Classification of acoustical signals by combining active learning strategies with semi-supervised learning schemes

In real-world cases, handling 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 in order to obtain a robust and, at the same time, accurate enough learning behavior. The main reason is the existence of much unlabeled data that are ignored by conventional supervised approaches, reducing the chance of enriching the final formatted hypothesis. However, the majority of the proposed methods that operate using both kinds of these data are oriented toward exploiting only one category of these algorithms, without combining their strategies. Since the most popular of them regarding the classification task are Active and Semi-supervised Learning approaches, we aim to design a framework that combines both of them trying to fuse their advantages during the main core of the learning process. Thus, we conduct an empirical evaluation of such a combinatory approach over three problems, which stem from various fields but are all tackled through the use of acoustical signals, operating under the pool-based scenario: gender identification, emotion detection and automatic speaker recognition. Into the proposed combinatory framework, which operates under training sets with small cardinality, our results prove the benefits of adopting such kind of semi-automated approaches regarding both the achieved predictive correctness when reduced consumption of resources takes place, as well as the smoothness of the learning convergence. Several learners have been examined for reaching to more general conclusions, and a variant of self-training scheme has been also examined.