Social Media Signals and Consumer Purchase Prediction Using Multimodal Behavioural Analytics

Social-commerce prediction increasingly requires evaluation of behavioural interactions together with relational signals, yet the incremental value and robustness of such information remain uncertain. This study assessed whether user–friend social-network and user–friend–item (UFI) sharing/exposure signals add predictive information beyond behavioural history. Using Beibei for development and Beidian for replication, an offline implicit-feedback framework compared logistic models, Random Forest, XGBoost, LightGCN, and AMBF-Net, supported by TreeSHAP, modality ablation, heterogeneity, missing-modality, fusion, and ranking analyses. On Beibei, XGBoost outperforms with the highest PR-AUC (0.8871) and Random Forest achieves the highest NDCG@20 (0.9382). Social and UFI signals were relatively minor additional gains over behaviour, and were more sensitive in sparser-history groups; prediction was more sensitive to missing UFI information. Adaptive fusion did not result in consistent ranking benefits, although there was a small advantage at the pair level. Evaluation was based on sampled implicit feedback candidate sets, that is, sets of user–item pairs that were not observed because the user did not purchase the item, but were observed because the user did not reject it. Accordingly, findings are predictive and non-causal. The main contribution is a controlled multimodal evaluation framework for incremental value, robustness, heterogeneity, and cross-dataset assessment across heterogeneous data availability conditions rather than algorithmic superiority claims.