Title: StreamMARS: A streaming multivariate adaptive regression splines algorithm
Authors: Inci Batmaz - Middle East Technical University (Turkey) [presenting]
Niall Adams - Imperial College London and University of Bristol (United Kingdom)
Abstract: Computers and internet have become inevitable parts of our life in the 1990s, and afterwards, bulk of data are started being recorded in digital platforms automatically. To extract meaningful patterns from such data computational methods are developed in data mining and machine learning domains. Multivariate adaptive regression splines (MARS) is one such method successfully applied to off-line static data for prediction. In about last ten years, we face with the big data problem due to the steady increase in the size of the data. Streaming data is a kind of big data collected from sensor networks, production processes, twitter messages etc. Algorithms processing this type of data should consider both memory and time limitations as well as its changing nature with time. We develop a streaming version of a powerful predictive method MARS for estimating model parameters on-line in a temporarily adaptive manner using forgetting factors. Performance of the algorithm developed is tested on simulated data with different dimensions in static, abrupt and smoothly changing environments; as well as on real-life datasets, and also, compared with those of some benchmarking methods such as sliding windows. Results show that StreamMARS is a promising algorithm for predicting streaming big data.