Um algoritmo de nuvem de partículas para combinação de classificadores em aprendizado multi-visão

Abstract

The multi-view or multi-modality learning approach is becoming popular for providing different representations of a problem from which classifiers can learn from. Given the task of video classification, for example, the sound, the image and the subtitles may be considered views.The main idea behind multi-view learning is that learning from these representations separately can lead to better gains than merging them into a single dataset. Hence, a classification model is created for each view, and outputs provided by each of them must be combined to provide a final class for each example.This dissertation proposes a Particle Swarm Optimization (PSO) algorithm to combine the outputs coming from different views. The PSO works in two contexts: the first takes into account only the class/confidence assigned by a classifier in the categorization of an instance, applying weights to each view. The second, besides assigning weights to views, also assigns weights to each class.Experiments were performed in two datasets, each one with three views, and compared with three different methods from the literature: the majority vote, the Borda Count algorithm and the Dempster-Shafer theory. Then, a comparison was made with the approach that uses all views together into a single dataset. The PSO obtained statistically better results than the other approaches evaluated in the majority of the experiments.