I have a clinical data set. The goal is to develop a model that predicts race based on binary data, whether or not a gene is present.
I am struggling to find a classification model that works without continuous data. One model that might work is k-modes. Are there any other models that I should consider?
Logistic regression models can certainly be used with dichotomous features. It also provides coefficient estimates for each feature so that relationships between features and the target labels can be tested and interpreted. Predictions made from logistic regression models are probabilities rather than binary decisions which can be helpful if you have targets for Type II Error rates, False Omission rates, etc. which is often the case in clinical data.
Distance-based methods can also be used such as the k-nearest neighbor algorithm. With all binary features, it would make sense to use distance measures designed for dichotomous data, such as the Russell Rao distance metric. These models will make predictions based on the class labels of the k-nearest observations in the feature space.
You might want to consider decision trees or random forests, those classifiers can work with non-continuous data, and actually are really good.
They are implemented in scikit-learn.
I recommend using LDA (Latent Dirichlet allocation) which works efficiently with discrete data.
