Data stream classification in dynamic feature space using feature mapping

Stream learning in dynamic feature space has evolved into an immensely popular field. This problem assumes that each instance of the data stream may have different features, and the feature spaces of the classifier and the instances may also differ. Such assumptions are more relevant to real-world applications dealing with data streams, where dimensions are not predetermined and fixed. This study introduces a general algorithm for data stream classification in dynamic feature space using feature mapping. In contrast with the other studies, the proposed algorithm is not based on a specific classifier and can cooperate with any classifier best suited for an intended application. It discovers the relationship between the features and estimates the unavailable features previously observed by the classifier. This technique helps to exploit the full potential of the classifier. Furthermore, empirical experiments and comparisons with modern methods demonstrate that the proposed algorithm achieves higher accuracy.