Gap between prediction and results


I wonder if there is any particular reason to explain a huge gap between prediction and score received when submitting data.

My f1 score is close to some of the best ranked results, however when submitting results I obtain a really low (far too low) f1 score.

Any advice would be very much appreciated.


Your model does not generalized results well. It is probably overfitting issue. You should first look how you split the data into the training and testing sets. Then explore what could cause low variance in your ML model. In addition you may want to construct/engineer some new data from the existing ones and use it as a replacement to the data you use for this purpose. For example if you are trying to predict MPG (miles per gallon) and your original data only shows the GPS x&y (Northings &Eastings), you should create a new data vector as distance and use it instead x&y. Why? Because the MPG is a function of distance and not of the direction of driving. If your test/train data contain information for driving to Canada and your submission data are for driving to Mexico, your submission score will be low if you use only x&y location data. There are other things you may want to try, but I will leave them to you to Google. Good luck!