A heterogeneous couplings and persuasive user/item information model for Next Basket recommendation

Next Basket Recommender Systems (NBRS) are session-based recommenders that leverage a user’s sequential behaviour and context to predict the next items they could buy during the next shop visit. This field is popular now. Current research works rely primarily on item–user relations, ignoring complicated heterogeneous couplings and persuasive item/user information, causing sparsity and cold-start, stifling NBRS performance. This research implements a Heterogeneous Coupling and Persuasive User/Item Information Model (HCPIM) comprising of heterogeneous couplings and a persuasive embedding layer (HCPEL), the consecutive basket layer, and a basket prediction layer. The HCPEL learns heterogeneous couplings (user–user, user–item, item–item) and persuasive user information for every basket to model the composite intra-basket associations across the items inside a basket given a user’s behaviour sequences. The consecutive basket layer learns inter-basket associations from the HCPEL using a Gated Recurrent Unit network. In the basket prediction layer, an improved Particle Swarm Optimization (PSO) algorithm optimizes the network’s weight and bias to iteratively learn heterogeneous couplings and persuasive information inside and across baskets to recommend the next basket. In PSO, we incorporate the exploratory Gravitation Search Algorithm to balance exploration and exploitation. Our architecture includes an Adaptive Response to Particle Adjustment Strategy to increase exploitation and prevent particles from being trapped. Extensive tests are run to evaluate HCPIM on TaFeng and the IJCAI 2015 dataset. The HCPIM improves performance by 27.87, 29.5, and 27.29 percent when density is 50, 100, and 150 on the IJCAI-15 dataset, and 8.53, 10.57, and 4.85 percent on the TaFeng dataset.