Classificação semissupervisionada baseada em densidade com reconhecimento de anomalias

Abstract

In the context of data mining, the task of anomaly detection is important because observations that deviate from the majority can negatively affect machine learning models or represent the main object of interest in various real-world scenarios. At the same time, semi-supervised classification tasks are essential in situations where labeled data are scarce. In this work, we suggest unifying these two tasks into a single integrated process: we propose combining a state-of-the-art density-based clustering algorithm capable of detecting outliers with two well-known density-based semi-supervised classifiers, with the goal of producing hybrid methods capable of performing both tasks. Experiments conducted on 42 semi-synthetic datasets with different proportions of labeled objects and two distinct types of anomalies showed that the investigated anomaly detection method outperforms similar approaches, especially on datasets containing global anomalies. The results also demonstrate that when the outlier detection method is combined with the semi-supervised classifiers, there is only a minor impact on classification quality. Thus, we show that the proposed hybrid approaches constitute viable alternatives to their respective original methods, enabling explicit identification of anomalies without significantly compromising classification performance.