[Paper Review] Distributed and parallel time series feature extraction for industrial big data applications

Maximilian Christ, Andreas W. Kempa-Liehr+1 more|arXiv (Cornell University)|Oct 25, 2016

Fault Detection and Control Systems14 references233 citations

TL;DR

The paper introduces FRESH, a scalable, feature-based time series extraction and filtering method that uses non-parametric hypothesis tests and Benjamini-Yekutieli FDR control to select relevant features for classification or regression in industrial big data.

ABSTRACT

The all-relevant problem of feature selection is the identification of all strongly and weakly relevant attributes. This problem is especially hard to solve for time series classification and regression in industrial applications such as predictive maintenance or production line optimization, for which each label or regression target is associated with several time series and meta-information simultaneously. Here, we are proposing an efficient, scalable feature extraction algorithm for time series, which filters the available features in an early stage of the machine learning pipeline with respect to their significance for the classification or regression task, while controlling the expected percentage of selected but irrelevant features. The proposed algorithm combines established feature extraction methods with a feature importance filter. It has a low computational complexity, allows to start on a problem with only limited domain knowledge available, can be trivially parallelized, is highly scalable and based on well studied non-parametric hypothesis tests. We benchmark our proposed algorithm on all binary classification problems of the UCR time series classification archive as well as time series from a production line optimization project and simulated stochastic processes with underlying qualitative change of dynamics.

Motivation & Objective

• Motivate efficient, scalable feature extraction and selection for multivariate time series in industrial settings.
• Propose a feature extraction framework that combines established feature mappings with a significance-filtering step.
• Control false extraction rate when selecting features across large feature sets and datasets.
• Enable parallelizable implementation suitable for distributed data environments.

Proposed method

• Apply a comprehensive set of 111 time series feature mappings to multivariate time series data.
• Perform a dedicated hypothesis test for each feature to assess its relevance to the target (binary or continuous).
• Use non-parametric tests (Fisher, KS, Kendall) tailored to the data codomain to compute p-values.
• Aggregate feature p-values with the Benjamini-Yekutieli procedure to control the false extraction rate (FER).
• Provide variants with PCA-based dimensionality reduction to reduce feature redundancy (FRESH_PCAb, FRESH_PCAa).
• Discuss parallelization and distributed computation to ensure scalability in big data contexts.

Experimental results

Research questions

• RQ1 Can a universal, scalable feature extraction pipeline compete with state-of-the-art shape-based methods on large time series feature sets?
• RQ2 Does a hypothesis-test-based feature selection approach (FRESH) improve predictive performance while controlling false discoveries in industrial big data?
• RQ3 What is the impact of PCA-based redundancy reduction on feature selection and classification accuracy?
• RQ4 How does FRESH perform across diverse datasets, including UCR time series archive and industrial production data?
• RQ5 Is the method amenable to distributed/parallel implementation for large-scale time series analysis?

Key findings

• FRESH and its PCA variants achieve competitive accuracy with AdaBoost and DTW_NN on many datasets, including outperforming some baselines on several UCR and industrial data tasks.
• Best-performing classifiers were AdaBoost and DTW_NN, achieving the highest mean accuracy on 21 of 32 datasets (all from UCR) in the evaluation.
• FRESH-based pipelines often improve or maintain accuracy for Random Forest classifiers across datasets, while reducing the feature set size.
• All feature-based methods scale linearly with time series length; FRESH and FRESH_PCAa scale linearly with the number of features, supporting big data scalability.
• FRESH is designed to be distributed and parallelizable, with the FER controlled by the Benjamini-Yekutieli procedure for dependent hypotheses.
• In iPRODICT industrial data, Boruta-based approaches achieved the highest mean accuracy among the feature-filtered methods (around 72%), while AdaBoost without filtering remained competitive.