Aprendizado ativo em modo batch ordenado

Abstract

With the large amount of information generated every day on the internet, it is getting harder, if not impossible, to manually administrate and process such data. In order to overcome this problem, Machine Learning algorithms are becoming widely used in different domains. The Active Learning field arose as a way to optimize the training phase of some machine learning methods. The main idea is that algorithms can achieve better results with smaller training sets if they are allowed to select which instances should be labeled. This assumption is specially interesting in scenarios in which unlabeled data is abundant and there is a cost, not negligible, associated with instance labeling. In its original form, one instance is labeled and incorporated by the Active Learning algorithm at each iteration, thus making it impossible to use multiple oracles in parallel. In order to solve this problem, Batch-Mode Active Learning methods arose by being able to select more than one instance at each iteration. Although this methods allow the use of multiple oracles, it is still necessary to run the algorithm at each annotated batch. With the increasing use of these methods in business environments, it is important to reduce the necessity of constant iterations, and consequently, the analysts idle time when waiting for a new batch to be created. In this dissertation, the Ranked Batch-Mode Active Learning problem is described. It relaxes traditional Batch-Mode Active Learning methods by generating a query that is an ordered list of instances, thus allowing batches to be of arbitrarily large sizes. In this way, the algorithm iterations can be spaced according to the user needs. This characteristics allow that a sufficiently large instance list (for a full work day) be generated outside working hours, then avoiding frequent stops for batch construction. In addition to the formal definition of this problem, one solution is presented that consists of a framework which iteratively builds the instance list by weighting the instance utility for the classifier(uncertainty) and the diversity brought to the model regarding already labeled and selected instances. The experimental evaluation shows that the Ranked Batch allows the reduction of the algorithm executions while maintaining the quality of the selected instances. In some cases, using only unlabeled data, the results obtained are better to the ones of traditional methods. In other words, an ordered list of instances, generated using unlabeled data, was able to guide a labeling process with results statistically better or equal to traditional active learning algorithms without their limitations.