Search Ranking via Machine Learning Models and Weak Supervision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing search algorithms on websites often return irrelevant results, causing users to spend time filtering through non-relevant items before finding what they are looking for, and current methods require extensive human-labeled training data, limiting the scalability of machine learning models.
Innovation Solution
Applying trained machine learning models, specifically relevance and engagement models, to determine item relevance to a search query using weak supervision, which generates semantic relevance scores to rank items based on their attributes and user engagement, allowing for real-time, more accurate search result ranking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional search algorithms are used to return search results, then the search process is simple to implement, but the relevance of search results to user intent deteriorates
Solution Approach 1:
The search algorithm is segmented into multiple independent machine learning models (relevance model, engagement model, semantic relevance model) that each handle specific aspects of search result ranking. This allows the system to improve overall relevance by combining specialized models rather than using a single complex algorithm.
Solution Approach 2:
Semantic relevance scores are introduced as an intermediary component that bridges the gap between user queries and search results. These scores serve as a mediator that enhances the connection between search intent and relevant items, improving overall search result relevance.
2Measurement precision
If machine learning models are trained with extensive human-labeled training data to improve search result accuracy, then the measurement precision of item relevance improves, but the time and resources required for data labeling worsen
Solution Approach 1:
The system uses weak supervision to enable machine learning models to self-train using automatically generated labels from search query logs and user behavior data. This eliminates the need for extensive manual human labeling, as the system generates its own training data from operational logs.
Solution Approach 2:
The approach changes the training data parameters from requiring high-quality human-labeled data to using large volumes of automatically logged search queries and user interactions. This parameter change allows training on much larger datasets without proportional increases in human labeling time.
3Productivity
If search results are ranked to prioritize more relevant items, then the user experience and purchase likelihood improve, but the complexity of the ranking system worsens
Solution Approach 1:
Multiple machine learning models (relevance model, engagement model, semantic relevance model) are merged into a unified ranking system that combines their outputs to determine final search result rankings. This merging allows the system to leverage multiple factors for improved search efficiency while managing complexity through integrated architecture.
4Quantity of substance
If the system processes and ranks a large number of search results in real-time, then the comprehensiveness of search results improves, but the processing time worsens
Solution Approach 1:
Machine learning models are trained in advance on large datasets, and semantic relevance scores are pre-computed for items. This preliminary action allows the system to quickly retrieve and rank search results in real-time without performing extensive computations during the actual search operation.
Data Source
AI summary
This application relates to apparatus and methods for applying trained machine learning models to determine an item's relevance to a search query. In some examples, a query and data identifying a plurality of items are received. Item attributes for each of the plurality of items are obtained, and features are generated based on the item attributes. Further, a score is generated for each item by applying a trained machine learning model to the corresponding features and the query. Matching attributes are determined for each of the plurality of items based on the corresponding item attributes and the query, and the score of each of the plurality of items is adjusted based on the matching attributes. Further, the ranking data is generated based on the adjusted score of each of the plurality of items. The ranking data may be transmitted to a web server for display of the items.