[Paper Review] Multivariate Time Series Classification using Dilated Convolutional Neural Network

Omolbanin Yazdanbakhsh, Scott Dick|arXiv (Cornell University)|May 5, 2019

Time Series Analysis and Forecasting21 references34 citations

TL;DR

The paper proposes a dilated CNN-based method that converts multivariate time series segments into one-channel images and classifies them, achieving competitive results on WISDM datasets without hand-crafted features.

ABSTRACT

Multivariate time series classification is a high value and well-known problem in machine learning community. Feature extraction is a main step in classification tasks. Traditional approaches employ hand-crafted features for classification while convolutional neural networks (CNN) are able to extract features automatically. In this paper, we use dilated convolutional neural network for multivariate time series classification. To deploy dilated CNN, a multivariate time series is transformed into an image-like style and stacks of dilated and strided convolutions are applied to extract in and between features of variates in time series simultaneously. We evaluate our model on two human activity recognition time series, finding that the automatic features extracted for the time series can be as effective as hand-crafted features.

Motivation & Objective

• Motivate multivariate time series classification and reduce reliance on hand-crafted features.
• Propose an image-based representation of multivariate segments processed by dilated CNNs.
• Demonstrate that learned features can match hand-crafted spectrogram and statistical features on benchmark data.

Proposed method

• Convert multivariate time series segments into one-channel images with M rows and K columns.
• Apply stacks of dilated and strided convolutions to learn intra- and inter-variate relations.
• Use a fully connected classifier with softmax for final class probabilities.
• Train with cross-entropy loss using Adam optimizer and L2 regularization.

Experimental results

Research questions

• RQ1 Can a dilated CNN operating on image-like representations of multivariate time series achieve competitive classification performance without hand-crafted features?
• RQ2 How well does the model capture inter-variate and temporal dependencies compared to prior approaches using spectrogram or hand-crafted features?
• RQ3 What is the impact of architecture choices (dilation, stride, layer depths) on recognition accuracy across activities?

Key findings

• On WISDM v1-split, our method achieves higher F1-scores than the cited spectrogram-based approach for Walking, Jogging, Sitting and Standing, with Upstairs and Downstairs being the weaker ones.
• On WISDM v1-individual, the method shows competitive recall across activities compared to prior CNN-based feature extraction methods.
• On WISDM v.2, the model matches or slightly trails the spectrogram-based method for Walking and Jogging, but outperforms for Sitting, Standing, and Lying Down.
• Overall, automatically learned features via the dilated CNN can be as effective as hand-crafted features in several activities, with some exceptions in stairs-related classes.