Personalized recommendation methods for e-commerce based on multi-source user behavior data fusion
Аннотация
Insufficient modeling of the heterogeneity and interdependencies within multi-source user behavior data limits the ability of e-commerce recommendation systems to fully capture user interests, thereby affecting recommendation stability and accuracy. To address this issue, this paper proposes a personalized recommendation method based on multi-source user behavior data fusion, aiming to enhance the completeness and consistency of user interest representation. The proposed framework performs unified and standardized modeling of diverse behavioral signals, including browsing, clicking, adding to cart, and purchasing. Under temporal semantic constraints, differentiated feature representations are constructed to capture both preference intensity and temporal dynamics. A collaborative fusion mechanism is introduced within a shared representation space to jointly model multi-source behavioral information, generating user interest embeddings that are structurally consistent and informationally complementary. Based on these fused representations, accurate user–item matching is achieved for personalized recommendation. Experimental results on the publicly available Amazon multi-behavior dataset demonstrate the effectiveness of the proposed approach within the adopted evaluation setting, achieving a Precision@10 of 0.412, Recall@10 of 0.356, and NDCG@10 of 0.437. For inactive user subsets, the NDCG@10 reaches 0.398, indicating that the proposed approach retains its recommendation performance under the evaluated sparse interaction conditions. These findings indicate that the proposed multi-source behavior fusion method improves user interest representation within the evaluated Amazon multi-behavior dataset, while the applicability of the learned representations to other e-commerce platforms and behavioral distributions requires further empirical validation.