Rule-Based Similarity for Classification

This paper presents an ongoing research on the problem of assessing a similarity between objects in the context of classification. A new model of similarity is presented, called Rule-based Similarity (RBS), in which the similarity is expressed in terms of higher-level binary features of objects. Those features may be associated with decision rules derived from data and can be interpreted as arguments for a similarity or for a dissimilarity of the examined objects. The model was motivated by the feature contrast model of Amos Tversky. Its main aim is to simulate the human way of perceiving similar objects and at the same time to achieve a high accuracy in real life classification tasks. The partial results of conducted experiments confirm that the RBS is an interesting alternative to the commonly used distance-based similarity models.