A Network-Based Semi-supervised Outlier Detection Technique Using Particle Competition and Cooperation

Outlier detection is a classical topic of data mining acting as an essential task for discovering knowledge. Its aim is to detect patterns that deviate from normal behaviour. Numerous outlier detection techniques have been developed but little work has been done in the context of semi-supervised learning. Semi-supervised outlier detection techniques are relatively new and include some labels of normal instances for improving the accuracy of classifier with respect to outliers. In this paper, we introduced a new outlier score using a stochastic network-based semi-supervised technique. The original algorithm builds a network on the data set and exploits some available normal labels in order to classify the whole network by means of a particle competition and cooperation mechanism. Particles visit vertices in order to conquer as many vertices as possible. Our assumption behind such a score is that an outlier is statistically improbable, and the number of visits received by an outlier vertex is significantly different from the remaining vertices. Experimental results suggest that our technique is capable of detecting many of outliers and outperforms some traditional techniques.