Member-only story
KDD2024 | Research on Privacy Risks in Interest Point Recommendation
Hey, remember to give it to me“ Machine Learning and Recommendation Algorithms ”Add Star symbol
**TLDR:**This article first questions the necessity of graph convolution in the training stage of recommendation systems through experiments, and then proposes a post training graph convolution framework and lightweight graph ordinary differential equations as alternatives to traditional graph convolutional neural networks. This method avoids the most time-consuming message passing in traditional training methods and has extremely high performance and efficiency on large-scale graphs.
paperhttps://arxiv.org/abs/2407.18910
Warehouse:https://github.com/DavidZWZ/LightGODE
1. Research background
In this section, we first explore the necessity of graph convolution for recommendation systems and analyze the key reasons why the matrix factorization (MF) model enhanced by graph convolution after post training unexpectedly achieves superior performance. Subsequently, we identified the issue of embedding differences when constructing deeper graph convolutional layers, and clarified the factors that need to be balanced when training graph convolutions after design.
