A Comparative Study for Outlier Detection Strategies Based On Traditional Machine Learning For IoT Data Analysis.

Elmahalawy, Amina Mohamed; Mousa, Hayam; Amin, Khalid

doi:10.21608/ijci.2021.91957.1059

A Comparative Study for Outlier Detection Strategies Based On Traditional Machine Learning For IoT Data Analysis.

Document Type : Original Article

Authors

¹ Information Technology Department Faculty of Computers and Information Menoufia University, Egypt

² Information Technology dept., Faculty of Computers and Information, Menoufia University, Egypt

10.21608/ijci.2021.91957.1059

Abstract

Internets of Things (IoT) systems are increasing very fast. They have different types of wireless sensor networks (WSN) behind them. These networks have many applications that are a portion of our life such as healthcare, agricultural, mechanical, and military systems applications. Therefore, a massive amount of data was collected. Outlier detection is one of the essential fundamental problems in these applications. It helps to discover erroneous, imperfect, and noisy nodes. There are various techniques used to detect this outlier. Machine learning algorithm-based approaches are exceptionally much valuable and successful among them. This paper is concerned with the study of outlier detection techniques. It categorizes them into different approaches, such as Statistical, Nearest_Neighbor, Clustering, Subspace, Ensemble-based, and other approaches. These approaches are examined in detail. This study is concerned with determining the best outlier detection method that can be used to detect the outlier in the IoT data analysis. In conclusion, the experimental results show that the Isolation Forest, HBOS, and CBLOF approaches give better performance in terms of precision, Area under the curve (AUC), and execution time than other algorithms.

Keywords