Introduction
Knowledge-based recommender systems, frequently employed for complex domains like financial services or high-end electronics, rely on explicit user requirements to filter catalogs. A common failure mode occurs when users specify mutually exclusive criteria—such as a high-return investment with zero risk and immediate liquidity. In these "no results found" scenarios, traditional systems often leave users stranded or provide "minimal diagnoses" that lack personalization. This paper introduces a novel approach that combines model-based diagnosis with collaborative filtering to suggest "plausible repairs": specific, individualized changes to user criteria that lead to items the user is actually likely to want.
Methodology: Personalized Repair via Guided Search
The core of the proposed solution is a "best-first" repair algorithm that integrates Model-Based Diagnosis (MBD) with similarity-based ranking. Unlike standard logic-based systems that prioritize the smallest number of changes (cardinality), this approach prioritizes repairs that lead to products most similar to the user's original intent.
The process involves four key steps:
- Nearest Neighbor Integration: The system identifies the $n$ items in the catalog most similar to the user's initial (albeit inconsistent) requirements using a similarity metric.
- Guided HSDAG Search: The algorithm utilizes a Hitting Set Directed Acyclic Graph (HSDAG) to identify conflicts. Rather than a standard breadth-first search, it employs a best-first search guided by the similarity of "supporting items."
- Similarity-Based Ranking: Repair paths are prioritized based on their proximity to high-similarity items identified in the first step.
- Conflict Identification: The QuickXPlain algorithm is used to efficiently find minimal conflict sets, ensuring the resulting repair suggestions are concise and actionable.
Empirical Evaluation and Performance
The authors evaluated their approach through a user study involving 493 subjects using a digital camera dataset (90 items) and a synthetic performance benchmark (5,000 items).
User Preferences
The study compared the personalized approach (Type 2) against a standard cardinality-based approach (Type 1). The personalized algorithm demonstrated significantly higher precision (0.551 vs. 0.307) and lower average deviation from the user's eventual choice (0.474 vs. 0.599). Users selected the top-ranked repair significantly more often when personalization was applied ($p < 0.001$).
Computational Efficiency
A critical requirement for interactive systems is real-time performance. Despite the underlying NP-hard nature of the problem, benchmark tests show that for typical recommender scales, personalized repairs can be calculated in under a second, making the approach viable for interactive web interfaces.
Discussion and Future Work
The primary insight of this research is that personalization is most valuable at the point of search failure. By shifting the focus from "minimal" repairs to "plausible" ones, systems can better navigate the trade-offs users are willing to make.
Future research directions include:
- Diversity and Selection Probability: Incorporating a wider range of repair types to avoid early pigeonholing of user preferences.
- Automated Weighting: Learning the relative importance of different requirements (e.g., "must-haves" vs. "nice-to-haves") to further refine similarity metrics.
- Scaling to Large Catalogs: While the current system handles thousands of items, extremely large e-commerce catalogs may require more aggressive search graph pruning.
Conclusion
By integrating logic-based diagnosis with similarity-based personalization, this approach transforms a frustrating "no results" dead-end into a guided navigation experience. This significantly improves user satisfaction and reduces the manual effort required to find suitable products in complex domains.
Summary updated