A novel learning method for feature evolvable streams

Recently, many researchers have focused on a novel type of data stream known as a feature evolvable stream, wherein the existing features may become obsolete while new features emerge simultaneously. The occurrence of this phenomenon leads to a degradation in the classification ability of the constructed model. To address this problem, we propose a novel method FENSL for classifying feature evolvable streams. First, we employ fuzzy membership values to enhance the accuracy and reliability of model. Then, we utilize a twin support vector machine to train a classification model on the instances with their existing features. Moreover, as new features emerge, we develop a mapping matrix between two heterogeneous feature spaces through the locally weighted linear regression algorithm. This enables our previously well-trained model to effectively adapt to the new feature space. Finally, by reusing the old model, we construct a stable classification model that is capable of handling data with new features. Experimental results obtained on synthetic datasets show that our proposed method exhibits adaptability in terms of learning from feature evolvable streams and demonstrates great antinoise performance.