Performance Aware Layer Combination for Graph Convolutional Network
Authors
Lybech, Mathias Møller ; Tróndarson, Jens Petur ; Schrøder, Frederik Valdemar
Term
4. term
Education
Publication year
2021
Submitted on
2021-06-10
Pages
32
Abstract
Recommender systems suggest items to users. A common approach is Graph Convolutional Networks (GCNs), which learn from connections between users and items represented as a graph. In this work, we study how the way information from multiple network layers is merged (layer combination) affects recommendation quality. We conduct an ablation study of GCF to understand why it outperformed LightGCN on their datasets. Our main focus is LightGCN, a streamlined version of Neural Graph Collaborative Filtering (NGCF) that previously outperformed NGCF and other leading methods. We introduce two new extensions to LightGCN—Aggressive Layer Combination (ALC) and Balanced Layer Combination (BLC)—which replace LightGCN’s weighted summation for layer combination. These extensions achieve better results on most datasets than both GCF and LightGCN. We also show that, in some cases, using the embedding (the learned vector representation) from a single layer outperforms ALC, BLC, GCF, and LightGCN.
Anbefalingssystemer foreslår elementer til brugere. En udbredt teknik er grafkonvolutionelle netværk (GCN), som lærer fra forbindelser mellem brugere og elementer repræsenteret som en graf. I dette arbejde undersøger vi, hvordan måden at kombinere information på tværs af netværkets lag (lagkombination) påvirker kvaliteten af anbefalinger. Vi udfører et ablationsstudie af GCF for at forstå, hvorfor det klarede sig bedre end LightGCN på deres datasæt. Vores hovedfokus er LightGCN, en forenklet udgave af Neural Graph Collaborative Filtering (NGCF), som tidligere har overgået NGCF og andre førende metoder. Vi foreslår to nye udvidelser til LightGCN, Aggressive Layer Combination (ALC) og Balanced Layer Combination (BLC), som erstatter LightGCN’s vægtede summering ved lagkombination. Disse gav bedre resultater på de fleste datasæt end både GCF og LightGCN. Vi viser også, at det i nogle tilfælde er bedst kun at bruge indlejringen (den lærte vektorrepræsentation) fra ét enkelt lag, som overgår ALC, BLC, GCF og LightGCN.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords
recommendation ; graph ; convolutional ; network ; GCN ; lightgcn ; ngcf ; gcf ; bitgcf ; collaborative filtering