Recommendation Model Training Using Category Representation Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current recommendation algorithms face a trade-off between recommendation accuracy and diversity, where improving diversity often reduces accuracy and user experience, and vice versa, leading to issues like information cocooning.
Innovation Solution
A training method for a recommendation model that processes data to obtain category-independent and category-dependent representations, using a discriminator to adjust parameters based on loss values, effectively separating these representations to enhance both accuracy and diversity without adding complexity or noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If recommendation algorithms optimize for recommendation accuracy, then recommendation accuracy is improved, but recommendation diversity deteriorates
Solution Approach 1:
The patent segments the recommendation representation into two independent parts: category-independent representation (capturing user preferences) and category-dependent representation (capturing item category information). This segmentation allows the model to optimize for both accuracy (through category-independent features) and diversity (through category-dependent features) without interference between the two objectives.
Solution Approach 2:
The patent introduces a new dimensional perspective by adding category-dependent representation as an additional dimension to the traditional recommendation representation. This enables the model to simultaneously consider both category-independent user preferences and category-specific characteristics, resolving the accuracy-diversity trade-off through multi-dimensional feature integration.
2Adaptability or versatility
If recommendation algorithms optimize for recommendation diversity, then recommendation diversity is improved, but recommendation accuracy deteriorates
Solution Approach 1:
The patent segments the recommendation representation into two independent parts: category-independent representation (capturing user preferences) and category-dependent representation (capturing item category information). This segmentation allows the model to optimize for both accuracy (through category-independent features) and diversity (through category-dependent features) without interference between the two objectives.
Solution Approach 2:
The patent applies local quality by making different parts of the representation serve different functions: the category-independent representation focuses on user preference patterns (accuracy), while the category-dependent representation focuses on item category characteristics (diversity). Each part has specialized quality optimized for its specific purpose, achieving both accuracy and diversity simultaneously.
3Adaptability or versatility
If post-ranking algorithms like DPP and MMR are used to improve diversity, then recommendation diversity is improved, but device complexity increases
Solution Approach 1:
The patent performs preliminary action by extracting and separating category-dependent representations during the initial representation learning phase, rather than applying complex post-ranking algorithms later. This preliminary separation of category information enables diversity to be built into the recommendation foundation, eliminating the need for additional complex post-processing steps like DPP or MMR.
Solution Approach 2:
The patent takes out the category-dependent information from the mixed representation and extracts it as a separate component. This extraction allows the model to utilize category information for diversity enhancement without incorporating the complexity of post-ranking algorithms, achieving diversity through a simpler, integrated approach.
4Adaptability or versatility
If rectification algorithms remove category features to avoid popular articles, then recommendation diversity is improved, but loss of information increases
Solution Approach 1:
The patent takes out the category-dependent representation as a separate, preserved component rather than removing category features. This extraction allows the model to maintain category information for diversity enhancement while avoiding the information loss associated with removing category features entirely, as the category representation is preserved and utilized in the recommendation process.
Solution Approach 2:
The patent changes the parameter approach by transforming category information from a potentially harmful factor (that causes popular article recommendations) into a useful dimension (category-dependent representation) that can be controlled and utilized selectively. This parameter transformation allows diversity enhancement without information loss by precisely controlling how category information is weighted and applied.
Data Source
AI summary
The present disclosure relates to a recommendation model training method, an article recommendation method and system, and a related device. The recommendation model training method includes: processing data for training by using a recommendation model to obtain a category-independent representation and a category-dependent representation, wherein the data for training comprises a feature of a user and a feature of an article; processing the category-independent representation and the category-dependent representation respectively by using a discriminator to obtain discrimination results; determining a prediction result according to at least one of the category-independent representation or the category-dependent representation; and training the recommendation model and the discriminator according to training targets comprising the category-independent representation not corresponding to any one of the plurality of categories, the category-dependent representation corresponding to the pre-marked category, and the prediction result matching with the pre-marked recommendation information.