Just the Fax - Differentiating Voice and Fax Phone Lines Using Call Billing Data

We consider the problem of classifying a telephone line as either a voice or fax telephone line based on its long-distance calling patterns (i.e., the time of day and duration of its long distance calls) and based on knowledge about other lines with which calls are placed. We show that both kinds of information are important. In particular, neither choosing new learning algorithms nor collecting more detailed calling pattern data can compensate for the lack of the most important attributes, call duration and neighbor identity. Thus, in our application using the algorithms we studied, we find feature selection to be more important than the choice of learning algorithm.