A Similarity-Based Approach to Attribute Selection in User-Adaptive Sales Dialogs

Cooperative Query Rewriting for Decision Making Support and Recommender Systems

“…The induced similarity-based ranking will rank equally all these exactly matching products. Conversational recommender systems are characterized by the fact that, when faced with these failure problems, they initiate an interaction with the user in order to refine the current query (Aha and Breslow 1997;Göker and Thomson 2000;Kohlmaier et al 2001;Gupta et al 2002;Shimazu 2001). These systems have mostly focused on the large results set problem, and when the user query returns too many results, these systems pose questions to the user to obtain additional preferences to further specify the current query.…”

“…These systems have mostly focused on the large results set problem, and when the user query returns too many results, these systems pose questions to the user to obtain additional preferences to further specify the current query. The questions are determined using feature-selection methods (Doyle and Cunningham 2000;Kohlmaier et al 2001;Shimazu 2001) that exploit information theory principles (e.g., information gain [Quinlan 1986;Witten and Frank 2000]), to identify the attributes with higher discriminatory power. In other words, they select a feature and generate a corresponding question.…”

“…Features with higher scores should be asked to the user because they are the most useful to identify the required product. In fact, many conversational systems (Kohlmaier et al 2001;Shimazu 2001) have applied this idea. But, in order to apply techniques similar to those used in the induction of decision trees, these systems identify a particular feature as the ''class'' of the product or consider each product as belonging to a different class.…”

“…Feature-selection methods for recommender systems have usually been evaluated with respect to the interactions length observed in simulated interactions (Aha and Breslow 1997; Kohlmaier et al 2001;Shimazu 2001). A typical simulation proceeds by first selecting a target product (test item in the FeatureSelectionEval() algorithm), and then searching for those products that are most similar to the target with respect to a subset of all the product features.…”

Supporting User Query Relaxation in a Recommender System

“…In previous researches on knowledge-based recommender systems the problem of dealing with the failure of an over-constrained query was typically addressed by exploiting similarity-based retrieval [7,8]. In Trip@dvice [2], as well as in other proposals [9], it is argued that the system should not autonomously determine the best attainable approximate match, as in a similarity-based retrieval, but should actively support the user and let him understand what is the best relaxation/compromise.…”