Adaptive Framework for Multivariate Stream Data Processing in Data-Centric Sensor Applications

We introduce an adaptive framework for multivariate sensor stream data reduction. The proposed method takes as input a sliding window of multivariate stream data, classifies the data in each window, and chooses reduction strategies that are most appropriate for the window. In the classification step, it discretizes the stream data into a string of symbols that characterize the signal changes and then applies classification algorithms to classify the transformed sensor stream data. In the second step, depending on the classification labels assigned to each window, it applies most appropriate data reduction techniques and reduction ratios to the window. For classification, we considered supervised methods including Naïve Bayes Model and SVM, and unsupervised methods including Jaccard, TFIDF, Jaro and JaroWinkler. For data reduction, we compared Wavelet, Sampling, SVD and Hierarchical clustering. In our experiments, SVM and TFIDF outperformed the other classification methods and SVD and Sampling showed the best result in data reduction.