Sorry for the delay guys, here's the 2nd place solution. "https://github.com/nikhilmishradevelop/zindi-aws-sansa-hack/", most of the credit still goes to John :D. What worked for me, I guess was using only positive examples instead of both positive and negative examples from the shp file. Initially I saw that my model was predicting mostly negative examples and only a few percentage of positive examples,(less than 5%). Sorry its badly documented, if you have any queries, please ask.