I am trying to do time series forecasting following this [1] paper. My data is represented by the total CPU load of a computer as my time series (one data point per second). Secondary features are represented by the UNIX time collected at the same time and converted into hours, minutes and the day of the week.

Obviously, collecting this data requires real time passage. I am doing this project as part of my internship and I am time restrained when it comes to collecting this data (at the moment I only have 9 days of continuous usage).

I came up with the idea of using the generative part of a GAN or maybe a variational autoencoder on my time series in order to simulate more data for training purposes. My question is: would this approach yield any real, useful results?

[1] https://www.sciencedirect.com/science/article/pii/S1877050917326170?viewFullText=true


This question might well be closed as "opinion-based". So I'll give you my opinion.

In my opinion, simulated data are always problematic, because you always get out what you put in. If you can't imagine a data generating process having a certain peculiarity, then the data you simulate won't exhibit it - because you won't even think of simulating it. This is a good case of an "unknown unknown". When you do start seeing such unthought-of features, they will be a Black Swan to you.

GANs can in principle simulate data extremely well, so this is certainly better than simulating your own data. Of course, they are still limited to what they saw when they were trained, or small extrapolations. No European GAN pre-18th century would have generated a black swan. Plus, if you don't have enough data for your actual use case, one could ask whether the amount you have was enough to train a GAN.

Bottom line: it's always better to use actual data. You could ask at OpenData.SE, or consider using similar data to the kind you are actually interested in. I see that the paper you are looking at actually forecasts electricity loads, although you want to apply the models to CPU loads - if you are good with using electricity data, there should be tons of datasets out there, e.g. at Kaggle or in the GEFCom competitions organized by Tao Hong.

If there simply is not enough data and you are pressed for time, then yes, by all means, do simulate your own. GANs are certainly better than simulating your own data. But if so, include a nice, big section on limitations of your study.


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.