Research On Session-based Recommendation Based On Cross-session Knowledge Graph

Session-Based Recommendation aims to learn users’ dynamic preferences based on anonymous user sessions(short sequences)to recommend the next item.Although it shares similar goals with recommendation methods that have been fully investigated for many years,major differences exist in problem definition and research ideas,and thus it is a current research hotspot and cutting-edge technology.Since session data exhibits data sparsity,noise and other characteristics,existing session-based recommendation methods only model user’s preferences based on the current session,ignore cross-session information and item knowledge,and cannot effectively capture the complex dependencies between items,resulting sub-optimal recommendation performance.In fact,the interaction data with other sessions,rich entity and relationship information in knowledge graphs and self-supervised signals exert great potential to fully capture the dependencies between items and alleviate the data sparsity problem.To this end,we investigate session-based recommendation on cross-session knowledge graphs from two main perspectives,one is to combine the side information of cross-session information and knowledge graphs,and the other comprehensively consider both subjective user interactions and objective item attributes.The following results are achieved:1.A knowledge-enhanced cross-session recommendation method is proposed.Specifically,firstly,the cross-session knowledge graph is constructed by effectively incorporating the cross-session information and the rich entity relationship information into the knowledge graph.Then,a knowledge-aware attention mechanism is designed to learn item node representations and model users’ long-term and short-term interests using gated recurrent unit and attention networks.Finally,the similar session reference circle is constructed for the target session by taking the similar session as a strong signal reflecting user preference to extract high-quality items for session-based recommendation.2.A cross-session recommendation method based on self-supervised graph cotraining is proposed.Specifically,two independent and complementary views,i.e.,crosssession graph and item attribute hypergraph,are firstly constructed from both subjective and objective perspectives to capture pairwise and beyond pairwise relations among items.Then,a query-aware attention mechanism and a hypergraph convolutional neural network are designed to learn the item and session embedding matrix.Finally,two contrast learning losses are designed to maximize the consistency between the two views and trained under a collaborative training scheme to improve the effect of session-based recommendation.Extensive experiments on multiple real-world recommendation datasets demonstrate that the proposed methods in this paper outperform the baseline methods,validating the effectiveness and reasonableness of the models.