Facing the problems of lagging identification of in-session transactions, sparse positive class samples and dispersed evidence from multiple sources on e-commerce platforms, this paper constructs a multi-source data-driven prediction model of consumers’ purchase intention, MSF-PIN.The study takes session as the basic object, and uniformly organizes behavioral logs, product attributes, evaluative sentiment, price promotions, and access contexts into 162,840 session-level samples, among which 24,912 purchase samples account for 15.30% of positive class. Among them, there are 24,912 purchase samples, accounting for 15.30% of positive categories. At the model level, behavioral sequence coding, cross-source gating fusion, and click-add-purchase multi-task learning are introduced to improve the ability to portray the strength of short-term interest paths and heterogeneous evidence. Experiments are conducted using temporal order slicing and deployment-oriented evaluation protocols. The results show that MSF-PIN has an AUC of 0.902, a PR-AUC of 0.523, an F1 of 0.587, a Logloss of 0.140, and an ECE of 0.032, which is overall better than the comparative models of LightGBM, CatBoost, DeepFM, AutoInt, DCN V2, and BST. Results obtained by checking each scenario one by one indicate that the model obtains an AUC of 0.913 in the scenario of returning visitors and a PR-AUC of 0.541 in the scenario of promotion. The deployment analysis further gives out that when we screen the top 10% high-intention sessions through predictive probability, the model can cover 48.9% of actual purchase samples, and the average inference delay is 34.9 ms for each 1000 sessions. The outcome illuminates that sample-level weight redistribution using many kinds of proof can promote the utilization degree of purchase intention forecast, thus give a directly gotten probability foundation for recommendation weight adjustment, coupon start, and customer service help.
