One-class classification with node embedding type features

One-class classification (OCC) deals with building predictive models when data from only one of the classes is available. Generating new features in traditional binary classification approaches has shown to produce better classifiers than the original feature space. However, new feature generation techniques have not been adequately addressed for OCC algorithms. Graph-based approaches have been very effective in anomaly detection problems. However, these approaches cannot be applied to OCC problems as they require creation of a graph of all the points, including the testing point, which is computationally expensive. In graph-based approaches, nearest neighbors of a data point are useful to compute the class of a data point. This suggests that nearest neighbors can be used to generate new features. Two different feature generation algorithms that are based on nearest neighbors are proposed for OCC problems. In these approaches, no graph is generated. The motivation is that the nearest neighbors of an outlier data point will be at larger distances as compared to the nearest neighbors of a normal data point. One of these algorithms also uses the nearest neighbors of the nearest neighbors of a data point to generate features. It is expected, that this additional step will add more information to the new features. OCC algorithms can then be applied on these new feature spaces. It is expected that OCC algorithms will perform better on these feature spaces than on original features. Extensive experiments were conducted with one of the most successful OCC algorithms, Isolation Forests(IF). The results suggest that IF with new feature spaces can produce better results than with the original features. Experiments were also performed to understand the effect of various parameters on the performance of IF, such as the number of nearest neighbors, similarity measures, and the combination of new feature spaces with original features. The comparative study of outlier scores generated by different OCC algorithms for normal data points and outlier data points suggests the effectiveness of the proposed approach in generating excellent decision boundaries. The results also suggest that the new features are useful for another very successful OCC algorithm, autoencoders, showing that the proposed techniques can also be effective for other OCC algorithms.