Illustration of the overfitting of the missing data that could appear in the autoencoder process and the solution proposed to overcome it. The input light curve is composed of different magnitudes (m₀,  m₁,  m₂ m₃) and missing values represented by zero values. In case 1, the algorithm has completely overfitted the missing data by replacing them at the same position on the light curve. So the loss function, , is ideally low. In case 2 the algorithm has completed the missing data by interpolating them. However, as the computation of the loss is made between the new values of magnitudes, (), compared to zero values, the value of the loss is overestimated. The solution that we provided is to multiply the interpolated light curve by a mask M before the computation of the loss, .