I have a situation where the misclassification cost depends on the instance, i.e. on the independent variables. In my training set I have for each instance the independent variables plus a vector of costs for each possible class, where the cost for the correct class is 0 and the cost for incorrect classes can be any value > 0.
It is easy to see how those costs could be used directly for a number of machine learning algorithms, but I could not find a tool which would directly support this. For example Weka only seems to support a global misclassification cost matrix, but not per-instance costs. I think the same is true for SciKit-Learn.
Is there any way to make this work with Weka or SciKit learn or are there other tools / frameworks where this can be done out of the box?
Is there a standard way to represent a training set fro such a scenario in an external file, e.g. in a format similar to ARFF format?