Novel Class Detection in Non-stationary Streaming Environment with a Discriminative Classifier
Robert Šajina, Nikola Tanković, Darko Etinger · 2020
In a data streaming environment, one of the biggest challenges for a machine learning classifier is to detect the changes in the concepts that the data corresponds to - a phenomenon called Concept Drift. It can manifest in different modes: existing classes can continually evolve, experience a sudden shift, or a novel class can emerge. Algorithms for novel class detection using cluster-based techniques and statistical approaches applied to the model outputs rely on the assumption that the feature space of the data posses some distance metric governing the class affiliation. Therefore, the novel class will correspond to a significant distance from the known clusters. Most of the time, these assumptions are correct, but the resulting algorithms are challenging to apply on higher-dimensional data such as images. In this paper, we present a novel approach called Discriminative Classifier Detector (DCD) for detecting concept evolution. DCD trains alongside the classification model. The primary model used to evaluate our approach is a Convolutional Neural Network (CNN) classifier for which the proposed DCD is a densely layered neural network. DCD is applied to the outputs of CNN's penultimate layer to achieve its independence against the number of output classes, which enables the underlying CNN model to evolve independently. DCD requires no structural change when the detected novel class is added to the CNN classifier. We demonstrate the effectiveness of our approach on several well-known datasets.