Similarity-based network models and how evaluate them

Similarity-based network models have been used in many data mining tasks, such as classification and clustering. These models are applied in non-relational data, where each example is represented by a vector of characteristics, creating a relational representation based on similarity among the examples. Using this representation, it is possible to use complex network measures, in a relational data mining context, incorporating more information than the traditional propositional data mining. In this technical report we present a new network model based on similarity, the Extended Minimum Spanning Tree network (EMST), and three measures to evaluate the network models, based on neighborhood, clusters and outliers preservation. The model proposed here is non-parametric and present good results for all evaluation measures when compared with the other models.