I am working on a basic fraud detection model. I have about 10 independent features and I am trying to predict if a given transaction is genuine or fraud. Most of the features are categorical and each categorical variable has hundreds of factors.

As each feature has so many factors what I am currently trying to bag factor levels into larger groups. For example assume feature 1 has about 1000 unique factors. Each factor appears multiple times in the training data. I calculated the percentage of fraud for each factor. All factors with less than 10% fraud transactions were put in group A, factors with less than 20% fraud transactions were put in group B and so on.

As I am using the dependent variable for bagging the independent variables I am not sure if my strategy is legit? Can some one help me with this? Alternatively what are other approaches for reducing the number of factors



Browse other questions tagged or ask your own question.