Noise-Resistant Complementary Item Recommendation via Gaussian Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing complementary item recommender systems face challenges in accurately identifying real complementary relationships between items due to noisy labeling, particularly from co-purchase data, where popular items are frequently co-purchased with non-complementary items, leading to unreliable model performance and decreased business metrics.
Innovation Solution
The system represents items as Gaussian embeddings with mean vectors and covariance matrices, modeling co-purchases as a Gaussian distribution to distinguish between true complementary relationships and noise, and uses an independence test-based method to generate trustworthy labels for evaluation, thereby improving the accuracy of complementary item recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If co-purchase data is used to label complementary item pairs, then the recommender system can operate without ground truth labels, but the labels become noisy and unreliable
Solution Approach 1:
The patent introduces an independence test as an intermediary mechanism to filter and validate co-purchase labels. This statistical test acts as a mediator between the raw co-purchase data and the final labels used for training, selectively accepting only those item pairs that pass the independence test threshold, thereby reducing label noise while maintaining operational feasibility
Solution Approach 2:
The patent changes the parameter used for label validation from simple co-purchase frequency to statistical independence test results. By using the independence test statistic (chi-square value) as the filtering parameter, the system can objectively distinguish between true complementary relationships and spurious co-purchases, improving label reliability without requiring ground truth data
2Measurement precision
If popular items are removed from recommendations to reduce noise, then label quality improves, but business metrics like Gross Merchandise Value decrease
Solution Approach 1:
The patent applies local quality by treating different item pairs differently based on their specific characteristics. Instead of uniformly removing all popular items, the system evaluates each item pair through the independence test, which considers the specific relationship between that pair. This allows popular items to remain in recommendations when they have genuine complementary relationships, while filtering out only the noisy co-purchases
3Quantity of substance
If co-viewed but not co-purchased items are used as negative labels, then more training data is available, but the data remains noisy and unreliable
Solution Approach 1:
The patent extracts and removes the noisy portion from the co-viewed data by applying the independence test as a filtering mechanism. Instead of using all co-viewed item pairs as negative labels, the system extracts only those that pass the independence test, effectively separating the reliable negative examples from the noisy ones, thereby improving data reliability while maintaining sufficient training data quantity
Data Source
AI summary
Systems and methods for providing noise-resistant complementary item recommendations are disclosed. A trained model is generated based on transaction data to represent each item of a set of items as a Gaussian distribution with a mean vector and a non-zero covariance matrix. An anchor item is to be displayed to a user via a user interface executed on a user device of the user, and is represented as a Gaussian distribution with an anchor mean vector and an anchor non-zero covariance matrix. A complementarity score for each item is computed based on a distance between the mean vector of the item and the anchor mean vector to generate a ranking for the set of items based on their respective complementarity scores. A plurality of top items are selected from the set of items based on the ranking as recommended complementary items, which are displayed with the anchor item on the user interface.