Abstract—Outlier detection is massively consist of large number of application domain. Previously the outlier is known as noisy data, but in recent days it became more important in various field because of its usage to detect the unpredicted and unidentified. The outlier detection are used in certain areas like fraud detection of credit cards, calling cards, discovering computer intrusion and criminal behaviors etc. Aim: the main motive of this research is to propose feature selection and subset based outlier detection. Methods: This paper proposes Optimal Feature selection based Neural Network (OFS-NN) an effectual outlier detection approach accompanied with prior feature optimization strategy. Initially, preprocessing stage formats all data instances available in the dataset utilized and deployed in a SPARK architecture. The datasets are preprocessed and divided into subsets. Furthermore, an Artificial Bee Colony Optimization gets employed in determining for an optimal set of features among the wholesome feature set available. Artificial Bee Colony that exclude outliers on the basis of a feature set. Moreover, an Expectation maximization clustering approach involves in clustering maximum similar data. As a final point, the neural network classification is used for outlier detection. Results: The efficacy of OFS-NN regarding outlier detection gets exemplified by evaluating Area Under Curve (AUC), CPU utilization time, execution time, detection accuracy and memory consumption against existing outlier detection methodologies. OFS-NN evidently proves to be efficacious than other approaches in terms of mitigated execution time under minimum and maximum dataset size. 


Index Terms—ABC algorithm, expectation maximization clustering, neural network classification

Cite: Karthikeyan G and Balasubramanie P " OFS-NN: Optimal Features-Neural Network Based Outlier Detection for Big Data Analysis," Journal of Communications, vol. 13, no. 7, pp. 396-405, 2018. Doi: 10.12720/jcm.13.7.396-405.