Finding Preferred Query Relaxations in Content-Based Recommenders

Summary

In many content-based approaches to product recommendation, the set of suitable items is determined by mapping the customer’s needs to required product characteristics. A ‘failing query’ in that context corresponds to a situation in which none of the items in the catalog fulfills all of the customer requirements and in which no proposal can be made. ‘Query relaxation’ is a common technique to recover from such situations which aims at determining those items that fulfill as many of the constraints as possible. This chapter proposes two new algorithms for query relaxation, which aim at resolving common shortcomings of previous approaches. The first algorithm addresses the problem of response times for computing user-optimal relaxations in interactive recommendation sessions. The proposed algorithm is based on a combination of different techniques like partial evaluation of subqueries, precomputation of query results and compact in-memory data structures. The second algorithm is an improvement of previous approaches to mixed-initiative failure recovery: Instead of computing all minimal ‘conflicts’ within the user requirements in advance – as suggested in previous algorithms – we propose to determine preferred conflicts ‘on demand’ and use a recent, general-purpose and fast conflict detection algorithm for this task.¹