Search Model

With respect to the search mechanism, we first plan to investigate alternative techniques for using item similarity values to determine which to return, for example by cutting off items at the point at which similarity drops off most steeply, instead of our current use of a threshold. We also note that work such as that of [23] on learning to rank instances could apply nicely to this work, augmenting our current item ranking scheme. Additionally, we plan develop a version of the system that generates alternative items or values in an over-constrained situation [60]. One way to do this would be to use the preferences to estimate the strength of a stated constraint, or to merge our preference-based similarity metric with a more traditional domain-specific similarity metric [59]. We also plan to evaluate the effect of making even stronger assumptions about user preferences. For example, if the system is certain enough about a value preference, it may not have to ask a question about the associated attribute. A final improvement of the search mechanism concerns the techniques for ranking attributes for constraining and relaxing. For attribute constraint ranking, we have implemented but not yet evaluated a conditional entropy measure [37]. The system selects the attribute to constrain by determining the attribute with the highest conditional entropy among the unconstrained attributes. This scheme would not be useful for ranking attributes to relax. Therefore, the system simply determines the size of the case base that would result if each attribute were relaxed, ranks these case bases from smallest to largest, and orders the attributes accordingly, excluding those attributes that, if relaxed, would still result in an empty case base. We also plan to investigate the combination of the user model with information gain, as well as with alternative attribute ranking techniques such as the one used by [2]. Another option is to add personalization to or otherwise adapt the variable selection techniques used by constraint-satisfaction solvers.