Incremental Time Series Prediction Using Error-Driven Informed Adaptation
| Autor: | Viera Rozinajová, Anna Bou Ezzeddine, Petra Vrablecová |
|---|---|
| Rok vydání: | 2016 |
| Předmět: |
Scheme (programming language)
Sequence Concept drift business.industry Computer science 02 engineering and technology Machine learning computer.software_genre 01 natural sciences Data modeling 010104 statistics & probability Identification (information) 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing Artificial intelligence Adaptive learning Data mining 0101 mathematics Time series business Adaptation (computer science) computer computer.programming_language |
| Zdroj: | ICDM Workshops |
| DOI: | 10.1109/icdmw.2016.0065 |
| Popis: | This paper presents an approach to predictive modeling of sequentially arriving data, also known as a stream. Because of their unique properties, this kind of data requires different mining techniques. The ultimate limitations are the memory and the time. Since the number of records can be infinite, it is not possible to store them all in memory or read them more than once. Hence, the prediction method should work incrementally. Another important aspect of these data is that their characteristics change over time. The identification of the ongoing change in the monitored data sequence, also called "concept drift", can significantly help to improve prediction accuracy by prediction model adaptation to the drifts. The challenge is to perform these model adaptations online. We have proposed an incremental adaptive method for time series prediction. Our approach is based on the adaptive learning scheme "predict – diagnose – update". The main concern was to find out whether our error-driven informed adaptation approach can equal the traditional blind adaptation approaches in accuracy and required resources such as time and memory. The results showed that informed adaptation can achieve comparable accuracy but uses less computational resources. We focused specifically on power demand forecasting but we showed that the approach is applicable also on time series with similar characteristics from other domains. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|