This is log for my iterations of baseline methods. Multiple features are included throughout iterations.
I used plain SVM, or sklearn.svm.svr, achieving really poor performance. Don't know why.
I used linear model, slkearn.linear_model.LogisticRegresseion to predict. Poor result even worse than Follower baseline
.
Transformed features by log(1+x), gets worse than BaselineSolution.
Use model.predict_proba
instead of model.predict
, yeilding 0~1 probability rather than {0,1} labels, achieving desired baseline method performance.