I get some metrics on validation data while training a model , and in my case the they are :

(0.25, 0.31, 0.46, 0.57, 0.65, 0.75, 0.77, 0.78, 0.84, 0.84, 0.85, 0.84, 0.84, 0.84, 0.82, 0.8, 0.8, 0.79, 0.78, 0.77, 0.77, 0.77, 0.75, 0.74, 0.73, 0.73, 0.73, 0.73, 0.73, 0.73)

They can described like this :

enter image description here

In my view , the ideal result should be like : enter image description here

Is it a matter of overfitting ?

Unfortunately , I tried few times to change the regular coefficients to avoid overfitting , and adjust learning rate coefficients to slow down , but it was still "convex" .

How can I achieve the ideal result showed above ?

Much appreciated if anyone would give me some constructive tips ?

  • $\begingroup$ What model are you training? $\endgroup$ Dec 9, 2016 at 8:34
  • $\begingroup$ General Linear model , such as logic regression $\endgroup$
    – joe
    Dec 9, 2016 at 8:38
  • $\begingroup$ Size of dataset? Type of validation? Still see same effect in k-fold cross-validation? $\endgroup$ Dec 9, 2016 at 9:37
  • $\begingroup$ I used the "toy data" to validate my recommendation system with Spark , so I simplified the whole process . The size of dataset is small (15K) . I tried few tens of times to adjust coefficients , and almost each result came out to be "convex" . $\endgroup$
    – joe
    Dec 9, 2016 at 11:04

1 Answer 1


Yes, what you are seeing is a classic case of overfitting.

You stated that you use a linear model such as logistic regression. To regularize these types of models, usually L1 and/or L2 regularization is applied. L1 regularization is simply $||W||_1$ and L2 is $||W||_2^2$ usually.

Another method is to alter the labels of the model in a specific way, which is a method of regularization I created (shameless plug). Here is the link to the paper: https://arxiv.org/abs/1609.06693

Hope this helps.


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.