University of Illinois Urbana-Champaign

Demystifying graph neural networks in recommender systems

Ji, Houxiang

Content Files
JI-THESIS-2021.pdf

Permalink

https://hdl.handle.net/2142/113920

Description

Title
Demystifying graph neural networks in recommender systems
Author(s)
Ji, Houxiang
Issue Date
2021-12-06
Director of Research (if dissertation) or Advisor (if thesis)
Torrellas, Josep
Department of Study
Computer Science
Discipline
Computer Science
Degree Granting Institution
University of Illinois at Urbana-Champaign
Degree Name
M.S.
Degree Level
Thesis
Date of Ingest
2022-04-29T21:34:54Z
Keyword(s)
  • Graph neural network
  • Recommender systems
  • Microarchitecture
Abstract
Recommender systems have become indispensable tools for many applications with the explosive growth of online information. Recommender systems learn user's interest based on their profile and historical behavior and then recommend the item with the highest predicted ratings. Graph Neural Networks (GNNs) is a powerful graph representation learning method and have shown their unprecedented performance on many scenarios including natural language processing and computer vision. Most data required in the recommender systems is naturally and essentially represented with graph structure, for example, user-item interactions can be represented as a bipartite graph. With the superiority in graph learning, GNNs are gaining more and more attention in the field of recommender systems. This thesis presents our study of GNNs, which are employed in the recommender systems, to characterize their runtime behavior from various levels. To this end, we carefully profile GNN models on training and identify the most expensive operations which are worthy of attention for overhead reduction. We observe that the memory-intensive aggregation phase in the GNNs dominate the overall runtime different from the conventional deep neural network models. Further, we investigate the GNN's behavior at the micro-architectural level, which have received little attention so far, and summarize several key observations. Based on these insightful observations, we propose and discuss software and possible hardware mechanisms to optimize GNN-based recommender systems.
Graduation Semester
2021-12
Type of Resource
Thesis
Permalink
http://hdl.handle.net/2142/113920
Copyright and License Information
Copyright 2021 Houxiang Ji

Owning Collections

Graduate Dissertations and Theses at Illinois PRIMARY
Graduate Theses and Dissertations at Illinois