Optimized Online Feature Selection Algorithm for Improving Performance of the Classifier

Social media platforms generate a continuous stream of data with diverse structures and high dimensionality. The conventional feature selection methods face challenges in handling the inherent complexity of datasets with such elevated dimensions. To tackle this, the proposed approach incorporates The Optimal Online Feature Selection technique, aiming to effectively address the complexities presented by the continuous data flow. This technique incorporates sparsity regularization and truncation methods in combination with a gradient descent rule. This strategic combination efficiently updates feature weights, enabling the selection of a subset from a vast number of features in extensive data streams. The experiment is conducted in 10 different datasets which were gathered from diverse repositories including UCI, Kaggle, LIBSVM, and KDD, Before analysis, preprocessing steps encompass the datasets which include data cleansing, attribute conversion, and min-max scaling. Then online feature selection method is applied and its outcome is tested with three different classifiers such as Random Forest, Hoeffding, and MLP, and their results are compared using performance metrics, revealing distinctions among them. Results indicate that the Random Forest Classifier, coupled with the Online Feature Selection algorithm, demonstrates minimal classification errors, outperforming than existing classification methods.