Negative Training Example Generation via Query Proximity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning algorithms face inefficiencies in generating effective negative training examples, which are crucial for training supervised learning models, as random negative examples often fail to accurately represent non-relevant data, leading to suboptimal performance in predicting relevance.
Innovation Solution
The method employs a query proximity approach to identify and generate negative training examples by analyzing search queries with high textual overlap but low result overlap, utilizing user interaction data to determine similarity scores and select appropriate query pairs for training, thereby improving the accuracy of negative example generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If random negative examples are used for training, then the training process is simple and fast, but the prediction accuracy deteriorates because random examples fail to accurately represent non-relevant data
Solution Approach 1:
The patent changes the parameters of negative examples from random selection to query-proximity-based selection. By modifying how negative examples are generated (using query similarity metrics instead of random sampling), the system maintains training efficiency while significantly improving prediction accuracy. The query proximity approach transforms negative examples into more representative non-relevant data points that better reflect actual search behavior patterns.
Solution Approach 2:
The patent introduces query similarity metrics as an intermediary mechanism between the training data generation process and the final prediction accuracy. This intermediary layer (query proximity analysis) acts as a bridge that transforms simple random examples into sophisticated, behavior-aligned negative examples without requiring complex manual curation, thus maintaining productivity while improving precision.
2Measurement precision
If query proximity approach is used to generate negative examples, then the prediction accuracy improves, but the device complexity increases due to additional analysis steps
Solution Approach 1:
The patent makes the query analysis system multi-functional by using the same query proximity metrics for both generating negative training examples and for the actual relevance prediction task. This universal approach means that one set of computational mechanisms serves dual purposes: training data generation and operational prediction, thereby reducing overall system complexity despite the advanced analytics involved.
Solution Approach 2:
The system performs self-service by using its own query analysis capabilities to automatically generate its training data. The query proximity metrics that will be used for prediction are the same metrics used to identify suitable negative examples, creating a self-contained training pipeline that reduces external dependencies and simplifies the overall system architecture.
3Reliability
If high-quality negative examples are generated through query proximity analysis, then the relevance prediction performance improves, but the processing time increases due to additional computational steps
Solution Approach 1:
The patent applies preliminary action by pre-computing query similarity metrics and identifying negative examples in advance during the training phase. By performing the computationally intensive query proximity analysis beforehand rather than in real-time during prediction, the system achieves high prediction reliability without incurring time penalties during operational use. The heavy lifting is done upfront, allowing fast predictions later.
Data Source
AI summary
There is disclosed a computer-implemented method and system for generating a set of training objects for training a machine learning algorithm (MLA) to determine query similarity based on textual content thereof, the MLA executable by the system. The method comprises retrieving, from a search log database of the system, a first query and other queries with associated search results. The method then comprises selecting a subset of query pairs such that: a query difference in queries in the pair is minimized and a results difference in respective search results is maximized.