How to predict similarity of unseen data to the training set?

Question

I have a time series of human pose data which are recorded from real humans. I want to train the model with unsupervised learning on the training data. Let's call this the "real" training data. The fake data is generated from moving/rotating joints of the pose.

After the model is trained on the real data, I would want to feed the model "fake" or "real" data and let it tells me how likely is the data to be real. E.g. if the data looks very real, the model tells me a probability value close to 1. If it's fake, return something close to 0.

I want to do this so that I can iteratively adjust the input data such that it maximizes this probability value. The application is to have a fake data, adjust it enough times until it looks real.

I know about GAN but I don't know how to apply it to tabular data or data that are time series (not images)

Please suggest me what kind of model or problem definition that suits my needs. Also any topic that I need to learn about.

Answer 1

The thing you can do is extract some "patterns" from your "real" data, and then compare incoming data to this pattern (take 1 - distance_between_pattern_and_input as your score).

You can find more about time series patterns extraction in this document

Alternatively, you can turn this problem into a supervised learning problem by generating negative examples and train your model to detect whether an input is real or fake.