Graph Neural Network Recommender System Leakage Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Graph neural network (GNN) based recommender systems suffer from overfitting due to feature leakage, which occurs when training data contains information that the model is trying to predict, leading to degraded performance during inference.
Innovation Solution
A message correction algorithm is used to modify user and item node embeddings generated by the GNN, specifically applying a correction function to exclude the contribution of positive target edges during training, allowing for parallel processing and reducing the need for multiple adjusted input graphs, thereby addressing feature leakage and improving computational efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If GNN-based recommender systems use training data containing user-item interaction history, then the model can learn user preferences and generate accurate recommendations, but feature leakage occurs where training data contains information the model is trying to predict, leading to overfitting and degraded performance during inference
Solution Approach 1:
The patent extracts and removes the positive target edges from the training graph to create a corrected graph structure. By taking out the edges that connect users to their positive target items, the system eliminates the feature leakage path while preserving other useful interaction patterns for training, thus preventing overfitting while maintaining predictive accuracy
Solution Approach 2:
The patent introduces a correction function as an intermediary component that processes the node embeddings after they are generated by the GNN. This correction function adjusts the embeddings by removing the contribution from positive target edges, acting as a mediator between the raw GNN output and the final training data, thereby eliminating feature leakage without losing predictive information
2Measurement precision
If the system trains GNN-based recommender systems for extended periods to improve accuracy, then predictive performance increases, but computing resources are consumed and training time increases
Solution Approach 1:
The patent performs preliminary action by pre-processing the training graph to remove positive target edges before training begins. By preparing the corrected graph structure in advance, the system eliminates the need for extensive training iterations to compensate for feature leakage, thereby reducing training time while maintaining accuracy
Solution Approach 2:
The patent changes the structural parameter of the training graph by removing positive target edges from the adjacency matrix. This parameter change fundamentally alters how information flows during training, preventing leakage while allowing the model to converge faster to accurate representations, thus reducing training time and resource consumption
3Measurement precision
If the system uses multiple adjusted input graphs for training to address feature leakage, then predictive accuracy improves, but computational complexity and resource requirements increase
Solution Approach 1:
The patent merges the feature leakage correction operation with the existing GNN training pipeline by integrating the correction function as a single operational step. Instead of requiring multiple separate adjusted input graphs, the system combines the correction mechanism into one unified process that operates on a single corrected graph structure, thereby reducing computational complexity while maintaining accuracy
Data Source
AI summary
Systems, methods, and computer-readable media provide a graph processing system that incorporates a graph neural network (GNN) based recommender system (RS), as well as a method for training a GNN based RS to address feature leakage that leads to overfitting of the trained GNN based RS. A message correction algorithm is used to modify a user node embedding and a positive item node embedding generated by the graph neural network when generating mini batches of training triples used to train the GNN based RS. The GNN message passing operations are performed on one graph only, in contrast to existing approaches which typically run GNN message passing operations on multiple adjusted input graphs constructed for multiple training triples.