Modelo de segmentação Clusterwise com protótipos híbridos

Abstract

This dissertation presents a methodology that combines prediction and grouping techniques called the Clusterwise Segmentation Model with Hybrid Prototypes (CSMoH), which aims to segment the data in clusters so that each cluster is represented by a predictive model, such as a regression model or machine learning algorithm (prototype), among a list of predefined methods. The choice of the best prototype for each cluster is intended to minimize an objective function. In addition to the implementation of the CSMoH method estimation algorithm, we consider different allocation techniques for new observations in order to assess the predictive performance of the algorithm. A proof of convergence is presented, as well as the application of the proposed method in synthetic data and in real databases. A new allocation method based on KNN, called KNN-combining clusters, is proposed, presenting interesting results. In the experiment with synthetic data, the CSMoH algorithm is compared with another algorithm in 6 different scenarios, with an satisfactory performance. In the validation of the CSMoH algorithm with real data, the proposed method presents a relevant performance when compared to 3 other algorithms (Linear K-means, Hybrid K-means and Clusterwise Linear Regression), as well as the evaluation of 5 different allocation methods.