Machine Learning Reranker for Ranking Virtual Experience Items
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine learning-based rerankers for recommending virtual experiences in a user interface do not effectively incorporate the context of co-impressed items, leading to suboptimal ranking quality when multiple items are displayed together, as they focus on individual item rankings rather than the context of other items displayed alongside them.
Innovation Solution
A machine learning-based reranker model is trained using training data that includes both user-selected and co-impressed items, calculating a pointwise loss and a comparator loss to adjust its parameters, incorporating context and improving ranking accuracy by considering the relationships between items displayed together.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning-based rerankers focus on individual item rankings, then the model complexity remains relatively simple, but the ranking quality deteriorates due to lack of context from co-impressed items
Solution Approach 1:
The patent merges the pointwise ranking function with the pairwise comparison function into a single unified model. The model simultaneously learns to rank individual items based on their features and to compare items relative to each other in the context of co-impressed items. This integration allows the model to capture both individual item quality and relative item relationships, improving ranking quality while maintaining a single coherent model structure rather than requiring multiple separate models.
Solution Approach 2:
The patent adds a new dimension to the ranking problem by introducing pairwise comparisons between co-impressed items. Instead of only considering individual item features, the model now operates in an expanded feature space that includes relationships between items displayed together. This dimensional expansion allows the model to capture contextual information about how items relate to each other, thereby improving ranking quality without excessively increasing model complexity.
2Measurement precision
If the reranker model incorporates context of co-impressed items by calculating both pointwise loss and comparator loss, then the ranking accuracy improves, but the computational cost increases
Solution Approach 1:
The patent combines the pointwise loss calculation and comparator loss calculation into a single unified training process. Rather than separately training one model for pointwise ranking and another for pairwise comparisons, the unified model learns both functions simultaneously through a combined loss function. This merging reduces the total computational cost compared to running multiple separate models, while still achieving improved ranking accuracy through the incorporation of co-impressed item context.
3Ease of operation
If traditional rerankers display multiple items without considering their contextual relationships, then the user interface implementation is simpler, but the user experience deteriorates due to suboptimal ranking
Solution Approach 1:
The ranking model automatically learns to incorporate contextual relationships between co-impressed items through self-supervised learning from user interaction data. The model uses the implicit feedback from user selections and rejections to automatically adjust its ranking behavior, eliminating the need for manual curation or complex business rules. This self-service capability improves user experience by adapting to user preferences while keeping the implementation relatively simple compared to manual ranking approaches.
Data Source
AI summary
A computer-implemented method includes receiving training data that includes groups of items and a respective user associated with each group, where each group includes a first item selected by the associated user and one or more second items rejected by the associated user from a user interface in which the first item and the one or more second items are presented together in ranked order. The method includes, for each group in the group of items: generating feature embeddings, calculating a pointwise loss for each item in the group based on the feature embeddings, calculating a comparator loss for a set that includes the first item and at least one of the one or more second items, and adjusting one or more parameters of the machine learning model based on the pointwise loss and the comparator loss. The method further includes obtaining a trained machine learning model.