Work

Denoising Long-Tail Augmented Contrastive Network for Multi-Behavior Recommendation

Year: 2024

Type: article

Abstract: Recommendation systems in multi-behavior recommendation systems integrate various user behavior data (e.g., favorites, cart) for representation learning, often alleviating the problem of data sparsity and enabling a more comprehensive user profile compared to traditional recommendation systems. Recent studies have proposed using graph convolutional networks combined with unsupervised learning to capture the dependencies between multi-behavior data, providing additional supervision signals for the target behavior. However, there are still two major issues that need to be addressed: i) The long-tail problem in multi-behavior recommendation. Existing research on multi-behavior recommendation only alleviates data sparsity globally without precise data augmentation for low-exposure nodes. As a result, high-exposure (i.e., higher degree) nodes continue to receive more attention from the model, while low-exposure nodes are easily overlooked, leading to the persistence of the long-tail problem.ii) The negative transfer problem of auxiliary behaviors. Current research primarily models the data associations within a single behavior, neglecting the correlations between multiple behaviors and the target behavior. To address these issues, we propose a new model called Denoising Long-tail Augmented Contrastive Network (DLACN) to mine low-noise auxiliary behavior information for precise long-tail enhancement. Specifically, we propose a correlation amplify strategy to generate augmented views, constructing a new graph augmentation contrastive learning task. This strategy uses denoised auxiliary behavior information for precise enhancement of long-tail nodes, mitigating the long-tail problem. Additionally, we build a lightweight inter-behavior gradient denoising weights to reduce the introduction of noisy auxiliary information during the learning of graph convolutional layers, thereby alleviating the negative transfer problem caused by auxiliary tasks. Experiments on three real-world datasets demonstrate that our model outperforms various state-of-the-art models.

Source: IEEE Access

DOI: 10.1109/access.2024.3497217

Authors: Jinle He, Chengyong Yang, Jiayi Liu, Jianlin Cheng

Institution: Guilin University of Technology

Language: English

Topic: Recommender Systems and Techniques

Subfield: Information Systems

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $1,995

Funder: National Natural Science Foundation of China