Coarse-to-Fine, Cost-Sensitive Classification of E-Mail

In many real-world scenarios, it is necessary to make judgments at differing levels of granularity due to computational constraints. Particularly when there are a large number of classifications that must be done in a real-time streaming setting and there is a significant difference in the time required to acquire different subsets of features, it is important to have an intelligent strategy for optimizing classification accuracy versus computational costs. Accurate and timely email classification requires trading off the classification granularity with the feature acquisition costs. To solve this problem, we introduce a Granular Cost-Sensitive Classifier (GCSC) which modulates the cost of feature acquisition with the granularity of the classification, allowing inexpensive classification at a coarse level and more costly classification at finer levels of granularity. Our approach can classify messages with greater accuracy while incurring a lower feature acquisition cost relative to baseline classifiers that do not make use of cost information.