I applied Xgboost and Randomforest to selected top 50 features from 500 features by taking the mean of feature importance returned by the models. And then, built the Xgboost model and got good results. In the second iteration, I added a second dataset with 65 new features to the existing 500 features and repeated the process to select top 50 features, in the selected features 40 were from the previous Top 50 and 10 where from the second dataset. However, when built the final Xgboost model keeping everything same the accuracy reduced. I was expecting accuracy should either improve or remain the same but in this case the accuracy reduced. Is there any wrong in what I'm doing? Thanks Continue reading...
