Pattern Matching to improve Accuracy and Efficiency of File Lifecycles Forecasting
Adrian Khelili, Sophie Robert Hayek, Soraya Zertal · 2023
Time series prediction is a crucial task with applications in various domains. In this research, we introduce a transfer learning approach that goes beyond the historical time series of individual instances to enhance prediction performance. Traditional time series prediction methods often focus solely on historical data from a single time series, limiting their ability to capture already known patterns and behaviors. Our proposed method addresses this limitation by constructing a curated database of time series representative of the behavior to predict. We apply our suggested method to the prediction of the I/O (Input/Output) behavior of scientific applications such as NEMO, NAMD, and LQCD. Our non-redundant database contains diverse time series, allowing us to extract valuable insights and patterns that might not be discernible from individual series. By leveraging similarities and patterns across those time series, our approach improves prediction accuracy and robustness. The results we obtained are highly promising, demonstrating that in term of time serie ranking prediction accuracy our proposed inter-timeseries method achieves from 72.9% to 83.8% accuracy in most difficult scenarios to 97% for the others. To compare our solution to existing methods, we show an improvement of 10.1% accuracy compared to ARIMA.