I am using a CNN based model to do sequence classification. Since training an entire dataset is very expensive, and I have a large set of features needed to try, its impossible for me to select features by full trainings.

Usually a good sanity check before training, is try to let model overfit a small set of training samples, to make sure the model is at least capable of remember a small sample size.

Borrowed from this idea, my question is, can I train a small subset of training data, and use its loss curve as a metric, to select best features? Each training is to test how quick would the training loss converges given a subset of selected features.

  • $\begingroup$ Regarding the details of our problem, I didn't understand few points. Your input is a sequence. Are you using RNN or CNN? Besides, how do you extract the features from the sequence? If the input to your algorithm is a fixed set of feature, having a sequence as their source is not an important issue. $\endgroup$ – DaL Nov 14 '17 at 7:44
  • $\begingroup$ I am actually using CNN to do stock market prediction, the input is a sequence of market behavior, features are extracted using varies methods, price differences, trading volume, indicators etc.. $\endgroup$ – Xer Nov 14 '17 at 7:51
  • $\begingroup$ And how do you get from the sequence to the features? Do you use the convolutions for feature extraction or another layer? $\endgroup$ – DaL Nov 14 '17 at 7:54
  • $\begingroup$ No, its all hand crafted features, hopping it can ease the training. $\endgroup$ – Xer Nov 14 '17 at 7:56

Congratulations! You have suggested independently the Wrapper method for feature selection. Yes, you can use this method. However, consider that the wrapper method is slow since you have to train a model for each iteration. More than that, feature selection is a NP-complete problem so don't expect the optimal subset.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.