kmdsy / dtcr

Tensorflow implementation of paper 'Learning Representations for Time Series Clustering' (NIPS 2019 accept paper).

Read me file

Chinese version

Tensorflow implementation of paper 'Learning Representations for Time Series Clustering' (NIPS 2019 accept paper). This code is not the official version.

Details

Ma, Q., Zheng, J., Li, S., & Cottrell, G. W. (2019). Learning representations for time series clustering. In Advances in neural information processing systems (pp. 3781-3791).

Bibtex

@inproceedings{ma2019learning,
  title={Learning representations for time series clustering},
  author={Ma, Qianli and Zheng, Jiawei and Li, Sen and Cottrell, Gary W},
  booktitle={Advances in neural information processing systems},
  pages={3781--3791},
  year={2019}
}

Some results

RI (Rand Index) is employed as performance (same as the paper). I used this version of the RI implementation since there is no official implementation method in sklearn package.

I run each experiment runs 5 times and report means and stand deviations. The best column represents the best performance in all the experiments. The paper column lists the RI reported by the paper.

Configs

Config1：encoder_hidden_units = [100, 50, 50], lambda = 1,

Config2：encoder_hidden_units = [100, 50, 50], lambda = 0.1,

Config3：encoder_hidden_units = [100, 50, 50], lambda = 0.01,

Config4：encoder_hidden_units = [100, 50, 50], lambda = 0.001,

Config5：encoder_hidden_units = [50, 30, 30], lambda = 1,

Config6：encoder_hidden_units = [50, 30, 30], lambda = 0.1,

Config7：encoder_hidden_units = [50, 30, 30], lambda = 0.01,

Config8：encoder_hidden_units = [50, 30, 30], lambda = 0.001.

Results

Data preprocessing method: N/A

Dataset	config1	config2	config3	config4	config5	config6	config7	config8	best	paper
ArrowHead	0.63103 ± 0.04962	0.64632 ± 0.02547	0.6402 ± 0.04928	0.66869 ± 0.02821	0.6562 ± 0.0493	0.67823 ± 0.04251	0.64906 ± 0.05363	0.6529 ± 0.03482	0.74023	0.6868 ± 0.0026
Beef	0.7669 ± 0.02558	0.76644 ± 0.02347	0.77471 ± 0.02122	0.77793 ± 0.02044	0.7577 ± 0.00926	0.74897 ± 0.00958	0.75954 ± 0.01854	0.76 ± 0.01204	0.81609	0.8046 ± 0.0018
BeetleFly	0.60526 ± 0	0.61684 ± 0.02316	0.68737 ± 0.10056	0.60526 ± 0	0.60526 ± 0	0.60526 ± 0	0.63053 ± 0.05053	0.67158 ± 0.08497	0.81052	0.9000 ± 0.0001
BirdChicken	0.66211 ± 0.07688	0.58211 ± 0.08346	0.74737 ± 0.03158	0.67632 ± 0.10017	0.54737 ± 0.06781	0.57789 ± 0.10082	0.59684 ± 0.06451	0.61474 ± 0.11087	0.81053	0.8105 ± 0.0033
Car	0.64667 ± 0.03581	0.68316 ± 0.03617	0.71537 ± 0.01632	0.71797 ± 0.01905	0.6304 ± 0.02426	0.65695 ± 0.01937	0.69153 ± 0.018	0.71073 ± 0.03539	0.77401	0.75.1 ± 0.0022
ChlorineConcentration	0.52175 ± 0.01628	0.51549 ± 0.01654	0.5276 ± 0.01301	0.53374 ± 0.00277	0.5222 ± 0.01634	0.51528 ± 0.01587	0.52575 ± 0.0123	0.53555 ± 0.00072	0.53659	0.5357 ± 0.0011
Coffee	0.68624 ± 0.17581	0.65132 ± 0.12575	0.78995 ± 0.10818	0.85397 ± 0.18698	0.58942 ± 0.11309	0.60741 ± 0.04073	0.79365 ± 0.1563	0.82381 ± 0.16011	1	0.9286 ± 0.0016

Data preprocessing method: Normalized

Dataset	config1	config2	config3	config4	config5	config6	config7	config8	best	paper
ArrowHead	0.61923 ± 0.05194	0.61398 ± 0.04337	0.65328 ± 0.02648	0.66475 ± 0.03845	0.6055 ± 0.03643	0.65639 ± 0.03132	0.67137 ± 0.02044	0.66328 ± 0.03323	0.71278	0.6868 ± 0.0026
Beef	0.70713 ± 0.00892	0.70575 ± 0.00497	0.71667 ± 0.01364	0.72337 ± 0.00217	0.70851 ± 0.01202	0.72138 ± 0.01457	0.71552 ± 0.01791	0.72414 ± 0.00291	0.74483	0.8046 ± 0.0018
BeetleFly	0.71842 ± 0.16428	0.67105 ± 0.08392	0.73509 ± 0.06021	0.74211 ± 0.09676	0.62842 ± 0.02836	0.75789 ± 0.10771	0.66421 ± 0.11789	0.74211 ± 0.10458	1	0.9000 ± 0.0001
BirdChicken	0.53579 ± 0.05702	0.58596 ± 0.05674	0.64511 ± 0.09452	0.67193 ± 0.08203	0.50877 ± 0.02796	0.56632 ± 0.09342	0.65 ± 0.10556	0.64868 ± 0.02507	0.81053	0.8105 ± 0.0033
Car	0.70927 ± 0.01742	0.71119 ± 0.02797	0.72249 ± 0.02928	0.7096 ± 0.02585	0.69085 ± 0.01935	0.70395 ± 0.01768	0.71073 ± 0.03181	0.71921 ± 0.01226	0.77288	0.75.1 ± 0.0022
ChlorineConcentration	0.50288 ± 0.00019	0.50821 ± 0.01159	0.51451 ± 0.0144	0.53447 ± 0.00096	0.50255 ± 0.00008	0.5083 ± 0.01156	0.51469 ± 0.01472	0.53519 ± 0.00106	0.053889	0.5357 ± 0.0011

Requirements

Tensorflow>=1.13.2

dtcr's Issues

Can you reproduce the paper's results?

program will suddenly stuck

After training to a certain epoch, the program is stuck without any output.
Forexample，when I train for dataset 'Beef', it will stuck at epoch 175，moreover，the RI doesn't change from the beginning.

update F matrix in k-means loss, epoch: 160.
RI: train: 0.6091954022988506   test:0.5724137931034483
loss: 0.9954168200492859, l_recons: 0.22520238161087036, l_cls: 0.6259787678718567, l_kmeans: 0.14423568546772003, epoch: 161
loss: 1.044954776763916, l_recons: 0.27587103843688965, l_cls: 0.6256064176559448, l_kmeans: 0.14347732067108154, epoch: 162
loss: 0.998985767364502, l_recons: 0.23104754090309143, l_cls: 0.6252334713935852, l_kmeans: 0.14270472526550293, epoch: 163
loss: 0.9973672032356262, l_recons: 0.23056352138519287, l_cls: 0.6248743534088135, l_kmeans: 0.14192931354045868, epoch: 164
loss: 0.9958614110946655, l_recons: 0.23019659519195557, l_cls: 0.6244871616363525, l_kmeans: 0.14117762446403503, epoch: 165
RI: train: 0.6091954022988506   test:0.5724137931034483
loss: 0.9945715665817261, l_recons: 0.2300342172384262, l_cls: 0.6240988373756409, l_kmeans: 0.14043846726417542, epoch: 166
loss: 0.993462860584259, l_recons: 0.23004977405071259, l_cls: 0.6237154006958008, l_kmeans: 0.13969765603542328, epoch: 167
loss: 0.9924675822257996, l_recons: 0.23015378415584564, l_cls: 0.6233416795730591, l_kmeans: 0.13897211849689484, epoch: 168
loss: 0.9914587736129761, l_recons: 0.23025156557559967, l_cls: 0.6229473948478699, l_kmeans: 0.13825981318950653, epoch: 169
loss: 0.9903848171234131, l_recons: 0.23028254508972168, l_cls: 0.6225518584251404, l_kmeans: 0.1375504434108734, epoch: 170
update F matrix in k-means loss, epoch: 170.
RI: train: 0.6091954022988506   test:0.5724137931034483
loss: 0.9892362356185913, l_recons: 0.23023679852485657, l_cls: 0.6221611499786377, l_kmeans: 0.13683833181858063, epoch: 171
loss: 0.9880603551864624, l_recons: 0.23014621436595917, l_cls: 0.6217778921127319, l_kmeans: 0.1361362338066101, epoch: 172
loss: 0.9868771433830261, l_recons: 0.230058491230011, l_cls: 0.6213670372962952, l_kmeans: 0.13545162975788116, epoch: 173
loss: 0.9857420325279236, l_recons: 0.23001040518283844, l_cls: 0.6209614872932434, l_kmeans: 0.13477016985416412, epoch: 174
loss: 0.9846595525741577, l_recons: 0.23001253604888916, l_cls: 0.6205600500106812, l_kmeans: 0.1340869665145874, epoch: 175
RI: train: 0.6091954022988506   test:0.5724137931034483

RI doesnt change as loss decrease.

The RI index hardly changes during the training procces.

I just ran the code with the config2 shown in your readme file but got a very stange performance. Is there anything need to be notived while runnnig the code? Thanks!