Product Recommendation System Using Multi-Dimensional Similarity Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing client-based recommender systems using collaborative filtering algorithms suffer from a lack of diversity in recommendations, often repeatedly suggesting the same handful of products due to limited methods of analyzing similarity among user history and recommended items.
Innovation Solution
An information-push method and apparatus that filter historical records to determine associations between products using a predefined similarity model, eliminating non-relevant items and improving recommendation diversity and accuracy by considering multiple aspects of product similarity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If collaborative filtering algorithm is used to recommend products based on user history, then recommendation accuracy is improved, but recommendation diversity deteriorates
Solution Approach 1:
The patent segments the recommendation process into multiple independent similarity calculation dimensions (category similarity, tag similarity, attribute similarity, etc.) rather than using a single collaborative filtering approach. Each dimension is calculated separately and then integrated to produce the final recommendation, thereby maintaining both accuracy and diversity
Solution Approach 2:
The patent introduces multiple similarity calculation dimensions beyond traditional collaborative filtering. Instead of relying solely on user-behavior-based similarity, it adds product-category similarity, product-tag similarity, product-attribute similarity, and other dimensional analyses to expand the recommendation space and improve diversity
2Device complexity
If limited methods are used to analyze similarity among products, then computational complexity is reduced, but recommendation diversity deteriorates
Solution Approach 1:
The patent divides the similarity analysis into multiple independent modules (category similarity module, tag similarity module, attribute similarity module, etc.), each handling a specific aspect. This segmentation allows the system to implement comprehensive diversity analysis without overwhelming computational complexity, as each module can be optimized independently
Solution Approach 2:
The patent performs preliminary classification and tagging of products before the recommendation process. By pre-organizing products into categories, tags, and attributes, the system reduces the computational burden during real-time recommendation generation, enabling diverse multi-dimensional analysis without excessive complexity
3Productivity
If same handful of products are repeatedly recommended, then processing efficiency is improved, but recommendation diversity deteriorates
Solution Approach 1:
The patent implements a dynamic recommendation strategy where the system adaptively adjusts which products to recommend based on user feedback, context, and multiple similarity dimensions. Instead of statically recommending the same top products, it dynamically selects from a broader pool by integrating results from multiple similarity calculations, maintaining efficiency while improving diversity
Solution Approach 2:
The patent changes the parameters used in recommendation by incorporating multiple similarity metrics (category similarity weight, tag similarity weight, attribute similarity weight, etc.) instead of relying on a single collaborative filtering score. This allows the system to efficiently retrieve results while presenting a more diverse set of recommendations by adjusting and balancing multiple parameters
Data Source
AI summary
A method and apparatus is described to push to-be-recommended products onto users' terminals based on similarities in the to-be-recommended products and products in the users' historical records of products. To improve the pushed recommendations various technics are applied, including: using predefined criteria to filter products provided from the users' historical records, applying temporal decay to provide real-time recommendations, removing already obtained products from the to-be-recommended products to be pushed, improving similarity calculations by determining similarity using pairwise variations between products both in and out of the candidates for to-be-recommended products, sorting the to-be-recommended product according to a preset recommendation policy, calculating pairwise variations using multiple types of similarity measures, using weights to determine the to-be-recommended products, and applying a preset recommendation policy based on a given scenario. Accordingly, the embodiments described herein avoid an excessive recommendation phenomenon (e.g., products previously obtained) and provide improved recommendations.