Facility Similarity Model Selection for Domain-Shifted Recommendations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information suggestion techniques face challenges in selecting a suitable model for a new facility due to domain shift, as they often lack sufficient data for performance evaluation, leading to decreased prediction accuracy when introduced in unfamiliar domains.
Innovation Solution
A method to select a model suitable for a new facility by evaluating similarity degrees between the new facility and training facilities using metadata-derived and facility-related information, allowing selection without direct performance evaluation on the new facility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a model is trained using data collected at a specific facility, then the prediction accuracy is improved at that facility, but the prediction accuracy decreases when the model is introduced to a different facility due to domain shift
Solution Approach 1:
The patent applies preliminary action by pre-training multiple models at different facilities before deployment. Instead of training a single model at the target facility (which requires sufficient data), the system prepares multiple candidate models in advance at various source facilities, then selects the most suitable one for the target facility based on domain similarity assessment.
Solution Approach 2:
The patent changes the parameter of model selection criteria by introducing domain similarity assessment based on metadata features. Rather than relying solely on prediction performance metrics (which require target facility data), the system evaluates and selects models based on similarity between source and target facility characteristics, such as user demographics, item categories, and behavioral patterns.
2Measurement precision
If behavior history data is collected for performance evaluation, then model selection accuracy is improved, but the system cannot be applied when sufficient behavior history data is unavailable at the new facility
Solution Approach 1:
The patent introduces metadata as an intermediary for assessing domain similarity. Instead of directly using behavior history data (which may be unavailable) to evaluate model performance, the system uses metadata features (user attributes, item attributes, facility characteristics) as a mediator to indirectly assess whether a model trained at a source facility will perform well at the target facility.
Solution Approach 2:
The patent copies metadata features from the target facility to assess domain similarity with source facilities. By copying and comparing structural characteristics (user profiles, item categories, facility attributes) rather than requiring actual behavior history data, the system can evaluate model suitability without needing sufficient target facility data for direct performance measurement.
3Reliability
If multiple models are trained at different facilities to handle domain shift, then robustness against domain shift is improved, but the complexity of model selection increases
Solution Approach 1:
The patent extracts key metadata features that characterize each facility's domain properties. Instead of considering all possible model parameters and performance metrics, the system extracts and compares specific relevant features (user demographics, item categories, behavioral patterns) to simplify the model selection process while maintaining robustness against domain shift.
Data Source
AI summary
An information processing method executed by one or more processors, in which a plurality of models, which are trained by using one or more of datasets including a behavior history of a user on an item, which is collected at each of a plurality of first facilities different from each other, are prepared, and the one or more processors include acquiring a characteristic of a second facility, which is different from the plurality of first facilities and acquiring a characteristic of each of the plurality of first facilities, evaluating a similarity degree between the acquired characteristic of the second facility and the characteristic of the first facility where the dataset, which is used for the training of the model, is collected, and selecting a model suitable for the second facility from among the plurality of models based on the similarity degree.