Fully Convlutional Neural Networks for state-of-the-art time series classification
Switch branches/tags
Nothing to show
Clone or download
Latest commit 196a996 Mar 23, 2017

README.md

Deep Learning for Time Series Classification

As the simplest type of time series data, univariate time series provides a reasonably good starting point to study the temporal signals. The representation learning and classification research has found many potential application in the fields like finance, industry, and health care. Common similarity measures like Dynamic Time Warping (DTW) or the Euclidean Distance (ED) are decades old. Recent efforts on different feature engineering and distance measures designing give much higher accuracy on the UCR time series classification benchmarks (like BOSS [1],[2], PROP [3] and COTE [4]) but also let to the pitfalls of higher complexity and interpretability.

The exploition on the deep neural networks, especially convolutional neural networks (CNN) for end-to-end time series classification are also under active exploration like multi-channel CNN (MC-CNN) [5] and multi-scale CNN (MCNN) [6]. However, they still need heavy preprocessing and a large set of hyperparameters which would make the model complicated to deploy.

This repository contains three deep neural networks models (MLP, FCN and ResNet) for the pure end-to-end and interpretable time series analytics. These models provide a good baseline for both application for real-world data and future research in deep learning on time series.

Before Start

What is the best approach to classfiy time series? Very hard to say. From the experiments we did, COTE, BOSS are among the best and DL-based appraoch (FCN, ResNet or MCNN) show no significant difference with them. If you prefer white box model, try BOSS first. If you like end-to-end solution, use FCN or even MLP with dropout as your fisrt baseline (FCN also support a certain level of model interpretability as from CAM or grad-CAM).

However, the UCR time series is kind of the 'extremely ideal data'. In a more applicable scenario, highly skewed labels with very non-stationary dynamics and frequent distribution/concept drift occur everywhere. Hopefully we can address these more complex issue with a very neat and effective DL based framework to enable end-2-end solution with good model interpretability , and yeah, we are exactly working on it.

Network Structure

Network Structure Three deep neural network architectures are exploited to provide a fully comprehensive baseline.

Localize the Contributing Region with Class Activation Map

Another benefit of FCN and ResNet with the global average pooling layer is its natural extension, the class activation map (CAM) to interpret the class-specific region in the data [7]. CAM

We can see that the discriminative regions of the time series for the right classes are highlighted. We also highlight the differences in the CAMs for the different labels. The contributing regions for different categories are different. The CAM provides a natural way to find out the contributing region in the raw data for the specific labels. This enables classification-trained convolutional networks to learn to localize without any extra effort. Class activation maps also allow us to visualize the predicted class scores on any given time series, highlighting the discriminative subsequences detected by the convolutional networks. CAM also provide a way to find a possible explanation on how the convolutional networks work for the setting of classification.

Visualize the Filter/Weights

We adopt the Gramian Angular Summation Field (GASF) [8] to visualize the filters/weights in the neural networks. The weights from the second and the last layer in MLP are very similar with clear structures and very little degradation occurring. The weights in the first layer, generally, have the higher values than the following layers. Feature

Classification Results

This table provides the testing (not training) classification error rate on 85 UCR time series data sets. For more experimental settings please refer to our paper.

Please note that the 'best' row is not a reasonable peformance measure. The MPCE score is TODO.

MLP FCN ResNet PROP COTE 1NN-DTW 1NN-BOSS BOSS-VS
50words 0.288 0.321 0.273 0.180 0.191 0.310 0.301 0.367
Adiac 0.248 0.143 0.174 0.353 0.233 0.396 0.220 0.302
ArrowHead 0.177 0.120 0.183 0.103 / 0.337 0.143 0.171
Beef 0.167 0.25 0.233 0.367 0.133 0.367 0.200 0.267
BeetleFly 0.150 0.050 0.200 0.400 / 0.300 0.100 0.000
BirdChicken 0.200 0.050 0.100 0.350 / 0.250 0.000 0.100
Car 0.167 0.083 0.067 / / / / /
CBF 0.14 0 0.006 0.002 0.001 0.003 0 0.001
ChlorineCon 0.128 0.157 0.172 0.360 0.314 0.352 0.340 0.345
CinCECGTorso 0.158 0.187 0.229 0.062 0.064 0.349 0.125 0.130
Coffee 0 0 0 0 0 0 0 0.036
Computers 0.460 0.152 0.176 0.116 0.300 0.296 0.324
CricketX 0.431 0.185 0.179 0.203 0.154 0.246 0.259 0.346
CricketY 0.405 0.208 0.195 0.156 0.167 0.256 0.208 0.328
CricketZ 0.408 0.187 0.187 0.156 0.128 0.246 0.246 0.313
DiatomSizeR 0.036 0.07 0.069 0.059 0.082 0.033 0.046 0.036
DistalPhalanxOutlineAgeGroup 0.173 0.165 0.202 0.223 / 0.208 0.180 0.155
DistalPhalanxOutlineCorrect 0.190 0.188 0.180 0.232 / 0.232 0.208 0.282
DistalPhalanxTW 0.253 0.210 0.260 0.317 / 0.290 0.223 0.253
Earthquakes 0.208 0.199 0.214 0.281 / 0.258 0.186 0.193
ECG200 0.080 0.100 0.130 / / 0.230 0.130 0.180
ECG5000 0.065 0.059 0.069 0.350 / 0.250 0.056 0.110
ECGFiveDays 0.03 0.015 0.045 0.178 0 0.232 0.000 0.000
ElectricDevices 0.420 0.277 0.272 0.277 / 0.399 0.282 0.351
FaceAll 0.115 0.071 0.166 0.152 0.105 0.192 0.210 0.241
FaceFour 0.17 0.068 0.068 0.091 0.091 0.170 0 0.034
FacesUCR 0.185 0.052 0.042 0.063 0.057 0.095 0.042 0.103
fish 0.126 0.029 0.011 0.034 0.029 0.177 0.011 0.017
FordA 0.231 0.094 0.072 0.182 / 0.438 0.083 0.096
FordB 0.371 0.117 0.100 0.265 / 0.406 0.109 0.111
GunPoint 0.067 0 0.007 0.007 0.007 0.093 0 0
Ham 0.286 0.238 0.219 / / 0.533 0.343 0.286
HandOutlines 0.193 0.224 0.139 / / 0.202 0.130 0.152
Haptics 0.539 0.449 0.494 0.584 0.481 0.623 0.536 0.584
Herring 0.313 0.297 0.406 0.079 / 0.469 0.375 0.406
InlineSkate 0.649 0.589 0.635 0.567 0.551 0.616 0.511 0.573
InsectWingbeatSound 0.369 0.598 0.469 / / 0.645 0.479 0.430
ItalyPower 0.034 0.03 0.040 0.039 0.036 0.050 0.053 0.086
LargeKitchenAppliances 0.520 0.104 0.107 0.232 / 0.205 0.280 0.304
Lightning2 0.279 0.197 0.246 0.115 0.164 0.131 0.148 0.262
Lightning7 0.356 0.137 0.164 0.233 0.247 0.274 0.342 0.288
MALLAT 0.064 0.02 0.021 0.050 0.036 0.066 0.058 0.064
Meat 0.067 0.033 0.000 / / 0.067 0.100 0.167
MedicalImages 0.271 0.208 0.228 0.245 0.258 0.263 0.288 0.474
MiddlePhalanxOutlineAgeGroup 0.265 0.232 0.240 0.474 / 0.250 0.218 0.253
MiddlePhalanxOutlineCorrect 0.240 0.205 0.207 0.210 / 0.352 0.255 0.350
MiddlePhalanxTW 0.391 0.388 0.393 0.630 / 0.416 0.373 0.414
MoteStrain 0.131 0.05 0.105 0.114 0.085 0.165 0.073 0.115
NonInvThorax1 0.058 0.039 0.052 0.178 0.093 0.210 0.161 0.169
NonInvThorax2 0.057 0.045 0.049 0.112 0.073 0.135 0.101 0.118
OliveOil 0.60 0.167 0.133 0.133 0.100 0.167 0.100 0.133
OSULeaf 0.43 0.012 0.021 0.194 0.145 0.409 0.012 0.074
PhalangesOutlinesCorrect 0.170 0.174 0.175 / / 0.272 0.217 0.317
Phoneme 0.902 0.655 0.676 / / 0.772 0.733 0.825
Plane 0.019 0 0 / / / /
ProximalPhalanxOutlineAgeGroup 0.176 0.151 0.151 0.117 / 0.195 0.137 0.244
ProximalPhalanxOutlineCorrect 0.113 0.100 0.082 0.172 / 0.216 0.131 0.134
ProximalPhalanxTW 0.203 0.190 0.193 0.244 / 0.263 0.203 0.248
RefrigerationDevices 0.629 0.467 0.472 0.424 / 0.536 0.512 0.488
ScreenType 0.592 0.333 0.293 0.440 / 0.603 0.544 0.464
ShapeletSim 0.517 0.133 0.000 / / 0.350 0.044 0.022
ShapesAll 0.225 0.102 0.088 0.187 / 0.232 0.082 0.188
SmallKitchenAppliances 0.611 0.197 0.203 0.187 / 0.357 0.200 0.221
SonyAIBORobot 0.273 0.032 0.015 0.293 0.146 0.275 0.321 0.265
SonyAIBORobotII 0.161 0.038 0.038 0.124 0.076 0.169 0.098 0.188
StarLightCurves 0.043 0.033 0.025 0.079 0.031 0.093 0.021 0.096
Strawberry 0.033 0.031 0.042 / / 0.060 0.042 0.024
SwedishLeaf 0.107 0.034 0.042 0.085 0.046 0.208 0.072 0.141
Symbols 0.147 0.038 0.128 0.049 0.046 0.050 0.032 0.029
SyntheticControl 0.05 0.01 0.000 0.010 0.000 0.007 0.030 0.040
ToeSegmentation1 0.399 0.031 0.035 0.079 / 0.228 0.048 0.031
ToeSegmentation2 0.254 0.085 0.138 0.085 / 0.162 0.038 0.069
Trace 0.18 0 0 0.010 0.010 0 0 0
TwoLeadECG 0.147 0 0 0.067 0.015 0.096 0.016 0.001
TwoPatterns 0.114 0.103 0 0 0 0 0.004 0.015
UWaveGestureLibraryAll 0.046 0.174 0.132 0.199 0.196 0.272 0.241 0.270
UWaveX 0.232 0.246 0.213 0.283 0.267 0.366 0.313 0.364
UWaveY 0.297 0.275 0.332 0.290 0.265 0.342 0.312 0.336
UWaveZ 0.295 0.271 0.245 0.029 / 0.108 0.059 0.098
wafer 0.004 0.003 0.003 0.003 0.001 0.020 0.001 0.001
Wine 0.204 0.111 0.204 / / 0.426 0.167 0.296
WordSynonyms 0.406 0.42 0.368 0.226 / 0.252 0.345 0.491
Worms 0.657 0.331 0.381 / / 0.536 0.392 0.398
WormsTwoClass 0.403 0.271 0.265 / / 0.337 0.243 0.315
yoga 0.145 0.155 0.142 0.121 0.113 0.164 0.081 0.169
Best 6 27 21 14 10 4 21 9

Dependencies

Keras (Tensorflow backend), Numpy.

Cite

If you find either the codes or the results are helpful to your work, please kindly cite our paper

[Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline] (https://arxiv.org/abs/1611.06455)

[Imaging Time-Series to Improve Classification and Imputation] (https://arxiv.org/abs/1506.00327)

License

This project is licensed under the MIT License.