Meta-Learning for Time Series-based Tasks in Software Engineering

Time series analysis is widely employed in software engineering (SE) for various critical activities, including capacity planning, anomaly detection, and performance testing. However, effectively incorporating time series analysis into SE processes can require extensive experimentation or specialized expertise, which may not often be readily accessible to software engineers. Meta-learning offers a promising avenue to reduce human intervention by automating several aspects of the time series analysis process, such as selecting the most appropriate algorithm. Although this approach has shown promising results across various time series domains, its applicability and potential impact within SE contexts still remain largely unexplored. In this paper, we report the first empirical evidence on the effectiveness of meta-learning for time series–based SE tasks. First, we introduce MALIAS, a framework inspired by the meta-learning principles, which leverages over 700 time series features, a random forest model, and best practices such as dimensionality reduction and hyper-parameter tuning, to automatically select and run the best-performing algorithm for a given software time series. Then, we instantiate this framework for three distinct time series-based SE tasks–forecasting of software telemetry data, anomaly detection in online software systems, and steady-state detection in software performance testing–and we conduct an extensive evaluation across six different real-world datasets. Results show that, when compared to the best-performing algorithm, MALIAS yields net improvements up to +17.7 percentage points (pp) for forecasting, up to +20pp for anomaly detection, and up to +19.8pp for steady-state detection. Furthermore, when compared to other automated approaches, such as established AutoML frameworks, MALIAS consistently outperforms them across all the considered tasks with very high statistical significance (p<0.0001), while showing competitive training and testing times, despite heterogeneous results. These findings highlight the potential of meta-learning for time series analysis in SE, paving the way for its broader adoption in both research and practice.