The goal is to classify three different cell types based on certain features (e.g. area, shape tensor etc.). However the amount of labelled training data I have is very small. Therefore, it was suggested to create a synthetic data set trying to mimic the real one which we could use to train the classifier. Hopefully, this would allow classification of the real cells.

The question is, how do I create proper synthetic data for this problem?

Someone suggested using averages of the limited data present and adding stochastic variations to the features. However, I don't think that I would really get new information based on this approach. Is this correct?

Therefore, my question is: Can I generate synthetic data based on the present real data or would I simply be adding more of the "same"? If this approach is worthless, are there suggestions for better ways to solve the classification problem?

  • $\begingroup$ A keyword here is Data Augmentation, search for that in context of your problem and you may find some established methods that can be applied. $\endgroup$ – jonnor Mar 31 at 10:07
  • $\begingroup$ The general idea of Data Augmentation is to find attributes of your data which can be changed (with in reasonable limits) without changing the class label. So sit down and think about your data and what you know of cells, and possible candidates for this $\endgroup$ – jonnor Mar 31 at 10:10
  • $\begingroup$ How many samples and how many features do you have? $\endgroup$ – jonnor Mar 31 at 10:11
  • $\begingroup$ Thanks, I have looked a little bit into Data Augmentation now. Some principles might be quite useful. Would you recommend me to completely abandon the mentioned idea regarding synthetic data based on the present data? Since I still have to wait for the data, I am not sure how many samples are actually present. I suppose roughly 50-100 labelled cells for each category. Regarding features, we haven't defined all of them yet but I guess we will have roughly 10 features per cell. $\endgroup$ – cenden Mar 31 at 13:13
  • $\begingroup$ 50-100 per category and 10 features is actually a quite reasonable amount of data for basic classification tasks, depending on intra-class variability. Do you have access to unlabeled data? Then semi-supervised learning techniques can possibly help. For instance you can do unsupervised learning, and use the low amount of labeled data only as the testset $\endgroup$ – jonnor Mar 31 at 13:28

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.