Page-Level Reranking Model for Multi-List Recommendation Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing reranking models, such as DHANR, are limited in capturing dynamic inter-list interactions and user behavior differences across lists, as they treat page-level information statically and fail to account for the format of the recommendation page, leading to suboptimal recommendations in multi-list scenarios.
Innovation Solution
A unified reranking model that incorporates hierarchical dual-side attention and spatial-scaled attention modules to extract intra-list and inter-list interactions, along with a multi-gated mixture-of-experts module to capture commonalities and differences in user behavior, while considering the format of the recommendation page to determine pairwise item influences and generate accurate reranking scores.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If list-level reranking models are used to improve recommendation accuracy within single lists, then intra-list cross-item influence is captured, but inter-list interactions and page-level format information are ignored
Solution Approach 1:
The patent merges list-level reranking with page-level reranking by integrating inter-list interaction modeling into the existing list-level framework. The reranking model simultaneously processes multiple candidate lists and captures cross-item influences both within lists and across lists, unified by page-level format information that coordinates the interaction between different list levels.
Solution Approach 2:
The patent transitions from single-list dimension to multi-list dimension by introducing page-level format information as an additional dimension. This allows the model to capture cross-item influences not only within individual lists but also across multiple lists on the recommendation page, effectively adding a new dimension of interaction modeling.
2Ease of manufacture
If static page representation is used to incorporate page-level information, then implementation simplicity is maintained, but dynamic inter-list interactions and user behavior differences are not captured
Solution Approach 1:
The patent introduces dynamic elements to page-level representation by incorporating user behavior data and interaction patterns. Instead of using fixed static representations, the model dynamically adjusts page-level format information based on user interactions, allowing it to capture evolving inter-list relationships and user preferences across different browsing sessions.
Solution Approach 2:
The patent implements feedback mechanisms where user interaction data is continuously incorporated into the page-level representation. The model uses user behavior feedback to refine its understanding of inter-list interactions, creating a loop where static page format information is enhanced with dynamic user-specific interaction patterns, improving both accuracy and personalization.
3Measurement precision
If multiple candidate lists are reranked simultaneously to capture inter-list interactions, then holistic recommendation quality improves, but computational complexity increases
Solution Approach 1:
The patent segments the reranking process into hierarchical levels: page-level format information extraction, list-level candidate scoring, and inter-list interaction modeling. This segmentation allows the system to process multiple lists simultaneously by breaking down the complex computation into manageable modules, where page-level information serves as a coordinating framework that reduces the overall computational burden.
Solution Approach 2:
The patent performs preliminary extraction of page-level format information before conducting detailed list-level reranking. By pre-processing and encoding the page structure and format characteristics in advance, the model reduces the computational complexity of subsequent inter-list interaction modeling, as the framework for coordinating multiple lists is already established.
Data Source
AI summary
A system is provided for reranking. The system comprises a user device and one or more servers. The system is configured to receive a plurality of candidate lists, rerank the plurality of candidate lists based on page-level information and a format of a recommendation page, generate recommendation results based on the reranked lists, and send the recommendation results to the user device. Each candidate list comprises a plurality of candidate items. The page-level information comprises interactions between the candidate items in each candidate list and between different candidate lists among the plurality of candidate lists. The reranking comprises using the format of the recommendation page to determine pairwise item influences between candidate item pairs among the candidate items in the candidate lists. The user device is configured to display the recommendation page with the recommendation results from the one or more servers.