Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift
Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift
Statistical properties such as mean and variance often change over time in time series, i.e., time-series data suffer from a distribution shift problem. This change in temporal distribution is one...
I found this paper when implementing TSMixer in darts and thought it was worth sharing.
As far as I know, it may not be relevant to non-timeseries problems. That being said, does anybody know of instance normalization outside of timeseries? It is obviously similar to batch normalization but is slightly different.
I added RIN to an N-HiTS model using the Energy benchmark dataset and got 21% and 22% reduction in MSE and MAE, respectively. This required two lines of code to be added to the model definition and increased compute time in no noticeable manner.