In practical recommendation scenarios, beyond the target behavior (e.g., purchase), there exist multiple auxiliary behaviors (e.g., click, favorite, and add to cart) that can provide additional information about user preferences. However, not all auxiliary behaviors are useful, with some potentially being noise, such as accidental clicks or arbitrary favorites, which may distort the recommendation model’s understanding of a user’s true preferences. Hence, removing these noisy behaviors is key to enhancing recommendation quality. Moreover, existing multi-behavior recommendation systems typically embed graph data into the Euclidean space for graph convolution operations. However, the Euclidean space presents structural distortions in representing the natural power-law distribution of graphs, leading to the suboptimal results for graph-based multi-behavior recommendation systems. To address these issues, we propose a Trust-Enhanced Twin Graph Convolutional Framework (TETCF). Speciffcally, we design a trust-aware auxiliary behavior denoising module to model the conffdence of auxiliary behaviors, and utilize a twin graph convolutional network module (consisting of the Euclidean GCN and Hyperbolic GCN) to compensate for the shortcomings of the traditional Euclidean space in representing graph data structures. During the training process, we tune the model through an adaptive denoising strategy and a learnable weight matrix fusion mechanism. Comprehensive experiments on three real-world datasets show a signiffcant improvement in the performance of our model in a multi-behavioral recommendation task, validating its effectiveness.
Discussion(0)
No comments yet. Be the first to comment.