I have 92 observations with 92 variables. Every observation is a binary outcome (0=no, 1=yes), indicating if that observation co-occurs with a given feature in the feature set. I have 18 classes which these observations can be classified.

I want to use probabilistic classifiers, to help explain potentially co-morbidity between patients. The data I have I encoded myself based on reports given from the literature. I only have one row per disorder, though, and the 18 classes represent "major classes" these disorders are currently categorized under.

I am getting 90% prediction errors using Naive Bayes (God awful!). This isn't inherently shocking, because the problem I'm trying to solve IS the classification framework, but I need to know how to improve my approach so I can better show what I'm trying to do.

My questions: 1. Is Naive Bayes good for binary data? 2. Are there any other classification probabilistic classifiers that may perform better than Naive Bayes for a multi class classification problem? As far as I understand, logistic regression is better for two class problems.

Ideally, I would get higher variances in class percentages predicted by Naive Bayes (or any probabilistic classifier) for patients with co-morbid disease states than I would for patients only diagnosed with one disease. However, what I'm getting right now is really confident Bayes network, with the wrong classifications. Again, this isn't inherently invalid - because the classification (the target labels) are probably not coming from a strong diagnostic framework, but I want to make sure my machine learning knowledge isn't also contributing to this poor classification performance.

Any thoughts will help!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.