Data Stream Learning Evaluation: Experimenting with Prequential Approach Over Real Data Streams

Abstract

Current IoT and sensor device-led computing systems are giving rise to enormous streaming data . This ever-growing streaming data necessitates data analytics to do ; on-the-fly decision making thus giving rise to stream learning enabled decisions. These new demands of data stream learning algorithms require evolving algorithms , evaluation methods and dynamic learning capabilities from the machine learning systems. Thus, changing data distribution over stream learning environments is  very less suitable to do with the existent and static environments of data analytics. This article addresses the prequential learning evaluation in data stream learning environment over real data streams like covertypeand airline with scikit multiflow stream learning APIs . The major objective of the paper is to analyse the better alternative to traditional hold-out evaluation . The classifier machine learning algorithms like  tree based adaptive classifiers and some ensemble-based alternatives are empirically evaluated over both hold-out and prequential methods. The results of the experimental evaluation suggest that prequential evaluation is able to achieve improvements  in accuracy. Further this article has evaluated time complexity which again suggests that GPU and similar computing environments if utilized , prequential evaluation may be the optimal solution for dynamic stream learning environments.