Contrastive Graph Learning with Positional Representation for Recommendation

Yi, Z., Ounis, I. and Macdonald, C. (2023) Contrastive Graph Learning with Positional Representation for Recommendation. In: Advances in Information Retrieval. Lecture Notes in Computer Science: 45th European Conference on Information Retrieval, ECIR 2023, Dublin, Ireland, April 2–6, 2023, Proceedings, Part II. Springer: Cham, pp. 288-303. ISBN 9783031282386 (doi: 10.1007/978-3-031-28238-6_19)

Abstract

Recently, graph neural networks have become the state-of-the-art in collaborative filtering, since the interactions between users and items essentially have a graph structure. However, a major issue with the user-item interaction graph in recommendation is the absence of the positional information of users/items, which limits the expressive power of graph recommenders in distinguishing the users/items with the same neighbours after propagating several graph convolution lay-ers. Such a phenomenon further induces the well-known over-smoothing problem. We hypothesise that we can obtain a more expressive graph recommender through graph positional encoding (e.g., Laplacian eigen-vector) thereby also alleviating the over-smoothing problem. Hence, we propose a novel model named Positional Graph Contrastive Learning (PGCL) for top-K recommendation, which aims to explicitly enhance graph representation learning with graph positional encoding in a con-trastive learning manner. We show that concatenating the learned graph positional encoding and the pre-existing users/items’ features in each feature propagation layer can achieve significant effectiveness gains. To further have sufficient representation learning from the graph positional encoding, we use contrastive learning to jointly learn the correlation be-tween the pre-exiting users/items’ features and the positional informa-tion. Our extensive experiments conducted on three benchmark datasets demonstrate the superiority of our proposed PGCL model over exist-ing state-of-the-art graph-based recommendation approaches in terms of both effectiveness and alleviating the over-smoothing problem.

• Item Type: Book Sections
• Status: Published
• Refereed: Yes
• Glasgow Author(s) Enlighten ID: Macdonald, Professor Craig and Ounis, Professor Iadh and Yi, Zixuan
• Authors: Yi, Z., Ounis, I., and Macdonald, C.
• College/School: College of Science and Engineering > School of Computing Science
• Publisher: Springer
• ISBN: 9783031282386
• Copyright Holders: Copyright: © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
• First Published: First published in Advances in Information Retrieval. ECIR 2023. Lecture Notes in Computer Science, vol 13981:: 288–303
• Publisher Policy: Reproduced in accordance with the publisher copyright policy