Recommendation Evaluation Using Orthogonal User Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation information systems, such as those described in Patent Document 1, often omit relevant content recommendations due to their reliance on predetermined extraction rules and direct user preference data, failing to provide users with necessary information, especially when browsing websites that encompass multiple types of content.
Innovation Solution
A recommendation information evaluation apparatus and method that generates a user characteristic vector orthogonal to the feature space separation plane between selected and non-selected items, allowing for accurate evaluation of unknown items without direct user preference data, using techniques like SVM and NN methods, and weighting based on selection frequency and recency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If recommendation information is extracted by matching processing based on predetermined extraction rules, then the system can operate with simple rules, but recommendation information omissions occur
Solution Approach 1:
The patent replaces the mechanical matching processing system with a neural network-based inference system. Instead of using predetermined extraction rules to match keywords, the system uses a neural network that processes user profile information and website content vectors to infer appropriate recommendation information, thereby eliminating information omissions while maintaining operational simplicity through automated learning.
Solution Approach 2:
The patent changes the parameter representation from discrete keyword matching to continuous vector space representation. User profiles and website contents are represented as vectors in a high-dimensional space, allowing the neural network to capture nuanced relationships and make accurate recommendations without relying on predetermined rules, thus preventing information omissions.
2Quantity of substance
If recommendation information is distributed based on access history, then the system can utilize user behavior data, but relevant content recommendations are omitted
Solution Approach 1:
The patent introduces a neural network as an intermediary between user access history and recommendation information. The neural network processes access history data and website content vectors through hidden layers, transforming raw behavioral data into meaningful recommendation insights. This intermediary processing enables the system to discover relevant content that would otherwise be omitted in direct matching approaches.
Solution Approach 2:
The patent transitions from two-dimensional keyword matching to high-dimensional vector space processing. By representing user profiles, access history, and website contents as vectors in a high-dimensional space, the neural network can capture complex patterns and relationships that go beyond simple access history analysis, thereby recommending relevant content without omissions.
3Productivity
If the browsed Web site is grasped in its entirety, then the system can process all content, but specialized filtering for specific topics is lost
Solution Approach 1:
The patent segments the website content into distinct topic areas and categories, represented as separate vectors in the feature space. The neural network processes these segmented representations to identify which topics are most relevant to the user's profile and access history. This segmentation allows the system to efficiently process all content while simultaneously providing specialized filtering and targeted recommendations for specific topics of interest.
Data Source
AI summary
A content evaluation apparatus and a content evaluation method that can distribute contents that a user potentially needs without any omissions. A history class separation unit generates a separation plane for separating in a separation plane a characteristic vector of a selected content and characteristic vectors of non-selected contents stored in a browsing history table. A user characteristic vector calculation unit generates a user characteristic vector based on an orthogonal vector perpendicular to the separation plane thus generated and stores the user characteristic vector in a user characteristic vector management table. Subsequently, when a recommendation request receiving unit receives a recommendation request, a content evaluation unit evaluates the contents based on the stored user characteristic vector.