Techniques for Fast Query Relaxation in Content-Based Recommender Systems

Abstract

‘Query relaxation’ is one of the basic approaches to deal with unfulfillable or conflicting customer requirements in content-based recommender systems: When no product in the catalog exactly matches the customer requirements, the idea is to retrieve those products that fulfill as many of the requirements as possible by removing (relaxing) parts of the original query to the catalog. In general, searching for such an ‘maximum succeeding subquery’ is a non-trivial task because a) the theoretical search space exponentially grows with the number of the subqueries and b) the allowed response times are strictly limited in interactive recommender applications.

In this paper, we describe new techniques for the fast computation of ‘user-optimal’ query relaxations: First, we show how the number of required database queries for determining an optimal relaxation can be limited to the number of given subqueries by evaluating the subqueries individually. Next, it is described how the problem of finding relaxations returning ‘at-least-n’ products can be efficiently solved by analyzing these partial query results in memory. Finally, we show how a general-purpose conflict detection algorithm can be applied for determining ‘preferred’ conflicts in interactive relaxation scenarios.

The described algorithms have been implemented and evaluated in a knowledge-based recommender framework; the paper comprises a discussion of implementation details, experiences, and experimental results.