Embedding-Based Query Clustering to Enrich Sparse Retrieval Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional online concierge systems face limitations in refining item retrieval results due to varying amounts of user interactions with different queries, leading to insufficient data for queries with limited interactions.
Innovation Solution
An online system generates embeddings for queries based on similarity measures, clusters similar queries, and combines interaction data across clusters to enhance the amount of stored information, allowing for improved item retrieval and display optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional online concierge systems maintain separate interaction data for different queries, then data storage and query processing are simpler, but the amount of interaction data available for refining results is insufficient for queries with limited interactions
Solution Approach 1:
The patent merges interaction data from multiple queries by clustering queries with similar embeddings. Interaction data from queries within the same cluster are combined, allowing the system to leverage data from semantically similar queries to improve result refinement for queries with limited direct interaction data.
Solution Approach 2:
The patent introduces embeddings as an intermediary representation to bridge between queries. Queries are transformed into embedding vectors, which serve as mediators for similarity comparison and clustering, enabling the system to identify and combine interaction data from semantically related queries without directly comparing query texts.
2Measurement precision
If the system combines interaction data from similar queries, then the accuracy of item retrieval is improved, but the computational complexity increases due to embedding generation and clustering operations
Solution Approach 1:
The patent performs preliminary actions by pre-computing and storing embeddings for queries and pre-clustering queries based on embedding similarity. This allows the system to avoid repeated computation during item retrieval operations, as the clustering structure is already established and can be directly utilized for data combination.
Solution Approach 2:
The patent uses embedding vectors as copies or representations of the original queries. Instead of working directly with the full query texts and their associated interaction data for every comparison operation, the system works with compact embedding representations, reducing computational overhead while preserving semantic information.
Data Source
AI summary
An online system leverages stored interactions with items made by users after the online system received queries to determine display of items satisfying the query. For example, the online system trains a model to predict a likelihood of a user performing an interaction with an item displayed after a query was received. As different items receive different amounts of interaction from users, limited historical interaction with certain items may limit accuracy of the model. The online system generates embeddings for previously received queries and uses measures of similarity between embeddings for queries to generate clusters of queries. Previous interactions with queries in a cluster are combined, with the combined data being used for determining display of items in response to a query.