A Multi-Head Attention Based Dual Target Graph Collaborative Filtering Network

• Published in: 2022 IEEE Smartworld, Ubiquitous Intelligence & Computing, Scalable Computing & Communications, Digital Twin, Privacy Computing, Metaverse, Autonomous & Trusted Vehicles (SmartWorld/UIC/ScalCom/DigitalTwin/PriComp/Meta) pp. 476 - 483
• Main Authors: Peng, Qinglong, Tang, Bin, Liu, Jinhuan, Cui, Shuang, Du, Junwei, Lu, Yan, Jiang, Feng, Yu, Xu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2022
• Subjects: Adaptation models; Bidirectional control; Collaboration; Collaborative filtering; Convolution; Cross-domain collaborative filtering; Graph neural network; Multi-head attention; Neural networks; Stacking
• DOI: 10.1109/SmartWorld-UIC-ATC-ScalCom-DigitalTwin-PriComp-Metaverse56740.2022.00086

Recently, cross-domain collaborative filtering (CDCF) has been widely used to solve the data sparsity problem in recommendation systems. Therein, the dual-target cross-domain recommendation becomes a research hotspot, which aims to improve the recommendation performance of both target and source domains. Most existing approaches tend to use fixed weights or self-attention in a single representation space for the bi-directional inter-domain transfer of the user representation. However, a single representation space leads to limited representation capability, which makes the transfer of the user representation coarse-grained and inaccurate. In this paper, Multi-head Attention Based Dual Target Graph Collaborative Filtering Network (MA-DTGCF) is proposed. The core of the model is the bi-directional transfer graph convolution layer, consisting of a graph convolution layer and a bi-directional transfer layer based on a multi-head attention mechanism. The latter can achieve fine-grained and adaptive transfer of user features in multiple representation subspaces. It is worth noting that by stacking multiple bi-directional transfer graph convolutional layers, we can get high-order user and item features and achieve adaptive transfer of each order user features. Experimental results on three real datasets show that the proposed MA-DTGCF model significantly outperforms the state-of-the-art models in terms of HR and NDCG.