I understand this is a wide question.
But there can be some suggestions.
I can try some methods which I do not know.
I think the model is already prefect on train data. But the test accuracy is a little low. So do you think there still can be some improvement on the deep model?
Tong Guo You can use regularization techniques, such as L2 regularization and dropout, to reduce overfitting.
Are you using k-fold cross-validation for training evaluation? Maybe your validation set is easy.
Tong Guo You can use regularization techniques, such as L2 regularization and dropout, to reduce overfitting.
Are you using k-fold cross-validation for training evaluation? Maybe your validation set is easy.
1. Look at your data. They should be clear (for example, age of consumers should be > 16, and < 100).
2. Look at some input variables:
if variance low or the values are imbalanced (5,5,5,5, 1, 5,5,5,5) they should be exclude
3. Try to use PCA - pricipal component analysis to extract few main components of input data, or may be
4. As Antonio Busson wrote: do not use the same train and test samples, try to resample them and use K-fold cross validation or something else.
Your model is best on train course of overfitting, very good for train data, but not for the real world.
5 Try to use varios metric (not only the accuacy), for example, precision / recall ,
ROC-curve etc.
... and various methods (NN, Regression, Decision Trees) :)
"The model is already prefect on train data. But the test accuracy is a little low. "
It is the first main problem when all works are done.
90% that you will get only test accuracy at all (train accuracy is high course of overfitting).
have you tried different approach to train and test the model?
Try hyper-parameter optimization using grid-search technique. Different hyper-parameters such as activation functions (relu, tanh, softplus), learning rate, optimizers (Adam, RMSProp, SGD), regularization (L1, L2), dropout rate could help.
- Badges
- Science topic