Online feature Selection using Pearson Correlation Technique

The large volume of data that is generated nowadays in the era of real-time data processing takes a long time. Streaming feature selection is one way to get out of this real time processing challenge. In stream-wise feature selection, new features are assessed one by one for insertion into a forecasting model. When there is a large pool of acceptable characteristics, stream-wise feature selection performs well over traditional feature selection approaches, in which the features are known in advance. Over fitting can be avoided by adjusting the feature addition threshold dynamically. Unlike classic forward feature selection methods like stepwise regression, which assesses all available features at each step and choose the best one, stream-wise feature selection considers each feature only once when it is formed. In our work, we have used Pearson correlation technique for feature selection of streaming data blocks and compared with complete feature list of streaming data blocks for human activity recognition (HAR) dataset. The proposed algorithm gives promising results in terms of accuracy 94.85(%) over complete feature list.