risk. The results on partner MSME data: ROC-AUC 0.965 (val)/0.946 (test), PR-AP 0.943/0.910; risk ROC-AUC 0.984, PR-AP 0.982, surpassing a CNN-BiLSTM baseline (0.93/0.91). Brier scores 0.161 (links) and 0.176 (risk) enable safer personalization. Going forward, we are focusing on per-segment calibration with ECE/MCE reporting, compute efficiency, multimodal expansion, ablation, and explainability (GNNExplainer, CNN saliency), as well as online retraining and drift monitoring to maintain production performance.
Kusanti et al. (Thu,) studied this question.