The main idea is to combine the MeanTeacher with the series saliency module. While improving the accuracy of the model, it can also enhance the interpretability quantitatively and qualitatively. Compared with the above work that only improves accuracy, it may provide more insights.
sklearn 0.22.1numpy 1.16.4pytorch >= 1.7torchgeometry 0.1.2
The data composed of 6 publicly available datasets downloadable from (Download link), and save them under the directory datasets/.
The following are the detailed parameters of the three data sets I have completed the experiment.
| Dataset | Train | Test | Dimension | Class |
|---|---|---|---|---|
| UWaveGestureLibraryAll | 2688 | 894 | 945 | 8 |
| CricketX | 458 | 156 | 300 | 12 |
| InsectWingbeatSound | 1320 | 440 | 256 | 11 |
mainOurs.py includes some options. The following are the important options, the python script takes along with their description:
-
--dataset- The experiments include 16 datasets. The previous papers mainly design experiments on the 6 datasets. Up until now, for each dataset, I ran for 5 times (random seed 0,1,2) and recorded the mean and variance. As shown in the experiments, there is a significant improvement compared with the previous SOTA results.
-
--model_name- It includes three opinions.
SupCE: The supervised training procedureSemiTime: The previous SOTA baselinesVT2: Our method ( VT2 + Series Saliency).
- It includes three opinions.
-
--label_ratio- The option is used to limit the proportion of labeled data.
-
--Saliency- The option is to indicate whether using series saliency module in the MeanTeacher.
-
Other parameters are some detailed parameters.
optim/- Under the
optim/directory, there are several semi supervised learning method.generalWay.pyincludes our implement methodpretrain.pyincludes the baseline
- Under the
model/- The mainly DL architecture is Temporal Convolution neural network
Dataloader/- The directory is necessary, including some data loaders that read the UCR time series classification data. The data used to calculate the consistency loss sample from both labelled and unlabeled data in our implementation.
After introducing the results of previous code, some examples for running commands.
## MeanTeacher
python mainOurs.py --model_name VT2 --dataset=CricketX --gpu=2 --label_ratio 0.4
## Semi time Method
python mainOurs.py --model_name SemiTime --dataset=CricketX --gpu=2 --label_ratio 0.4
## Supervised method
python mainOurs.py --model_name SupCE --dataset=CricketX --gpu=2 --label_ratio 0.4
The model architecture is intuitive, which used VT2 method to the semi-supervised learning of time series. We combine it with the previously proposed series saliency module. As shown in the figure, we can guess the design idea of the model. The implementation details are in code. At present, the algorithm significantly improves accuracy. On the other hand, we validated the series saliency module is helpful in semi-supervised learning. This is good news! 🎉 🎉 😄
The second part is to use the series saliency for interpretation in time series semi-supervised learning. I will implement the codes, and migrate from time series forecasting to time series classification. We'll provide more quantitative and qualitative analysis. The motivation is to observe learning procedure with increasing label size. The phenomenon may require more domain knowledge and cherry-pick some visualization.
Finally, I think that easy-to-implementation series saliency can significantly improve prediction accuracy and interpretability, contributing to the time series semi-supervised learning!
We mainly compare the latest two papers on time series supervised learning. The second paper reproduces the results of the first paper. There are some other baseline methods are implemented in second paper.(like Pi model or pseudo label). I only put the strongest baseline in the table. Therefore, we will compare their methods. The experiment results show that the series saliency is also an effective augmentation. Now the more visualization results will be added (like t-sne).
| Label Ratio | 10% | 20% | 40% | 100% |
|---|---|---|---|---|
| Dataset | ChinaTown | |||
| SemiTime | 44.88 (3.13) | 51.61 (1.22) | 58.71 (2.78) | 65.66 (1.58) |
| MeanTeacher | 45.54 (1.16) | 51.59 (1.98) | 62.87 (1.69) | 67.32(0.12) |
| MT w/ SS | 47.31(2.21) | 53.87(1.12) | 63.45 (1.28) | 69.31(0.11) |
| Dataset | MFPT | |||
| SemiTime | 54.96 (1.61) | 59.01 (1.56) | 62.38 (0.76) | 66.57 (0.67) |
| MeanTeacher | 56.33 (2.1) | 61.21 (2.17) | 63.37(0.92) | 67.53(1.98) |
| VT2 w/ SS | 57.24 (2.27) | 61.47 (1.91) | 64.9 (2.1) | 68.99(1.98) |
| Dataset | Epilep | |||
| SemiTime | 81.46(0.60) | 84.57(0.49) | 86.91(0.47) | 90.29(0.32) |
| MeanTeacher | 91.92 (1.52) | 92.11(0.32) | 94.37(0.30) | 95.13(0.21) |
| VT2 w/ SS | 92.28 (0.51) | 94.94(0.68) | 96.36(0.71) | 97.11(0.11) |
| Dataset | MFPT | |||
| SemiTime | 64.16(0.85) | 69.84(0.94) | 76.49 (0.54) | 84.33(0.50) |
| MeanTeacher | ||||
| MT w/ SS | ||||
| Dataset | Epilep | |||
| SemiTime | 74.86(0.42) | 75.54(0.63) | 77.01(0.79) | 79.26(1.20) |
| MeanTeacher | ||||
| MT w/ SS | ||||
The two papers were not presented at the top-tier conference. I think the main reason is the lack of further analysis for semi-supervised learning.
[1]SEMI-SUPERVISED TIME SERIES CLASSIFICATION BY TEMPORAL RELATION PREDICTION
[2]Self-Supervised Time Series Representation Learning by Inter-Intra Relational Reasoning
[3]Self-supervised Learning for Semi-supervised Time Series Classification