MSMEs need modeling to predict user behavior over time as well as the structure of relationships between entities. The problem with this study is that LSTMs are effective in capturing sequential dynamics but weakening them in infrequent sequences, while GNNs excel at modeling structural relationships between entities but do not explicitly represent temporal evolution. The research contribution combines LSTM, and GNN to measure the impact of multi-task learning and calibration on rating accuracy and probability reliability on the Long prediction horizon. The method used is existing work that combines LSTM and GNN for product design and security-conscious MSME e-commerce. Especially in terms of risk calibration and long-term evaluation. Therefore, this study determines how multi-task training and calibration strategies affect rating accuracy (ROC-AUC, PR-AP, P@K/R@K) and probability reliability (Brier/ECE) over the extended horizon. The aim of this study is to design and evaluate the GNN-LSTM hybrid architecture to improve the accuracy of recommendations while reducing risk. The results of the experiment on the data of the partner MSMEs that were kept private, showed strong ratings and reliability: the purchase prediction reached ROC-AUC 0.965 (val) and 0.946 (test) with a PR-AP of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.943 / 0.910$</tex>, and session risk detection reached a ROC-AUC of 0.984 and a PR-AP of 0.982. This outperformed the CNN-BiLSTM baseline (0.93 val/0.91 ROC-AUC assay). Losses decrease steadily without negative transfers, and adequate calibration (Brier 0.161 for links, 0.176 for risk), indicating safe personalization with low false alarms.