영문내용
(English Abstract) |
Advance in wireless sensor network made it possible to collect massive amounts of sensor-based data with temporal attribute at real-time, Existing prediction methods for time-seriess data have tried to predict the future using the past data or without updating a model. However, the methods are not suitable for stream data, because the data is characterized by high data rate and changes in data distribution over time. In this paper, we propose IMQR(Incremental Multiple Quadratic Regression), a novel prediction method using incremental model updation. To apply stream data to a multiple regression model, we gractionalize stream data in several attributes using fractal. The model is then updated using sliding window to reflect changes in data distrubution incrementally. This method that updates the model through a matrix having minimal information without maintaining the past data guarantees low space complexity. We evaluated error rate in IMQR using RME(Relative Mean Error) and RMSE(Root Mean Square Error). As a result, compared with other methods such as DES(Double Exponential Smoothing), SVR(Support Vector Regression), MLR(Multiple Linear Regression), and MQR(Multiple Quadratic Regression), RME and RMSE in IMQR decreased by 5.6% and 0.067% in average, respectively. |
