Discovering Highly Informative Feature Sets from Data Streams

How to select interesting feature sets from data streams is a new and important research topic in which there are three major challenges. First of all, instead of discovering features individually and independently, we are interested in comprehensively selecting a subset of features whose joint importance or weight is the highest. Secondly, we are concerned with the problem of selecting feature sets over dynamic, large and online data streams which are only partly available when we are selecting the features. This problem distinguishes itself over the data streams from the ones on the static data which is completely available before the feature selection. Finally, data streams may evolve over time, requiring an online feature selection technique which can capture and adapt to such changes. We introduce the problem of online feature selection over data streams and we provide a heuristic solution. We also demonstrate the effectiveness and efficiency of our method through experiments on real-world mobile web usage data.