0

I am training an autoencoder which takes sampled time series sensor data in range [-1024,1024] (0 values is possible). I use mean_squared loss and Adam optimizer. During the training MAE decreases and the network converges after 150 epochs. MAE starts from 9 and converges to 3 on validation set.

However, looking at MAPE the values does not make sense at all. The MAPE is in order [2.9,4.1]x10^6.

I assume MAPE is not a good metric for this dataset but what would be the explanation the very very high MAPE values?

Farnaz
  • 157
  • 4
  • You may have a data point with a value of 0.0001, and a predicted value of 500. The absolute percentage error is on the order of 5e6. The MAPE probably makes no sense at all for your data, and in any case, it makes no sense to optimize on one metric (MSE) and evaluate on another one (MAE, MAPE), or even to evaluate a single prediction on multiple metrics. See the duplicate and especially [Kolassa (2020, *IJF*)](https://doi.org/10.1016/j.ijforecast.2019.02.017) referenced there. – Stephan Kolassa Jan 04 '22 at 09:26

0 Answers0