Two-Transformer Search Ranking with Offline History Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current web search systems face challenges in accurately ranking search results in real time due to the need for extensive input data, which can overwhelm computational resources and lead to user-perceivable delays.
Innovation Solution
A two-transformer-based machine-learning model architecture is employed, where a first model generates a vector representation of larger historical data offline, and a second model determines relevance parameters using both recent and larger historical data to optimize computational efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If more input data is provided to the transformer-based ML model to improve ranking accuracy, then the accuracy of search results ranking is improved, but the computational load increases and real-time operation becomes challenging
Solution Approach 1:
The system segments the historical data into two distinct components: larger historical data (processed offline) and recent historical data (processed online). This segmentation allows the model to utilize comprehensive data for accuracy while managing computational resources efficiently in real-time operations.
Solution Approach 2:
The system performs preliminary processing of larger historical data offline before real-time search operations. By pre-processing and storing embeddings of larger historical data, the system reduces the computational burden during online operations, enabling real-time performance without sacrificing accuracy.
2Measurement precision
If more input data is processed in real time to improve ranking accuracy, then the relevance determination is improved, but user-perceivable delays occur
Solution Approach 1:
The system performs computationally intensive processing of larger historical data in advance (offline) before users submit queries. The processed embeddings are stored and readily available during real-time operations, eliminating delays that would otherwise occur during online processing.
Solution Approach 2:
The system maintains continuous improvement of ranking accuracy by continuously updating and refining the model with recent historical data while leveraging the continuously available offline-processed larger historical data, ensuring both accuracy and responsiveness.
3Measurement precision
If extensive input data is used for training the ML model to improve accuracy, then the ranking effectiveness is improved, but server computational resources are overwhelmed
Solution Approach 1:
The system divides data processing into offline and online segments, with larger historical data processed offline using available computational resources, and only lightweight recent historical data processed online. This segmentation prevents server resources from being overwhelmed during real-time operations.
Solution Approach 2:
The system performs extensive data processing and model training in advance during offline operations when computational resources can be fully utilized. The processed results are then deployed to servers, reducing the real-time computational requirements and preventing resource overload during active service.
Data Source
AI summary
A method and system for training a machine-learning algorithm (MLA), including first and second machine-learning (ML) models to rank digital documents at a digital platform. The method comprises: receiving first historical data of past searches conducted by the given user over a given past period; receiving second historical data, a given portion of the second historical data including data of past searches conducted by the given user over a given past user session; jointly training both the first and second ML models to rank the in-use digital documents, the training comprising: training, based on the first historical data, the first ML model to generate a vector representation of the first historical data; and training, based on the vector representation of the first historical data and the second historical data, the second ML model to determine a respective likelihood value of the given user interacting with a given in-use digital document.