Sequential Recommendation (SR) aims to predict the most likely coming user-item interaction based on their interaction history. Since bipartite graphs can easily represent SR records, GCN-based models proved to be very effective in SR. However, 1) directly importing high-order sequential patterns (SP) would improve performance but bring a huge complexity burden; 2) current GCNs are to learn nodes’ interaction characteristics in SR by transporting unary message carriers between adjacent nodes, which neglects that users and items are different. Thus, we proposed an Edge-view Dual message carriers GCN for sequential Recommendation (EDG4Rec), which regards users and items as dual heterogeneous message carriers delivering their influences on SR bipartite graph independently. Specifically, an interaction and its user’s 1-order SP and its item’s 1-order SP are joined directly as a basic domain, where the dual message carriers transported by our proposed DMT mechanism with three variants, i.e., CD-DMT, JSP-DMT, CI-DMT. Finally, an edge-based negative sampling mechanism (E-NSM) is proposed to make training faster. Extensive experiments on dense and sparse real-world datasets demonstrate the superior performance of our model. It shows that supplementary information is no longer necessary for our design. Besides, our model can work well in cold-start situations.

Utilizing user-item interaction dynamics is crucial for providing efficient and precise sequential recommendations (SR). Besides earlier works based on Matrix factorization, Markov Chains, and RNNs, GCNs taking advantage of attention mechanism and Knowledge-enhanced NNs digging heterogenous auxiliary information recently made improvements. However, there still exist some problems: 1) existing works mainly focus on capturing user preference without considering the importance of item popularity; 2) introducing too much auxiliary information denies the independence of interaction behavior; 3) favoring high-order chronological history with heavy memory while ranking score referring different bias is unappreciated. To tackle these problems, we proposed DraG4Rec (Dual rating enhanced attention-based GCN for Recommendation), which makes symmetry of user preference and item popularity manifest great value. Specifically, we first merge ranking score, time, and position encoding for user/item representation in a bipartite graph. Then, an edge-view message-passing mechanism is imported into the attention-based GCN to use implicit 2-order historical information with less memory. Finally, user preference and item popularity are learned jointly. Extensive experiments on dense and sparse real-world datasets demonstrate the superior performance of our framework over state-of-the-art baselines regarding two commonly-used metrics: i.e., Recall and NDCG. The ablation study illustrates the boosting effect of different components.