Anomaly Detection on Shuttle data using Unsupervised Learning Techniques

Many modern day applications require the ability to identify those observations or data that deviate from the ones that are considered to be normal by domain expert. Anomaly detection helps to identify these anomalies and once identified, then the system can take the necessary changes. In data mining, this problem is tackled using supervised and unsupervised machine learning techniques. Since in many practical applications, data used will have no labels, unsupervised learning techniques are well suited. This work was aimed at comparing various unsupervised anomaly detection techniques using performance metrics like precision, recall, F-score and area under the curve. The unsupervised learning techniques used in this work are One Class Support Vector Machine(OneClassSVM), Local Outlier Factor(LOF), Isolation Forest(IF) and Elliptic Envelope(EE). Shuttle and satellite datasets were used for experimentation. Performance of these unsupervised learning techniques were compared with supervised learning techniques such as SVM and k-NN. Results show that unsupervised learning techniques are on par or better for anomaly detection compared to supervised learning techniques for the shuttle and satellite datasets.