On-Device Recommendation Algorithm Selection Using Remote Result Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
On-device recommendation systems face challenges in selecting a well-performing recommendation algorithm when data is scarce, especially for new users or new services, as traditional algorithms may perform poorly on richer data sets.
Innovation Solution
A method and device that compare local recommendation algorithms with a remote recommendation system using a query-set to select the algorithm with the smallest distance to the remote system's results, utilizing techniques like Singular Value Decomposition, k-nearest neighbors, or deep neural networks, and compute distances between graph representations of recommendation results to identify the best local algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a recommendation algorithm is selected based on performance on scarce data, then the system can operate on-device with limited data, but the algorithm may perform worse than other algorithms when more data becomes available
Solution Approach 1:
The system performs preliminary evaluation of multiple recommendation algorithms using a query-set and compares their results against a remote recommendation system before final selection. This preliminary action ensures the selected algorithm is pre-validated for performance on scarce data while also assessing its potential on richer data through comparison with the remote system's performance.
Solution Approach 2:
The system uses feedback from comparing local algorithm results with remote system results to guide the selection process. By measuring the distance between local and remote recommendation results, the system receives feedback about which algorithms perform best in the local scarce-data environment while maintaining compatibility with broader data scenarios.
2Measurement precision
If multiple recommendation algorithms are evaluated and compared, then the selection accuracy improves, but the computational complexity and time required increase
Solution Approach 1:
The evaluation process is segmented into distinct components: querying multiple local algorithms with a query-set, obtaining results from a remote recommendation system, computing distances between results, and making a selection based on comparison. This segmentation allows systematic evaluation of multiple algorithms while managing computational complexity through structured processing.
Solution Approach 2:
The remote recommendation system acts as an intermediary reference point for evaluating local algorithms. Instead of directly comparing all algorithms against each other or against ideal performance, the system uses the remote system's recommendations as a mediator to assess and compare local algorithm performance, simplifying the selection process.
Data Source
Figure 1
Figure 2a~2b
Figure 3
AI summary
A query-set is used to obtain recommendation results from different recommendation algorithms available locally on a device. The same query-set is used to obtain recommendation results from a remote, well-established on- line recommendation system using an unknown recommendation algorithm. Distances are measured between the recommendation results from the online recommendation system and the recommendation results from each of the local recommendation algorithms. The local recommendation algorithm having generated recommendation results with a smallest distance to the recommendation results from the online recommendation system is selected for generation of recommendations based on a local recommendation algorithm.