Hint mechanism different in the paper? about gain HOT 11 CLOSED

Thank you for the quick answer.

Besides the convergence question, I still have the doubt about the 0.5.

In the paper I understood that the hint was indicating:

• 1 -> known original value
• 0 -> known imputed value
• 0.5 -> unknown

And the discriminator has to define if the 0.5 is an original or an inputed value.

But in this implementation, the hint shows:

• 1 -> known original value
• 0 -> unknown

So the hint is only helping in the known original values, but giving no hint about the missing values?

from gain.

In practice, providing 90% of the mask vector as the hints make the best performance. (Hint is only given to the known features)
In theory (in the paper), providing one feature as a hint converges to the optimal solution with MCAR setting.

from gain.

Yes.
In this code, the hint is only provided to the known variables.
Therefore, the discriminator has to determine if the 0 is an original or an imputed value.
We don't provide the imputed variables as the hint; therefore, we don't need to introduce 0.5 here.
Thanks.

from gain.

I have tested two types of hint : original paper vs. this code
and I found out that the performance of these two models were almost the same.
Although the variance of MSE test loss designed by the original paper(using 0.5 for hint) was bit higher, it didn't seem that meaningful.

from gain.

Usually, on missing completely at random setting, hint does not have a big impact on the results.

from gain.

1.The Imputed Matrix is equal to the Hat_New_X?
2.when i try to print the Hat_New_X, I find that some 0 positions are not imputed ,Is it 0 in the original data?
look forward to your reply

from gain.

1. Yes. G_sample is the output of the generator and Hat_New_X is the matrix that only missing values are replaced by G_sample.
2. Yes. some of them have 0 as the original values.
   Thanks!

from gain.

Thank you for the quick answer.

1. Is the letter data in your codes containing missing values? And has been filled with 0.
2. I can't compare the imputed data with the original dataset because there is no raw dataset
   I recently wrote a paper to quote your paper to impute the data, but the effect is not ideal

from gain.

1. No. The letter data is complete data.

• I introduce the missing in line 51-59 and 210.
• Please check those lines.

2. The original raw data is always there that you can compare.

• Please see line 233 and 186.

from gain.

In the paper, Figure 1 shows, you feed three matrixes, including data matrix, random matrix, mask matrix, but I do not see you feeding random matrix to the generator. What is the random matrix?

from gain.

You can see how we use random matrix in this link (https://github.com/jsyoon0823/GAIN/blob/master/gain.py#L168-L169)

from gain.