Recommendation Recall via Category Identifier Recognition Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation systems face challenges in improving the recall effect, particularly in dealing with new and cold objects, which leads to a low recall rate and the cold start problem.
Innovation Solution
A data processing method that involves determining user features and category identifiers to train a recognition model. This model is used to identify second category identifiers matching target users, constructing recall candidate sets, and determining target objects for recommendation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional recommendation systems use recall stages to select objects from massive data sets, then the system can handle large scale data, but the recall effect is poor especially for new and cold objects leading to low recall rates
Solution Approach 1:
The patent introduces category identifiers as intermediary elements that bridge users and objects. Instead of directly matching users with objects from massive data sets, the system first identifies relevant category identifiers based on user features, then retrieves objects associated with these identifiers. This intermediary approach enables effective recall even for new and cold objects by leveraging their category associations rather than relying solely on direct user-object interaction history.
Solution Approach 2:
The system performs preliminary action by pre-establishing category identifiers and their associations with objects before the actual recommendation process. Category identifiers are determined in advance based on object features and relationships, creating a pre-organized structure that enables rapid and accurate recall when user features are provided, without requiring extensive real-time computation or prior user interactions with specific objects.
2Measurement precision
If the recommendation system focuses on hot objects with existing interaction data, then the recall accuracy improves, but new and cold objects are neglected causing the cold start problem
Solution Approach 1:
The patent segments the object space into multiple categories using category identifiers, allowing the system to handle different types of objects (hot, cold, and new) separately and systematically. By dividing the massive object data set into category-based segments, the system can apply targeted recall strategies for each segment, ensuring that new and cold objects are not overlooked while maintaining accurate recall for hot objects through their shared category structures.
3Quantity of substance
If the system uses massive object data sets for recommendation, then the coverage of recommended objects increases, but the quality of selected objects deteriorates making it difficult to find matching objects
Solution Approach 1:
The system extracts key features from user data and matches them with category identifiers, then extracts objects associated with these identifiers. This extraction approach allows the system to work with massive object data sets by selectively pulling out relevant objects based on feature matching, rather than attempting to evaluate all objects. The recognition model extracts essential characteristics to identify high-quality matches from the larger pool of candidates.
Data Source
AI summary
A data processing method, an object processing method, a recommendation method, and a computing device are provided. At least one user feature of sample user(s) is determined. First category identifier(s) corresponding to sample object(s) matching the sample user(s) is/are determined. A recognition model is trained using at least one user feature of the sample user(s) and the first category identifier(s), wherein the recognition model is used to determine at least one second category identifier matching a target user based on at least one user feature of the target user, the at least one second category identifier is used to construct a recall candidate set corresponding to the target user, the recall candidate set includes at least one object hit by the at least one second category identifier, and is used for determining at least one target object for performing a recommendation operation to the target user.