Tenant-Specific Hierarchical Code Embeddings for Cold-Start Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing search systems face challenges in efficiently determining item category codes from user search terms due to the 'cold start' problem, where models require sufficient usage data and are tenant-specific, leading to confidentiality issues and reduced accuracy.
Innovation Solution
A system that generates tenant-specific hierarchical group code representations using embeddings based on user search terms, allowing for accurate category prediction without shared models, and boosts search results based on these codes, ensuring confidentiality and reducing maintenance costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are trained on usage log data to identify item category codes, then prediction accuracy improves, but the system suffers from cold start problem and requires sufficient usage data
Solution Approach 1:
The system pre-generates synthetic search terms and their corresponding category codes before actual usage data is available. This preliminary generation of training data allows the model to be trained immediately without waiting for sufficient usage logs to accumulate, thereby resolving the cold start problem while maintaining prediction accuracy
2Measurement precision
If tenant-specific machine learning models are trained separately for each tenant, then prediction accuracy for each tenant improves, but maintenance costs increase and model sharing becomes impossible
Solution Approach 1:
The system creates a universal synthetic data generation mechanism that can serve multiple tenants simultaneously. The same synthetic data generation process and base model architecture are reused across all tenants, allowing model sharing while maintaining tenant-specific customization through configurable parameters. This reduces maintenance complexity compared to completely separate models for each tenant
3Reliability
If tenant-specific models are trained separately, then confidentiality is maintained, but the system cannot leverage shared learning across tenants
Solution Approach 1:
The system generates synthetic search terms and category codes that replicate realistic usage patterns without copying actual tenant data. These synthetic copies capture the statistical properties and relationships of real data while containing no sensitive information, enabling shared learning across tenants while maintaining confidentiality of actual tenant data
Data Source
AI summary
Systems and methods include reception of a plurality of item descriptions associated with a respective hierarchical code which includes a plurality of group codes, extraction of a respective plurality of phrases from each item description, assignment, for each item description, a phrase embedding determined for each of the extracted respective plurality of phrases to each of the plurality group codes included in the hierarchical code associated with the item description, generation of a composite embedding for each of the plurality of group codes based on the phrase embeddings assigned to the group code, and storage of each composite embedding in association with the group code for which the composite embedding was generated.