I have a time series on which I've trained an autoencoder for anomaly detection (AD) purposes.
My time-series is 6 years worth of daily data, but for AD purposes, I am looking at one week window. This means I've only got 330 sub-sequences for my purposes.
I used 10% of those sub-sequences as my validation set, so 33 samples. It's not much, but it my model generalizes well to these few samples.
For AD, I want to estimate the parameters of the residual distribution. I need these estimates in order to calculate an anomaly score using Mahalanobis distance. However, I don't think 33 samples is enough to estimate such parameters. Using the entire dataset seems like a bad idea, a sort of data leak.
Is it possible to synthesize new data in order to expand my validation set?
I thought of adding "random" noise to some samples, but I'm pretty sure that would make the parameter estimates skewed as it might change the distribution in the wrong direction.