Re-ranking Model Selection Using Aggregated User Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems for personalized search and information access face challenges in creating effective re-ranking models due to insufficient data and sparsity of feature values in user search logs, leading to decreased search accuracy.
Innovation Solution
A re-ranking device that utilizes a plurality of pre-prepared re-ranking models based on common information from multiple users, allowing for a re-ranking process that selects and applies the most appropriate model to user search queries, thereby overcoming data shortages and enhancing accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a re-ranking model is created for each individual user using user-specific search logs, then personalization accuracy is improved, but data insufficiency occurs because the volume of individual user logs is too small
Solution Approach 1:
The patent merges search logs from multiple users to create a sufficient dataset for model training. By combining logs from plurality of users, the system accumulates enough data to train effective re-ranking models while still enabling personalized search results for individual users.
Solution Approach 2:
The patent creates a universal re-ranking model that can be applied to multiple users. This single model trained on aggregated data from multiple users serves the function of personalizing search results for any individual user, eliminating the need to create separate models for each user.
2Adaptability or versatility
If user-specific re-ranking models are created to provide personalized search results, then search personalization is improved, but feature value sparsity occurs because the items in user logs are too sparse
Solution Approach 1:
The patent combines feature values from multiple users' search logs to create a comprehensive feature set. By merging logs from plurality of users, the system overcomes the sparsity problem where individual user logs contain too few feature items, ensuring sufficient feature values for model training.
3Measurement precision
If multiple re-ranking models are prepared for different user groups, then model accuracy is improved, but device complexity increases due to model selection requirements
Solution Approach 1:
The patent segments users into different groups based on common characteristics and creates specific re-ranking models for each segment. This segmentation approach allows the system to prepare multiple specialized models for different user groups while managing complexity through organized model selection based on user characteristics.
Data Source
AI summary
Provided is a re-ranking device that enables a re-ranking process with less possibility of decrease in accuracy due to shortage of data. A re-ranking device 100 includes an input unit 101 that receives a search query from one user, a re-ranking model storage unit 106 that stores a plurality of re-ranking models prepared in accordance with common information of a plurality of users, a search unit 102 that performs a search on the basis of the search query and obtains a search result, and a re-ranking processing unit 107 that selects one re-ranking model on the basis of common information of the one user and performs a re-ranking process on the search result using the one re-ranking model.