Bipartite Graph Recommendation Weights for Popularity Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Recommendation systems based on collaborative filtering algorithms suffer from popularity bias, where popular items are over-recommended, and graph algorithms struggle to provide meaningful explanations for recommendations.
Innovation Solution
A method that calculates edge weights in a bipartite graph using the inverse of user and item node degrees to de-emphasize popularity bias, and uses path scores calculated through matrix multiplication to recommend items, while also generating explanations based on interior nodes in the paths between the target user and recommended items.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If collaborative filtering algorithms are used to build recommendation models from user behavior and similar users' decisions, then the system can generate personalized recommendations, but popular items are disproportionately recommended and highly active users' interactions are over-represented
Solution Approach 1:
The patent applies parameter changes by transforming the edge weights in the bipartite graph using the formula w' = w * (k / (deg(u) + deg(i))), where deg(u) and deg(i) are the degrees of user and item nodes. This parameter transformation reweights interactions to account for node popularity, thereby reducing the disproportionate influence of popular items and highly active users while maintaining the collaborative filtering recommendation capability
Solution Approach 2:
The patent introduces an intermediary mechanism by using a bipartite graph as a structured representation of user-item interactions. The graph structure with calculated edge weights serves as an intermediary that mediates between raw interaction data and final recommendations, allowing the system to capture nuanced relationship patterns while controlling for popularity bias through the graph's structural properties
2Adaptability or versatility
If graph algorithms are used to model user-item interactions, then the system can capture complex relationship patterns, but it becomes challenging to generate useful explanations for recommendations
Solution Approach 1:
The patent applies segmentation by decomposing the recommendation explanation into discrete path components in the bipartite graph. Each path from user to item is segmented into sequential edges representing specific interaction patterns, and the contribution of each path to the final recommendation score is separately calculated and can be individually explained, making the overall recommendation interpretable despite the complexity of the graph structure
3Object-generated harmful factors
If edge weights are calculated using inverse of node degrees to reduce popularity bias, then the proportion of relevant interactions is emphasized, but the computational complexity of calculating path scores increases
Solution Approach 1:
The patent replaces the mechanical computation of individual path scores with matrix multiplication operations. By representing the bipartite graph as adjacency matrices and using matrix operations to compute aggregated path scores, the system maintains accuracy while leveraging optimized linear algebra libraries to reduce computational overhead compared to explicit path enumeration
Data Source
AI summary
A method may include obtaining interactions between users and items, and calculating, for each edge in a bipartite graph, an edge weight using an inverse of the degree of a user node connected to the edge and an inverse of the degree of an item node connected to the edge. The bipartite graph includes user nodes corresponding to the users and item nodes corresponding to the items. The method may further include identifying paths each including an edge connecting the target user node and a common item node, an edge connecting a neighboring user node and the common item node, and an edge connecting the neighboring user node and a neighboring item node. The method may further include calculating, using the edge weights calculated for the edges, scores for the paths, and recommending, to the target user and using the scores for the paths, a recommended item.