Trouble shooting for final project

Trouble shooting for final project

by Teresa Kortz -
Number of replies: 1

Hello,

For my final project model, the outcome is known – in-hospital mortality – and it is binary, so I chose supervised machine learning algorithms (classification or regression). I also want a more interpretable model and one that provides a probability of mortality and not just predicted categories, so I explored a few different models: stepwise logistic regression, Bayesian network, C5 classification tree, CART classification tree. However, when I run these models, I get an AUC of ~0.5. I then ran the ‘Autoclassifer’ node without selecting models. SPSS selected XGBoost Tree, Random Tree, CHAID, and Decision List as the four highest performing models and the AUCs improved to 0.6-0.7 in the testing set. Is my problem  that I was selecting the wrong learning algorithms for the data or something else?

Thanks,
Teresa

In reply to Teresa Kortz

Re: Trouble shooting for final project

by Chuck Mcculloch -

I don't think it is the fault of the choice of algorithms.  C5 or logistic would be fine.  It is possible that an algorithm like C5 is not finding any useful predictors and therefore having no predictive power.  

You might try with a small number of input variables you know are associated.