2
$\begingroup$

Now I have a task to classify the imbalanced time series datasets using ML classifiers, such as Logistic Regression, Decision Tree, SVM, and KNN. I am not allowed to use the Deep Learning tools, such as CNN and RNN. The time series data is measurements of the Force-Displacement Curve from a production line. The dataset is extremely imbalanced (minority class: majority class= 1:100). So I want to use Data Augmentation techniques to enlarge the size of the minority class, in order to optimize the performance of the classifiers and avoid overfitting.

I have tried many tools in feature-space, such as oversampling, undersampling, SMOTE, ADASYN and so on. But their performance is not so perfect. And I wish to generate synthetic time series data using Data Augmentation techniques, based on the initial data. Similar to Image Recognition, I have also tried the exiting methods which have been applied to images, such as scaling, rotation, and jittering. But they are also not so useful.

So I want to ask if there are any other DA techniques to generate synthetic time series data. I have only some initial ideas, such as using DTW, Fournier Transform, Markov Chain and so on, but no papers or code about applying them.

Can anyone help me? I really appreciate your help. Thank you!

$\endgroup$
  • 1
    $\begingroup$ I think it might be hard to augment time-series data. Have you considered re-weighting your training samples as a strategy to overcome the class imbalance? Re-weighting would probably be easier than DA in this scenario $\endgroup$ – zachdj Nov 7 at 14:40
  • $\begingroup$ @zachdj Thank you zachdj for your comment. But my task is to augment the time series dataset. The optimization in algorithm-level or resampling strategies I have already tried, but they are not my demand. I only want to know how to generate synthetic time series data, namely curves, but preserving the labels and important information and features of initial dataset. $\endgroup$ – Qi Zhang Nov 7 at 14:46
  • $\begingroup$ In that case, this cross-validated post might be helpful: stats.stackexchange.com/questions/320952/… $\endgroup$ – zachdj Nov 7 at 14:54
  • $\begingroup$ The strategies in that post are for forecasting, but some of them may work for your classification problem as well. $\endgroup$ – zachdj Nov 7 at 14:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.