|
Which are you preferred methods for multinomial classification, ie k>2. I am particularly interested in regression methods which allow interpretation of the explanatory variables. |
|
Do you mean multiclass classification? A very popular method is regularized logistic regression: you have weight vectors w_k for each class, classify by choosing the largest dot product w_k·x, and train by maximizing the l2-regularized log likelihood of the data (P(y_k|x) = exp(w_k·x)/sum_i(exp(w_i·x))). This allows you to inspect the posterior for the explanatory variables by looking at the posterior class distribution for an x that only has that feature "turned on". You can also use l1 regularization, which will bring most of the feature weights to 0 and leave only a few relevant ones (at a loss of performance, however). |
