Exploiting Concept Clumping for Efficient Incremental E-Mail Categorization

We introduce a novel approach to incremental e-mail categorization based on identifying and exploiting “clumps” of messages that are classified similarly. Clumping reflects the local coherence of a classification scheme and is particularly important in a setting where the classification scheme is dynamically changing, such as in e-mail categorization. We propose a number of metrics to quantify the degree of clumping in a series of messages. We then present a number of fast, incremental methods to categorize messages and compare the performance of these methods with measures of the clumping in the datasets to show how clumping is being exploited by these methods. The methods are tested on 7 large real-world e-mail datasets of 7 users from the Enron corpus, where each message is classified into one folder. We show that our methods perform well and provide accuracy comparable to several common machine learning algorithms, but with much greater computational efficiency.