Combining Active Learning with Self-train algorithm for classification of multimodal problems

Abstract

In real-world cases, handling of both labeled and unlabeled data has raised the interest of several data scientists and Machine Learning engineers, leading to several demonstrations that apply data augmenting approaches to achieve an effective learning behavior. Although the majority of them propose either the exploitation of Semi-supervised or Active Learning approaches, individually, their combination has not been widely used. The ambition of this strategy is the efficient utilization of the available human knowledge relying along with the decisions driven by automated methods under a common framework. Thus, we conduct an empirical evaluation of such a combinatory approach over three problems, related to multimodal data operating under the pool-based scenario: Gender Identification, Recognition of Offensive Language and Emotion Detection. Into the proposed learning framework, which exploits initially labeled instances with small cardinality, our results prove the benefits of adopting such kind of semi-automated approaches regarding both the achieved predictive correctness and the reduced consumption of time and cost resources, as well as the smoothness of the learning convergence, mainly using ensemble classifiers.