Co-watch Graph Recommendation System Reducing Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current recommendation systems suffer from biases such as popularity bias, family bias, position bias, and multiple interest and interest shift bias, which can lead to recommendations that do not accurately reflect a user's real interests.
Innovation Solution
The use of a co-watch graph is proposed, where nodes represent items viewed by a user and edges represent the co-occurrence of these items within a predetermined time interval. The system adjusts edge weights to reduce bias, allowing for more accurate recommendations based on the user's view history.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional recommendation systems use popularity-based algorithms, then general public demand is reflected in recommendations, but user-specific real interests are not accurately captured
Solution Approach 1:
The patent segments the recommendation problem by creating a co-watch graph that separates individual user viewing behavior from general popularity metrics. Each node represents a specific item viewed by the user, and edges represent co-occurrence relationships, allowing the system to analyze user-specific patterns independently from overall popularity data.
Solution Approach 2:
The co-watch graph serves as an intermediary structure between raw viewing history data and recommendation outputs. It transforms raw view sequences into a normalized graph representation where edge weights are adjusted to remove bias, acting as a mediator that filters out popularity bias while preserving user-specific interest patterns.
2Ease of manufacture
If the system uses raw co-occurrence counts for edge weights, then simple implementation is achieved, but family bias and popularity bias remain in recommendations
Solution Approach 1:
The patent changes the parameter representation from raw co-occurrence counts to normalized edge weights. The normalization process transforms the weight values by dividing by the product of node weights, converting absolute counts into relative probabilities that reflect actual co-viewing likelihood while removing bias from the data.
3Quantity of substance
If the system analyzes long-term view history, then comprehensive user interest patterns are captured, but bias from popular items viewed by family members or general public increases
Solution Approach 1:
The patent extracts and isolates user-specific viewing patterns from the broader view history data. By constructing a co-watch graph that only includes items and edges relevant to the specific user account, it separates individual user interests from family or public viewing patterns, allowing long-term analysis without family bias contamination.
Data Source
AI summary
Disclosed are mechanisms for selecting a recommended item for a current item being viewed by a user account based on a view history of the user account with reduced bias. For a current item being viewed by the user account represented by a current node of a co-watch graph, embodiments can select a recommended item represented by an associated node in the co-watch graph likely being viewed by the user account, and determine a probability of the recommended item likely being viewed. The co-watch graph can be generated based on a view history of the user account. An edge between a first node and a second node of the co-watch graph can have a weight representing a number of co-occurrence times when the first item represented by the first node and the second item represented by the second node are viewed in sequence within a predetermined time interval.