Weighted Attribute Combinations Based Similarity Measures

In this paper we introduce three parameterized similarity measures which take into account not only the single features of two objects under comparison, but also all the significant combinations of attributes. In this way a great expressive power can be achieved and field expert knowledge about relations among features can be encoded in the weights assigned to each combination. Here we consider only binary attributes and, in order to face the difficulty of weights' elicitation, we propose some effective techniques to learn weights from an already labelled dataset. Finally, a comparative study of classification power with respect to other largely used similarity indices is presented.