Semi-supervised Learning Using Incremental Polynomial Classifier and Extreme Value Theory

The data in many real-world applications are streamed continuously which causes a variety of problems, e.g. Infinitely long data streams, concept drift, on-line or real-time classification and noise or outlier samples. To overcome these problems, the classifier should be updated continuously and it should have the ability to detect outliers. Since the size of the data set is growing with the duration of the data stream, the classifier should be updated incrementally without storing the whole training set. We present a polynomial classifier that efficiently detects the outliers using the extreme value theory in combination with confidence band intervals derived from regression techniques. All parameters are updated incrementally without requiring the old data. This approach makes the classifier suitable for on-line classification, since the processing time of the update is negligible with respect to the time required for processing the full training data set. In contrast to other novelty detection algorithms which work only with one-class systems, the proposed method can be applied in multi-class systems. The proposed algorithm is evaluated on an unbalanced multi-class gesture database. A comparison of the proposed method with the support vector data description classifier shows that it has superior properties.