Personalized Item Retrieval Using Two-Tower Embedding Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional retrieval systems struggle to identify relevant items based on a small number of input characters, leading to increased user effort and resource consumption due to large candidate item inventories and inefficient relevancy scoring.
Innovation Solution
A two-tower deep machine learning model is employed, comprising a context tower that processes user and query features, and an item tower that processes item features, to calculate relevancy scores and select or rank candidate items effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional retrieval systems process large candidate item inventories with limited input characters, then they can handle partial queries, but they consume excessive computational resources and fail to identify relevant items accurately
Solution Approach 1:
The system segments the candidate item inventory into multiple batches and processes them in parallel. Each batch is scored independently using the two-tower model, allowing distributed computation that reduces the computational burden on any single processing unit while maintaining accurate relevancy scoring across the entire inventory.
Solution Approach 2:
The system performs preliminary filtering of candidate items based on basic query matching criteria before applying the computationally intensive two-tower model. This preliminary action reduces the number of items that require full relevancy scoring, thereby reducing overall computational resource consumption while preserving accuracy for the most promising candidates.
2Measurement precision
If conventional systems require more input characters to achieve accurate retrieval, then relevancy scoring improves, but user effort and query input time increase
Solution Approach 1:
The system applies partial action by processing only a subset of candidate items in detail using the two-tower model, while applying lighter weighting or filtering to the remainder. This allows accurate identification of relevant items without requiring the user to input complete queries, as the system can achieve satisfactory results with partial processing of the candidate inventory.
Solution Approach 2:
The system dynamically adjusts the number of candidate items processed and the depth of analysis based on query characteristics and available resources. By changing parameters such as batch size, processing depth, and confidence thresholds, the system maintains high identification accuracy while adapting to varying user input lengths and resource availability.
3Measurement precision
If the system processes all candidate items with full relevancy scoring, then retrieval accuracy improves, but processing speed and system responsiveness decrease
Solution Approach 1:
The candidate item inventory is divided into multiple segments or batches that are processed in parallel. Each segment undergoes full relevancy scoring using the two-tower model, ensuring high retrieval accuracy within each segment. The parallel processing architecture maintains overall processing speed by distributing the computational load across multiple processing units.
Solution Approach 2:
The system implements periodic action by processing candidate items in successive batches or waves rather than all at once. Each batch is scored and ranked, with the highest-scoring items from each batch potentially being returned early. This periodic processing maintains retrieval accuracy while improving system responsiveness by providing results in stages.
Data Source
AI summary
Disclosed are system, method and/or computer program product embodiments that retrieve items for a user based on a query using a two-tower deep machine learning model. An example embodiment provides input to a context tower, wherein the input includes the query and one or more of a query embedding corresponding to the query or a graph user embedding corresponding to the user. The context tower generates a context embedding in a vector space based on the input. The model determines a measure of similarity between the context embedding and each of a plurality of item embeddings in the vector space that are generated by an item tower and represent a plurality of candidate items. A relevancy score is calculated for each candidate item based on the measure of similarity between the context embedding and the corresponding item embedding. The relevancy scores are used for item retrieval and/or ranking.