Thanks for your comment

For me, my CV scores are better for binary-classif problem

I've got 0.04 Log Loss using a single model, here are some tips:

1. Encoding your categorical variables - For most models, they are only able to use numeric data. Beware of encoding techniques that result in large expansion in the feature space that would hurt most models (See: Curse of dimensionality). How you encode the categorical variables impacts your models.

2. Feature engineering ---- Creating additional features from the dataset.

3. Hyperparameter Tuning --- Its not always sufficent to just use the default parameters for your model, you need to optimize the parameters based on what gets you the best results on you cross validation or development set. Using K Fold cross validation here is useful, along with RandomizedSearchCV for example.

3. Bias and Variance ---- If your model is performing poorly on both test and training then your model is not complex enough and maybe use a different model or lower the regularization parameter. If you are doing well on the training set and not the test set then maybe increase regularization.

4. Imbalanced Classes ---- Some models perform poorly when we have imbalanced classes. Try oversampling or undersampling techniques to help, and use metrics to score your models that handles well for imbalanced datasets. i.e Accuracy is a poor metric if your classes are imbalanced.

5. Error Analysis ---- Split into training and test and look at the precision and recall of the different predicted products. Is your model doing better for some products and not in others. Use this as a base to decide next steps as well.

Good tips... Thanks