Self-Adversarial Negative Sampling for Search Relevance Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing online systems face challenges in accurately determining the relevance of items to queries due to limited or noisy human-labeled data, leading to suboptimal performance in identifying relevant items in search results.
Innovation Solution
An online system employs self-adversarial negative sampling and negative-sample sharing to generate and optimize negative training examples, using a combination of query and item encoders with a fusion layer to determine relevance, and applies weighted error terms based on similarity measures during model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional training methods with limited human-labeled data are used, then the model training is computationally feasible, but the model accuracy in determining relevance between queries and items deteriorates
Solution Approach 1:
The system performs preliminary action by generating synthetic negative training examples through adversarial sampling before the main training process. This pre-generation of training data allows the model to learn from a more comprehensive dataset without requiring proportional computational resources during training, as the difficult work of creating diverse training scenarios is done in advance through the adversarial process.
Solution Approach 2:
The system introduces an intermediary component - the adversarial sampler - that generates synthetic negative examples to bridge the gap between limited human-labeled data and the need for comprehensive training data. This intermediary transforms the training process by creating additional learning signals without requiring proportional increases in computational resources.
2Reliability
If comprehensive negative training examples are generated to improve model training, then the model performance improves, but the computational complexity and memory requirements increase
Solution Approach 1:
The system segments the training data into positive human-labeled examples and synthetic negative examples generated through adversarial sampling. This segmentation allows the model to process different types of training examples through optimized pathways, reducing overall computational complexity while maintaining comprehensive training coverage.
Solution Approach 2:
The system changes parameters by dynamically adjusting the sampling distribution during adversarial generation based on current model predictions. This parameter adaptation allows the system to focus computational resources on generating the most informative negative examples, improving model performance without linearly increasing computational complexity.
3Measurement precision
If all generated negative training examples are used for training, then the model learns from diverse data, but the training time and computational load increase significantly
Solution Approach 1:
The system applies partial action by selectively using a subset of generated negative training examples rather than processing all possible examples. The adversarial sampler generates examples with varying degrees of difficulty and informativeness, and the system strategically selects the most valuable portions for training, achieving effective learning without the full computational burden of processing every generated example.
Data Source
AI summary
To train an embedding-based model to determine relevance between items and queries, an online system generates training data from previously received queries and interactions with results for the queries. The training data includes positive training examples including a query and an item with which a user performed a specific interaction after providing the query. To generate negative training examples for the query to include in the training data, the online system determines measures of similarity between items with which the specific interaction was not performed and the query. The online system may weight a loss function for the embedding-based model by the measure of similarity for a negative example, increasing the effect of a negative example including a query and an item with a larger measure of similarity. In other embodiments, the online system selects negative training examples based on the measures of similarities between items and queries in pairs.