Learning and engineering similarity functions for business recommenders

Abstract

We study the optimisation of similarity measures in tasks where the computation of similarities is not directly visible to end users, namely clustering and case-based recommenders. In both, similarity plays a crucial role, but there are also other algorithmic components that contribute to the end result. Our suggested approach introduces a new form of interaction into these scenarios that make the use of similarities transparent to end users and thus allows to gather direct feedback about similarity from them. This happens without distracting them from their goal – rather allowing them to obtain better and more trustworthy results by excluding dissimilar items. We then propose to use the feedback in a way that incorporates machine learning for updating weights and decisions of knowledge engineers about possible additional features, based on insights derived from a summary of user feedback. The reviewed literature and our own previous empirical investigations suggest that this is the most feasible way – involving both machine and human, each in a task that they are particularly good at.