A semi-supervised dynamic ensemble algorithm for IoT anomaly detection
Shudong Liu, Xiping Hao, Xu Chen · 2020
In general, anomaly detection is to identify the abnormal samples from large-scale regular data, which is an interesting research topic in data mining and machine learning domains. In this paper, we propose an anomaly detection algorithm based on semi-supervised extreme learning machine for large-scale IoT data. Unlike supervised learning algorithm, semi-supervised learning algorithm needs to deal with many unlabeled and labeled data, and its performance is better than the former. Because unlabeled data is much convenient to access than labeled data, semi-supervised learning algorithm is widely used in many realworld scenarios. In this paper, for anomaly detection task of large-scale IoT data, we propose a dynamic ensemble algorithm via a combination of semi-supervised extreme learning machine (SSELM) and mutual information criteria. As a base classifier, SSELM has the characteristics of simple structure, convenient training and excellent generalization, we select the base classifiers for the target instance based on maximum relevance and minimum redundancy. Experimental results on four real-life datasets from UCI show that our proposed algorithm for IoT anomaly detection outperforms three state-of-the-art methods in terms of average classification accuracy.