Behavioral Filter for Personalized Recommendations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation systems face challenges in filtering user behavior to prevent the recommendation of items similar to those already purchased, especially when user activity data is incomplete or miscategorized, leading to ineffective personalized recommendations and inappropriate advertisements.
Innovation Solution
Implementing a related history filter that selectively filters seed behaviors based on user purchases, categorizing and adjusting category designations to exclude items from recommendations, and filtering user behavior across categories to prevent miscategorized items from being recommended.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If user behavior data from third-party sites is used to generate recommendations, then the quantity and diversity of recommendation data increases, but the accuracy and reliability of recommendations deteriorates due to incomplete or miscategorized data
Solution Approach 1:
The patent introduces a behavioral filter as an intermediary component that processes user behavior data from third-party sites before it reaches the recommendation engine. This filter validates and cleanses the data, removing incomplete or miscategorized entries, thereby maintaining data reliability while still utilizing the expanded data quantity from external sources.
Solution Approach 2:
The patent replaces manual data validation mechanisms with an automated behavioral filter system that uses predefined rules and algorithms to automatically detect and remove low-quality data. This substitution enables efficient processing of large volumes of third-party data while maintaining consistent quality standards without requiring human intervention.
2Reliability
If purchase history filtering is applied to exclude purchased items from recommendations, then the relevance of recommendations improves, but the complexity of the recommendation system increases
Solution Approach 1:
The patent implements purchase history filtering as a preliminary action that occurs before the main recommendation generation process. By pre-filtering out purchased items and recently viewed items from the candidate pool, the system simplifies the subsequent recommendation algorithm while ensuring that irrelevant items are never presented to users, thereby improving relevance without significantly increasing overall system complexity.
3Measurement precision
If strict filtering rules are applied to exclude all similar items, then the accuracy of recommendations improves, but the versatility of the recommendation system deteriorates
Solution Approach 1:
The patent applies local quality by implementing differentiated filtering rules for different item categories and user contexts. Instead of uniform strict filtering, the system adjusts filtering intensity based on category characteristics and user behavior patterns, allowing versatile recommendations in some areas while maintaining precision in others. This enables the system to adapt filtering strictness to local requirements rather than applying a one-size-fits-all approach.
Data Source
AI summary
This disclosure relates to selectively filtering seed behavior, e.g., user activity used to generate item recommendations, based on the availability of item classifications. Seed behaviors and catalog items may be associated with categories in an electronic catalog, and a particular seed behavior may be used to generate user recommendations if it is more recent than a user's last purchase in the category of the seed behavior. For example, a user's activity in the TV category, e.g., viewing various TV models, may not be used to generate recommendations if the activity occurred prior to the user's purchase of a TV. As a result, additional TVs may not appear in the user's recommendations following the purchase of a TV. However, if classification information is unavailable relating to the purchase, seed behavior may be filtered across a set of categories, not just the TV category to reduce the chance of less effective recommendations.