Towards understanding the importance of time-series features in automated algorithm performance prediction
Accurate and reliable forecasting is a crucial task in many different domains. The selection of a forecasting algorithm that is suitable for a specific time series can be a challenging task, since the algorithms’ performance depends on the time-series properties, as well as the properties of the for...
Saved in:
| Published in: | Expert systems with applications Vol. 213; p. 119023 |
|---|---|
| Main Authors: | , , |
| Format: | Journal Article |
| Language: | English |
| Published: |
Elsevier Ltd
01.03.2023
|
| Subjects: | |
| ISSN: | 0957-4174, 1873-6793 |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Accurate and reliable forecasting is a crucial task in many different domains. The selection of a forecasting algorithm that is suitable for a specific time series can be a challenging task, since the algorithms’ performance depends on the time-series properties, as well as the properties of the forecasting algorithms. The methodology and analysis presented in this paper are contributing towards understanding the performance of time-series forecasting methods. Instead of using time-series meta-features only to obtain a good meta-model that can predict the performance of a forecasting algorithm, the methodology can link which features are important for which forecasting methods. We used time-series meta-features extracted using the tsfresh and catch22 libraries. We also found that the importance of the meta-features changes depending on the meta-model that is used. There are only a few meta-features that always appear important for a given forecasting method no matter which meta-model will be used for learning, which further provides opportunities to select a model-agnostic feature portfolio. In addition, different feature importance techniques can provide different results that are related to the methodology that is used by the meta-model. By using the feature importance obtained by a meta-model and a specified feature importance technique, we can define a representation of a forecasting method behavior, which can further provide an insight into which forecasting methods have similar behavior.
•Meta-learning performance prediction for time-series forecasting algorithms.•Analysis of features in the catch22 and tsfresh libraries.•Time-series feature importance analysis for algorithm performance prediction. |
|---|---|
| ISSN: | 0957-4174 1873-6793 |
| DOI: | 10.1016/j.eswa.2022.119023 |