Behavior Merging Graph Convolution Network for Multi-Behavior Recommendation

Encoding multi-behavior information into a single graph collaborative filtering vector is an emerging challenge, as different behaviors generate distinct graphs, each with its own embedding vector. To address this problem, recent approaches typically designate the embedding of some behaviors as primary embeddings and use the embeddings of other behaviors to enhance the primary behavior recommendation. However, these models may excel in recommending primary behaviors at the expense of degrading the performance of auxiliary behaviors. As a result, modern recommender systems often need to maintain multiple sets of collaborative filtering embeddings to achieve satisfactory recommendation performance across all behaviors. To alleviate this issue, we introduce the <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">B</u>ehavior <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">M</u>erging <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G</u>raphs. Instead of modeling each behavior separately, BMG uses a joint graph to capture potential behavior merging sets between nodes and applies the partial order theory to model the intricate structures and relational order among behavior merging sets. Based on BMG, we introduce the <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">B</u>ehavior <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">M</u>erging <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G</u>raphs <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</u>onvolutional <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">N</u>etworks (BMGCN), which aggregates neighbor information by integrating convolutional weights that account for the rank transformation of Behavior Merging Order across various behavior merging sets. Furthermore, BMGCN employs behavior merging-based sampling to guide the traditional BPR sampling process, enhancing embedding training. Experiments on three widely used datasets demonstrate that BMGCN achieves superior multi-behavior recommendation performance compared to state-of-the-art baselines.