Visualização de dados em processos de Machine Learning

Abstract

With the recent emergence of Big Data the amount of data being generated on a daily basis has never been greater. In order to extract knowledge from this data Machine Learning plays a key role, due to its ability to learn from historical data. Machine Learning techniques comprise data- driven methods that combine fundamental concepts from computer science with others fields like statistics, probability, optimization and aim to implement algorithms capable of learning, with little or no need for human assistance or intervention. Given the complexity of data sets, visual components are added to the Machine Learning process. The visual components aim to visually represent the information. Helping in the interpretation of data, taking advantage of the visual system of human beings to recognize patterns. Besides offering different ways to show the results in a more friendly way, helping in the decision making process. To demonstrate the relevance of data visualization during the Machine Learning process, this work develops and evaluates a Machine Learning process supported by data visualization tools. The case study method was used, since it is considered that each data set must be treated in a unique way, given its characteristics and distributions. The visualization tools selected were Yellowbrick, Seaborn, Ploty, and MatPlotLib. In this work, two distinct datasets were selected. The first data set, includes public data that contains the number of cases of those contaminated, recovered and killed by the new coronavirus in all continents. Through this set, numerical and categorical data visualization resources were explored, with maps and mixed graphics. The second dataset, widely used in studies in the area, was the Iris dataset. This dataset is commonly used to evaluate classification algorithms because it has a well-defined structure. Moreover, it constitutes an important element for studies in the area because its data present numerical and categorical characteristics (classes). This enables comparisons, analyses, and visualizations that, once understood in an illustrative scenario, can be replicated in other complex contexts. As observed results, we highlight the importance of visualization tools in the Machine Learning process, contributing to the understanding and validation of its steps. In particular, in data sets with Coronavirus, error analysis resources (mean, square, etc.) and statistical measures are the most used. On the other hand, on data sets with Iris, resources for analyzing the accuracy of the classifiers are more relevant. To conclude, it is considered that the results were satisfactory and that visualization tools, besides being necessary for decision making, help the human being in the data exploration process. [resumo fornecido pelo autor]