SWGCN: Synergy Weighted Graph Convolutional Network for Multi-Behavior Recommendation

TL;DR

The paper addresses sparsity and ambiguous signal in multi-behavior recommendation by proposing SWGCN, a graph-based model that jointly learns fine-grained intra-behavior edge weights and cross-behavior synergy. It introduces the Target Preference Weigher to quantify action-specific interaction strength and the Synergy Alignment Task to align target and auxiliary behavior preferences, optimized via a joint L$= \

Abstract

Multi-behavior recommendation paradigms have emerged to capture diverse user activities, forecasting primary conversions (e.g., purchases) by leveraging secondary signals like browsing history. However, current graph-based methods often overlook cross-behavioral synergistic signals and fine-grained intensity of individual actions. Motivated by the need to overcome these shortcomings, we introduce Synergy Weighted Graph Convolutional Network (SWGCN). SWGCN introduces two novel components: a Target Preference Weigher, which adaptively assigns weights to user-item interactions within each behavior, and a Synergy Alignment Task, which guides its training by leveraging an Auxiliary Preference Valuator. This task prioritizes interactions from synergistic signals that more accurately reflect user preferences. The performance of our model is rigorously evaluated through comprehensive tests on three open-source datasets, specifically Taobao, IJCAI, and Beibei. On the Taobao dataset, SWGCN yields relative gains of 112.49% and 156.36% in terms of Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG), respectively. It also yields consistent gains on IJCAI and Beibei, confirming its robustness and generalizability across various datasets. Our implementation is open-sourced and can be accessed via https://github.com/FangdChen/SWGCN.

Figures (8)

• Figure Fig. 1: A case of the synergistic relationships between the target behavior and auxiliary behaviors.
• Figure Fig. 2: The overall framework of SWGCN. (1) Input the user-item bipartite graphs of auxiliary and target behaviors. (2) Calculate interaction weights for each behavior using the Target Preference Weigher. The edge weights $w_{u_1,i_1;r},w_{u_1,i_2;r},w_{u_1,i_3;r},w_{u_1,i_4;r}$ represent the interaction strengths between user $u_1$ and items $i_1,i_2,i_3,i_4$ under $r$-th behavior, respectively. Likewise, $w_{u_1,i_1;R}, w_{u_1,i_2;R},$ and $w_{u_1,i_3;R}$ serve to quantify the engagement intensities associated with the $R$-th behavior. The auxiliary behavior graph undergoes the Auxiliary Preference Valuator to estimate interaction weights based on auxiliary behavior preferences, facilitating the computation of the Synergy Alignment Task loss, $L_{SAT}$. (3) Encode node features in the user-item bipartite graphs with interaction weights via graph neural networks. In each layer, nodes update their features by aggregating neighboring node features based on the interaction weights and self-loop weight $\lambda_s$. For instance, under the $r$-th behavior, user $u_1$ updates its $l$-th layer node feature $\mathbf{e}_{u_1,r}^{(l)}$ by aggregating the previous layer features of neighboring items, $\mathbf{e}_{i_1,r}^{(l-1)}, \mathbf {e}_{i_2,r}^{(l-1)},\mathbf{e}_{i_3,r}^{(l-1)},\mathbf{e}_{i_4,r}^{(l-1)}$, and its own upper layer features $\mathbf{e}_{u_1,r}^{(l-1)}$. A parallel process is executed for the target behavior. The $L$-th layer features are used as the encoded features under behaviors $1,\ldots,r,\ldots,R$ such as $\mathbf{e}_{u_1,1}^{(L)},\ldots, \mathbf{e}_{u_1,r}^{(L)}, ..., \mathbf{e}_{u_1,R}^{(L)}$. (4) Fuse multi-behavior user and item features through a self-attention mechanism, and summate to obtain the final feature representations, e.g., $\mathbf{e}_{u_1}^*, \mathbf{e}_{i_1}^*$. (5) Predict the preference scores through the dot product between the ultimate user and item embeddings, e.g., $\widetilde{y}_{u_1,i_1}$, denoting the preference score of user $u_1$ for item $i_1$. Then the BPR loss $L_{BPR}$ is computed using the true preference, e.g., $y_{u_1,i_1} \in \{0,1\}$, indicating the occurrence of interaction under the target behavior.
• Figure Fig. 3: Visualization of ablation experiment results.
• Figure Fig. 4: Visualization of self-loop weights affecting results.
• Figure Fig. 5: Visualization of the distribution of interaction weights for different behaviors.