Many of machine learning methods are considered as "black boxes". Examples of such methods are SVM, Neural Networks, Random forests etc. One may apply sensitivity analysis techniques (as described for example in this paper by Paulo Cortez) in order to assess the importance of each predictor to the entire model. However, this analysis does not answer the (very tricky) question "What are the top predictors that contributed to a given single prediction?".

Provided several criteria are met, one may obtain such answer from Naive Bayes models by comparing the individual posterior values. In linear models, comparison of $a_i \times x_i$ values will also provide good answer to the above question. But what about more complex models?

  • $\begingroup$ for RF: by.tree argument in varUsed() in randomForest package in R will provide part of this information; you can also extract an individual tree (getTree). The question is which single prediction are you interested in and why, and if so why not use CART. $\endgroup$
    – katya
    Oct 31 '14 at 18:54

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.