Query Similarity Prediction for Sparse Behavioral Signals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing search systems face challenges with sparse user behavioral signals, leading to unbalanced feature quality for ranking, especially for queries with limited interaction data, resulting in inefficient ranking models and hardware scalability issues.
Innovation Solution
The use of machine learning techniques to identify similar queries with better behavioral signals, employing a combination of representation-based and interaction-based models for initial filtering and re-ranking, respectively, to generate similarity scores and improve ranking models, thereby addressing the sparse signal issue and hardware limitations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning techniques use behavioral features for ranking, then ranking accuracy is improved, but the system fails when user signals are sparse
Solution Approach 1:
The patent introduces query similarity as an intermediary concept to transfer behavioral signals from well-served queries to sparsely-served queries. By identifying semantically similar queries through embedding models and transferring their behavioral features, the system can generate reliable ranking features even when direct user signals are sparse for the target query.
Solution Approach 2:
The system pre-computes query embeddings and identifies similar queries in advance to build a knowledge base of behavioral signals. This preliminary action allows the system to quickly retrieve and transfer relevant behavioral features when encountering queries with sparse signals, rather than computing everything in real-time.
2Measurement precision
If the system processes all queries with behavioral models, then ranking quality improves, but computational cost and hardware requirements increase
Solution Approach 1:
The patent applies different processing qualities to different queries based on their needs. Queries with sufficient behavioral signals undergo standard processing, while sparsely-served queries receive enhanced processing through signal transfer from similar queries. This local differentiation optimizes computational resources by applying complex models only where necessary.
Solution Approach 2:
Instead of applying full behavioral modeling to all queries, the system applies partial action by using query similarity transfer only for sparsely-served queries. This selective application reduces overall computational energy consumption while maintaining ranking quality where it matters most.
3Measurement precision
If the system stores extensive training data for all queries, then model performance improves, but data storage requirements become unsustainable
Solution Approach 1:
The patent creates virtual copies of behavioral signals by transferring them from similar queries. Instead of storing separate behavioral data for every possible query, the system copies relevant behavioral features from semantically similar queries, dramatically reducing storage requirements while maintaining model performance.
Solution Approach 2:
The query embedding space serves multiple functions: it enables similarity search, facilitates signal transfer, and provides a compact representation of query semantics. This universal representation approach replaces the need for extensive query-specific training data storage.
Data Source
AI summary
An improved search system can address sparse behavioral signals. Machine learning is used to determine queries with historical data that are similar to an input query, which have limited behavioral signals. Co-acquisitions of items for query pairs are used to generate a similarity value for the pairs, which are used as labels during training of machine learning models. Multiple machine learning models are used during inference such that a subset of similar queries is determined with a less computationally expensive model and a more computationally expensive model is used to re-rank the subset of queries. Accordingly, for input queries with limited behavioral data, the search system identifies similar queries with behavioral data and their corresponding search results can be presented to a user.