Abstract:
|
Machine learning algorithms can easily be implemented on streaming data. However, because learning algorithms have traditionally been used with static training and/or testing data sets, the temporal information included in streaming data often goes unused. In particular, many algorithms work under the strong assumption that sequential observations are independent. Leveraging the temporal information and structure into learning algorithms may help improve predictive analytic capabilities on streaming data. Applications for these methods are varied, ranging from real-time prediction of extreme hurricane events, to tracking traffic movement using GPS data, to online quality control monitoring of mass spectrometry data. We propose the use of statistical models incorporating temporal dependence to improve upon current state-of-the-art learning algorithms. We compare our results to the current methodologies to assess improvement of real-time forecasts in streaming data applications.
|