[Paper Review] Multi-Scale Convolutional Neural Networks for Time Series Classification

Zhicheng Cui, Wenlin Chen+1 more|arXiv (Cornell University)|Mar 22, 2016

Time Series Analysis and Forecasting15 references475 citations

TL;DR

MCNN is an end-to-end neural network for time series classification that learns multi-scale and multi-frequency features through a multi-branch architecture, achieving state-of-the-art performance on benchmark datasets.

ABSTRACT

Time series classification (TSC), the problem of predicting class labels of time series, has been around for decades within the community of data mining and machine learning, and found many important applications such as biomedical engineering and clinical prediction. However, it still remains challenging and falls short of classification accuracy and efficiency. Traditional approaches typically involve extracting discriminative features from the original time series using dynamic time warping (DTW) or shapelet transformation, based on which an off-the-shelf classifier can be applied. These methods are ad-hoc and separate the feature extraction part with the classification part, which limits their accuracy performance. Plus, most existing methods fail to take into account the fact that time series often have features at different time scales. To address these problems, we propose a novel end-to-end neural network model, Multi-Scale Convolutional Neural Networks (MCNN), which incorporates feature extraction and classification in a single framework. Leveraging a novel multi-branch layer and learnable convolutional layers, MCNN automatically extracts features at different scales and frequencies, leading to superior feature representation. MCNN is also computationally efficient, as it naturally leverages GPU computing. We conduct comprehensive empirical evaluation with various existing methods on a large number of benchmark datasets, and show that MCNN advances the state-of-the-art by achieving superior accuracy performance than other leading methods.

Motivation & Objective

• Motivate improving TSC accuracy and efficiency beyond traditional feature extraction+classifier pipelines.
• Propose an end-to-end MCNN architecture that jointly learns feature representations and classification.
• Integrate transformations across time and frequency domains to capture multi-scale information.
• Leverage data augmentation to improve generalization on smaller TSC datasets.

Proposed method

• Introduce MCNN with three stages: transformation, local convolution, and full convolution.
• Transformation stage creates multi-scale branches using down-sampling and multi-frequency spectra via moving averages.
• Local convolution stage applies independent 1-D convolutions to each branch with max pooling of multiple sizes.
• Full convolution stage concatenates branch outputs and applies additional convolutions and a softmax classifier.
• Train the model end-to-end with cross-entropy loss and backpropagation.
• Data augmentation via window slicing to increase training size for smaller datasets.

Experimental results

Research questions

• RQ1 Can MCNN automatically learn discriminative features for TSC without handcrafted features?
• RQ2 Do multi-scale and multi-frequency branches improve classification accuracy compared to single-scale CNNs?
• RQ3 How does MCNN compare to state-of-the-art TSC methods on standard benchmarks (UCR datasets)?

Key findings

• MCNN achieves superior accuracy across the majority of datasets, outperforming many leading methods.
• On 41 of 44 UCR datasets, MCNN outperforms a standard CNN with the same parameter count.
• MCNN has a mean rank of 3.95, competitive with the top ensemble method COTE and better than most baselines.
• MCNN significantly outperforms most methods (except BOSS and COTE) according to binomial and Wilcoxon signed-rank tests at the 1% level.
• Data augmentation via window slicing helps MCNN generalize on smaller datasets; MCNN remains strong without augmentation on larger datasets.