Deep Neural Network Ensembles for Time Series Classification

This is the companion repository for our paper also available on ArXiv titled "Deep Neural Network Ensembles for Time Series Classification". This paper has been accepted at the IEEE International Joint Conference on Neural Networks (IJCNN) 2019.

Approach

Data

The data used in this project comes from the UCR/UEA archive, which contains the 85 univariate time series datasets.

Code

The code is divided as follows:

• The main.py python file contains the necessary code to run all experiements.
• The utils folder contains the necessary functions to read the datasets and manipulate the data.
• The classifiers folder contains eight python files one for each deep individual/ensemble classifier presented in our paper.

To run a model on all datasets you should issue the following command:

python3 main.py

To control which datasets and which individual/ensemble classifiers to run see the options in constants.py.

You can control which algorithms to include in the ensemble by changing this line of code.

Prerequisites

All python packages needed are listed in pip-requirements.txt file and can be installed simply using the pip command.

• numpy
• pandas
• sklearn
• scipy
• matplotlib
• tensorflow-gpu
• keras
• h5py
• keras_contrib

Results

The following table shows the results of four ensembles, the raw results can be found here.

	Fine-tuned FCNs	NNE	ALL	ResNets
50words	66.81	80.00	80.00	77.14
Adiac	85.17	85.17	83.38	83.63
ArrowHead	84.00	86.29	86.29	86.86
Beef	76.67	76.67	80.00	76.67
BeetleFly	90.00	85.00	85.00	85.00
BirdChicken	90.00	95.00	85.00	90.00
CBF	99.78	99.44	98.56	99.78
Car	91.67	95.00	86.67	93.33
ChlorineConcentration	82.42	85.05	83.98	85.49
CinC_ECG_torso	85.87	89.71	92.90	83.55
Coffee	100.00	100.00	100.00	100.00
Computers	83.20	83.60	71.60	83.60
Cricket_X	78.97	82.05	77.95	81.54
Cricket_Y	79.23	84.36	78.72	82.05
Cricket_Z	82.05	83.85	79.49	82.05
DiatomSizeReduction	30.07	30.07	88.56	30.07
DistalPhalanxOutlineAgeGroup	71.94	72.66	76.26	73.38
DistalPhalanxOutlineCorrect	77.54	77.90	77.90	78.99
DistalPhalanxTW	71.22	65.47	67.63	66.19
ECG200	89.00	89.00	92.00	88.00
ECG5000	94.16	94.42	94.51	93.67
ECGFiveDays	99.54	99.88	99.65	98.61
Earthquakes	71.94	74.82	74.82	72.66
ElectricDevices	71.74	74.39	73.03	74.22
FISH	96.00	97.71	93.71	98.29
FaceAll	92.84	86.39	83.91	84.02
FaceFour	93.18	95.45	92.05	95.45
FacesUCR	93.95	95.76	95.51	95.90
FordA	90.67	93.70	94.22	92.56
FordB	88.04	92.90	92.33	92.16
Gun_Point	100.00	100.00	99.33	99.33
Ham	74.29	75.24	74.29	78.10
HandOutlines	92.70	95.14	93.78	93.78
Haptics	50.65	52.60	50.97	53.25
Herring	65.62	60.94	62.50	60.94
InlineSkate	40.55	38.36	38.00	38.55
InsectWingbeatSound	39.49	59.75	65.91	52.73
ItalyPowerDemand	96.11	96.50	96.89	96.40
LargeKitchenAppliances	89.60	90.93	83.20	89.60
Lighting2	80.33	80.33	77.05	78.69
Lighting7	89.04	90.41	83.56	83.56
MALLAT	96.93	96.93	95.44	97.40
Meat	91.67	95.00	93.33	96.67
MedicalImages	78.29	79.74	80.13	78.42
MiddlePhalanxOutlineAgeGroup	53.90	59.09	60.39	59.09
MiddlePhalanxOutlineCorrect	81.10	83.51	83.85	83.51
MiddlePhalanxTW	51.95	51.95	55.19	49.35
MoteStrain	93.37	93.93	93.45	93.05
NonInvasiveFatalECG_Thorax1	96.44	96.39	95.88	95.01
NonInvasiveFatalECG_Thorax2	95.73	96.18	96.54	95.01
OSULeaf	97.52	98.76	78.51	98.35
OliveOil	86.67	86.67	86.67	86.67
PhalangesOutlinesCorrect	83.57	84.27	83.57	84.97
Phoneme	32.65	35.13	30.91	34.81
Plane	100.00	100.00	99.05	100.00
ProximalPhalanxOutlineAgeGroup	84.39	84.88	85.85	85.37
ProximalPhalanxOutlineCorrect	92.10	91.75	90.38	92.10
ProximalPhalanxTW	79.51	77.56	80.98	78.54
RefrigerationDevices	50.40	53.07	53.33	52.80
ScreenType	65.07	62.13	52.27	62.13
ShapeletSim	86.11	81.11	70.56	93.89
ShapesAll	90.00	92.83	89.17	92.00
SmallKitchenAppliances	79.47	82.13	77.60	78.93
SonyAIBORobotSurface	95.84	94.68	78.04	96.17
SonyAIBORobotSurfaceII	98.22	97.69	88.88	98.11
StarLightCurves	96.78	97.92	97.79	97.38
Strawberry	97.84	98.11	97.57	98.11
SwedishLeaf	97.28	97.28	96.16	96.48
Symbols	95.68	95.88	91.06	91.56
ToeSegmentation1	96.49	98.25	81.58	96.05
ToeSegmentation2	90.77	92.31	93.08	91.54
Trace	100.00	100.00	98.00	100.00
TwoLeadECG	99.91	100.00	97.72	100.00
Two_Patterns	87.62	100.00	100.00	100.00
UWaveGestureLibraryAll	82.86	92.27	96.26	87.16
Wine	77.78	87.04	90.74	83.33
WordsSynonyms	55.96	66.93	68.97	62.85
Worms	76.62	81.82	62.34	83.12
WormsTwoClass	74.03	77.92	63.64	77.92
synthetic_control	98.67	100.00	100.00	100.00
uWaveGestureLibrary_X	76.13	82.10	83.28	79.51
uWaveGestureLibrary_Y	64.82	73.20	75.38	68.68
uWaveGestureLibrary_Z	73.12	78.03	77.41	76.19
wafer	99.61	99.84	99.81	99.90
yoga	87.10	89.33	88.57	88.17
Wins	18	38	29	27

Critical difference diagrams

If you would like to generate these diagrams, take a look at this code!

Reference

If you re-use this work, please cite:

@InProceedings{IsmailFawaz2019deep,
  Title                    = {Deep Neural Network Ensembles for Time Series Classification},
  Author                   = {Ismail Fawaz, Hassan and Forestier, Germain and Weber, Jonathan and Idoumghar, Lhassane and Muller, Pierre-Alain},
  booktitle                = {IEEE International Joint Conference on Neural Networks},
  Year                     = {2019}
}

Acknowledgement

We would like to thank NVIDIA Corporation for the Quadro P6000 grant and the Mésocentre of Strasbourg for providing access to the cluster.