Learning Model Parameter Update via Synthetic Click Logs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating learning models for search engines face challenges in achieving high prediction accuracy when the number of search targets is large, as they often rely on click logs with low search frequency, leading to insufficient training data for targets with low search frequency.
Innovation Solution
An information processing device and method that utilize target data with identification and attribute information to update learning model parameters through multiple training tasks, including adapter modules and an adapter fusion layer, allowing efficient fine-tuning and high prediction accuracy even with a large number of search targets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If click logs are used for training data, then the learning model can be trained with real search behavior data, but the prediction accuracy deteriorates when the number of search targets is large due to insufficient training data for low-frequency targets
Solution Approach 1:
The system performs preliminary action by generating synthetic click logs before actual searches occur. The attribute prediction model and category prediction model are pre-trained using artificially generated training data that simulates user search behavior. This allows the models to learn from sufficient data even when actual click logs for certain targets are scarce, thereby maintaining high prediction accuracy across all search targets including low-frequency ones.
2Adaptability or versatility
If the number of search targets is increased, then the search system becomes more comprehensive, but the prediction accuracy deteriorates due to insufficient training data for each individual target
Solution Approach 1:
The system applies universality by creating a multi-functional attribute prediction model that can handle diverse search targets across different categories simultaneously. The model is trained on synthetic data covering multiple product categories and attributes, enabling it to generalize well to any search target regardless of its frequency. This allows the system to maintain high prediction accuracy even when supporting a large number of different search targets with varying data availability.
3Loss of information
If actual click logs are used for training, then the training data reflects real user behavior, but data loss occurs for low-frequency search targets leading to reduced model performance
Solution Approach 1:
The system uses copying by generating synthetic click logs that replicate the structure and characteristics of actual user search behavior. Instead of relying solely on sparse real click logs, the system creates artificial copies of training data through simulation that preserve the essential patterns of user interactions. This synthetic data copying ensures that even low-frequency search targets have sufficient training examples, preventing data loss and maintaining model performance across all targets.
Data Source
AI summary
An information processing device that generates a learning model includes a processor and a memory. The memory stores target data. The learning model is configured to output attribute information of a target based on a search query that has been input to search for the target. The processor is configured to execute a process that updates some of parameters included in the learning model by giving the target data to one or more training tasks and executing the one or more training tasks.