OPOSSAM: Online Prediction of Stream Data with Self-Adaptive Memory

Main Article Content

Akihiro Yamaguchi
Shigeru Maya
Tatsuya Inagi
Ken Ueno

Abstract

Data streams such as traffic flows, stock prices, and electricity consumption are endless time-series data from time-varying environments, and concept drift in non-stationary data streams is an important problem. To forecast the shortrange future values of such data streams accurately in real time, we propose an online prediction method called Online Prediction Of Stream data with Self-Adaptive Memory (OPOSSAM). OPOSSAM introduces adaptive memory management consisting of short term memory and long-term memory to manage time-series segments, and forecasts future values by local regression based on similar time-series segments. In order to deal with concept drift, OPOSSAM automatically adjusts the prediction model learned from short-term memory by considering the prediction model learned from the entire memory as the prior model. In addition, OPOSSAM keeps long-term memory consistent by reducing redundant samples with large prediction errors. Experimental results showed a reduction in prediction errors compared with baseline methods on real-world datasets in the different domains of traffic flow, stock prices, and electricity consumption.

Article Details

Section
Articles