performance decreasing after increasing training data

ssv112

Hello my model was giving better performance by using 60;40 ratio of dataset. However when i made it 70:30 , performance decreased and even for any other combination except 60:40 was best. Can anyone explain whats going on??

varunm1

Hello @ssv112

There may be several reasons, the 10 percent data which you moved from test to train might change the model dynamics during training. The 30 percent test data might not capture all underlying relations in original dataset. Models are sensitive to the input data and even the performances are sensitive to the test data. Change in test data changes model performance, this is one disadvantage of split. It might give superb performance for one split but worse for another split. Cross validation is recommended for this same reason.

Hope this helps.