Authenticated Negative Items in Recommendation Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current reinforcement learning models for online recommendations struggle to distinguish between items not selected due to lack of popularity or quality versus those not selected for other reasons, leading to irrelevant recommendations and a diminished user experience.
Innovation Solution
The method involves generating recommendation sets by incorporating authenticated negative items, which are items known to be disfavored by users, to improve the model's understanding of user preferences and reduce the likelihood of recommending undesirable items.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the reinforcement learning model includes unpopular items in recommendations to distinguish non-selection reasons, then the model's ability to distinguish between items not selected due to lack of popularity versus other reasons is improved, but the user receives irrelevant recommendations and the user experience is diminished
Solution Approach 1:
The patent segments the set of non-selected items into two distinct categories: authenticated negative items (which provide training value but are not shown to users) and regular non-selected items (which may be shown to users). This segmentation allows the system to use unpopular items for model training while preventing them from appearing in user-facing recommendations, thus resolving the contradiction between improving measurement precision and maintaining ease of operation.
2Reliability
If the model recommends items based on past selections to improve relevance, then the likelihood of user selection and purchase is improved, but the model cannot distinguish between temporary non-selection and permanent disfavor
Solution Approach 1:
The patent introduces an intermediary mechanism - the authenticated negative items database - that acts as a mediator between user non-selection behavior and recommendation generation. Instead of directly using all non-selected items as negative signals, the system first authenticates them through the intermediary database before incorporating them into the model, thereby preserving accurate user preference information while maintaining recommendation reliability.
3Measurement precision
If the model avoids recommending items related to non-selected items to improve accuracy, then irrelevant recommendations are reduced, but items that users may want later are excluded
Solution Approach 1:
The patent applies preliminary action by pre-authenticating negative items and storing them in a dedicated database before the recommendation generation process. This preliminary authentication step allows the system to later retrieve and use these authenticated negatives for training without directly blocking related items from recommendations, thereby maintaining both recommendation accuracy and temporal flexibility in user preferences.
Data Source
AI summary
One embodiment of the present invention sets forth techniques for generating recommendation sets for a first client device. A recommendation system receives, from the first client device, a first selection of a first recommended item included in a plurality of recommended items. The recommendation system identifies a second recommended item included in the plurality of recommended items that has not been selected. The recommendation system retrieves an authenticated negative item from a plurality of authenticated negative items. The recommendation system stores one or more entries in a log file comprising a plurality of entries, based on at least one of the first recommended item, the second recommended item, and the authenticated negative item. One advantage of the disclosed techniques is that the use of authenticated negative examples, also referred to herein as authenticated negative items, provides a more relevant set of recommendations for the user.