I am using the random Forrest algorithm to classify data points based upon 10 features. The accuracy score from these forests has been low (5-15% above random). I have a theory that there is a subset of highly predictable data within the dataset that shares similar features, but that the rest is essentially random - giving low accuracy meterics. How can I test this theory?
One approach I have thought of involves clustering the data and seeing if any clusters co-localize with predictions. For example, individuals within cluster X have a 95% chance of being class Y. What do you think? Thanks so much for the advice!