1
$\begingroup$

After reading this and this , I tried it on mine by fitting the 2-input model i.e. text and numerical. The result remains similar even several attempts on tuning the hyperparameters e.g. embedding dimension, epoch, activation. What is causing this behaviour? How can it be improved? The data is imbalanced with yes=65% and no=35%.

Improve this question
$\endgroup$
3
  • $\begingroup$ You need to elaborate on "I tried it on mine". $\endgroup$
    – gunes
    Jun 5, 2022 at 9:10
  • $\begingroup$ Sorry, I mean my data. $\endgroup$
    – user357565
    Jun 5, 2022 at 9:15
  • $\begingroup$ The results look like your model hasn't learned much of anything, predicting basically the prior (every point as 65% chance of being positive). It's really hard to say much else without the data or architecture or hyperparameters. Can you get anything useful from either text or numeric inputs alone? Can a simpler model produce anything useful? $\endgroup$
    – Ben Reiniger
    Jun 5, 2022 at 12:41

0

Reset to default