2

I have trained a regression model with 7 features for a given problem. Now, I have another regression problem (quite similar to the previous one) where I have only 6 samples in hand, but with 3 more features than the first model (7+3). The correlation between the target value and these 3 additional features is very high. So I would like to use transfer learning method to create a new model based on the old model; but I can not find a way to integrate the additional features into the model, since the old one was trained only on X features. On the other hand I must employ the 3 additional features in the new model which means that I cannot just remove or ignore them.

Thanks.

jojo
  • 153
  • 4

1 Answers1

1

One way: Use your old model's prediction as a new feature and combine with the additional three you have, i.e. a four-length feature vector for each sample. Then, train with these new features. So, the old model will act like a static feature generator for you. Of course, you may have challenges with that low number of samples.

gunes
  • 49,700
  • 3
  • 39
  • 75
  • the reason of which I've turned to transfer learning is the issue of few samples of the second problem. with your proposed solution I think that I will encounter the same challenge – jojo Sep 24 '21 at 19:11
  • If you are going to train & modify a transferred model, even there is no additional features, you need a data volume. Otherwise, you'll end up using the transferred model as is. e.g. I can't imagine continuing training Inception with a few images. – gunes Sep 24 '21 at 19:14
  • 1
    @jojo with only 6 samples and 3 new features you are at very high risk of overfitting, getting a model that maybe works OK on those samples but doesn't generalize well. The suggestion in this answer (+1) might be the best you can do without getting more data. – EdM Sep 24 '21 at 19:30