Least-Ask Recommender System Using Random Walk for Preference Assessment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional conversational recommender systems (CRS) face challenges in minimizing user interaction and quantitatively evaluating user experience, as they often require repeated interactions and fail to provide diverse recommendations that meet user preferences effectively.
Innovation Solution
The implementation of a Least-Ask mechanism within the CRS, utilizing a random walk algorithm to model user preferences and feedback, which minimizes the number of questions asked by re-ranking items based on user feedback and adjusting the recommendation strategy to ensure optimal product recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If navigation-by-asking strategy is adopted to collect user requirements, then the system can identify user preferences, but it would take an infinite time to navigate the user to all products
Solution Approach 1:
The system pre-computes a product graph structure and similarity relationships between products before user interaction. This preliminary preparation allows the system to quickly navigate and recommend products without having to ask users about all products sequentially, resolving the time consumption issue while maintaining preference assessment accuracy.
Solution Approach 2:
The patent replaces the mechanical navigation-by-asking process with a computational recommendation engine that uses a product graph and similarity algorithms. Instead of systematically asking users about all products (mechanical approach), the system uses computational methods to predict and recommend products based on user preferences and product relationships.
2Productivity
If top k items are presented based on preference estimation, then the system provides recommendations quickly, but the recommended products may be similar to each other and far away from an optimal product
Solution Approach 1:
The system introduces a new dimension of product similarity space by constructing a product graph where products are nodes and edges represent similarity relationships. This graph-based dimension allows the system to navigate the product space more effectively, finding diverse yet relevant recommendations rather than relying solely on traditional top-k ranking methods.
Solution Approach 2:
The patent dynamically adjusts the recommendation parameters by using a product graph traversal approach that considers both similarity and diversity. The system changes from static top-k selection to dynamic graph-based navigation, where the recommendation parameters (which products to recommend) are adjusted based on the user's preferences and the graph structure, improving both accuracy and diversity.
3Manufacturing precision
If diversity-oriented CRS is used to overcome imprecision in preference assessment, then the system provides diverse recommendations, but it fails to quantitatively evaluate the user experience
Solution Approach 1:
The system implements a feedback mechanism where user interactions with recommended products (clicks, purchases, ratings) are collected and used to update the product graph and refine future recommendations. This feedback loop allows the system to quantitatively evaluate user experience by measuring engagement metrics and using them to improve the recommendation algorithm, thus recovering the lost information about user satisfaction.
Data Source
AI summary
A method for a conversational recommender system includes receiving at least one user preference from a user, generating an abstract explanation reflecting the at least one user preference and based on the abstract explanation, and retrieving a list of items to be evaluated by the user from an item database. The items on the list have a relevance with the user preference. The method further includes determining an initial ranking order of the items to be evaluated by the user, and ranking the items to be evaluated by the user based on a random walk algorithm and the initial ranking order. Items not ranked are remaining items. The method further includes receiving a user feedback of the ranked item, and re-ranking the remaining items based on the user feedback of the ranked items and the random walk algorithm.