I'm participating in a Kaggle contest "What's cooking". The goal is to know wich kind of cuisine we have, depending on some ingredients. So it's a multiclass classification problem. I have an existing model, and I have been trying to improve it for 2 weeks without result. I am using scikit learn, and my existing model is

pipelineLR = Pipeline([
('vect', CountVectorizer(max_df=0.75)),
('lr', LogisticRegression(C=0.4,solver='lbfgs',multi_class='ovr',fit_intercept=True,warm_start=True)),

It's really basic. But the problem of this dataset is that we have unbalanced data. Also we have a a huge sparse matrix because of the CountVectorizer. I tried a lot of different model, with and without TF-IDF and this model is the best I can get. I have also confusion matrix, classification report and learning curve for it.

What I can notice is that is have high-variance. I think I can do something by preprocessing the data but I am really stuck. I don't know if my model is good enough to be improved, or if I have to find a totally new approach to it. Also what I can get from the classification report is that the model is working well for classes where I have a lot of data, but it's making a lot of mistakes for classes with less data.

I am very new to data science and I am trying to learn by myself. But I think I really need to have another point of view with this problem, because I can't manage to improve my score.

Classification report Learning curves


1 Answer 1


But the problem of this dataset is that we have unbalanced data

I think that the way to fix your problem is to use something like SMOTE or one of its variants. (My favorite is SMOTE-ENN, but go with what works the best). Here is an implementation that you can use to fix the class imbalances. I don't know if this will solve your specific problem, but it is worth a shot.


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.