Contrastive Learning for Shilling-Resistant Recommendation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current recommendation systems face challenges in accurately detecting fake users and maintaining effective recommendations under shilling attacks, with existing methods exhibiting low detection accuracy and poor defense capabilities.
Innovation Solution
A method and system utilizing contrastive learning to enhance defense against shilling attacks by establishing a user-item interaction matrix, performing data augmentation, and jointly training a contrastive learning model, user detection model, and recommendation model using weighted dot products and Bayesian personalized ranking loss functions to improve user and item confidence values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If graph neural network models are used to aggregate high-order neighbor information for more accurate recommendations, then recommendation accuracy is improved, but the system becomes more vulnerable to shilling attacks
Solution Approach 1:
The patent extracts and removes fake user nodes from the graph structure before performing graph neural network computations. By detecting fake users through contrastive learning and excluding them from the neighborhood aggregation process, the system maintains the benefits of high-order neighbor information while eliminating the harmful influence of shilling attacks.
Solution Approach 2:
The patent introduces an intermediary detection layer that uses contrastive learning models to identify fake users before the main recommendation process. This intermediary layer acts as a filter between the raw graph data and the recommendation output, preventing fake user information from propagating through the graph neural network.
2Device complexity
If traditional machine learning methods are used to detect fake users, then the detection process is simple, but the method performs poorly with large data volumes due to poor generalization ability
Solution Approach 1:
The patent replaces traditional machine learning detection methods with contrastive learning-based deep neural networks. This substitution enables the system to handle large data volumes effectively by using contrastive learning to capture complex patterns and relationships in the data, significantly improving generalization ability and detection accuracy.
Solution Approach 2:
The patent changes the detection approach by introducing contrastive learning with specific loss functions (binary cross-entropy and focal loss) and weighting mechanisms. These parameter changes enable the model to focus on difficult-to-detect fake users while maintaining simplicity in the overall detection process.
3Device complexity
If graph neural network models generate feature vectors for detection, then the detection process is integrated, but cross-propagation of information between fake users and normal users results in low detection accuracy
Solution Approach 1:
The patent extracts and removes fake user nodes from the graph structure before performing graph neural network computations. By detecting fake users through contrastive learning and excluding them from the neighborhood aggregation process, the system maintains the benefits of high-order neighbor information while eliminating the harmful influence of shilling attacks.
Solution Approach 2:
The patent introduces an intermediary detection layer that uses contrastive learning models to identify fake users before the main recommendation process. This intermediary layer acts as a filter between the raw graph data and the recommendation output, preventing fake user information from propagating through the graph neural network.
Data Source
AI summary
The present disclosure discloses a method and system based on contrastive learning for assisting a recommendation system to defend against shilling attacks, and belongs to the field of recommendation system security. The method includes: establishing an interaction matrix using interaction data; achieving data augmentation by using the target user's neighbor nodes as positive samples and a non-target user's neighbor nodes as negative samples; establishing a contrastive learning model, a user detection model, and a recommendation model resistant to shilling attacks, performing joint training on the three models, computing user-item interaction scores after training, and taking the top-20 scoring interacted items to form a recommendation list for the user. Based on a recommendation system, the present disclosure combines contrastive learning and a multi-layer perceptron to detect fake users and utilizes confidence values as weights to implement a recommendation system method resistant to shilling attacks.