Collective Classification of Posts to Internet Forums

We investigate automatic classification of posts to Internet forums. We use collective classification methods, which simultaneously classify related objects — in our case, the posts in a thread. Specifically, we compare the Iterative Classification Algorithm (ICA) with Conditional Random Fields and with conventional classifiers (k-Nearest Neighbours and Support Vector Machines). The ICA algorithm invokes a local classifier, for which we use the kNN classifier. Our main contributions are two-fold. First, we define experimental protocols that we believe are suitable for offline evaluation in this domain. Second, by using these protocols to run experiments on two datasets, we show that ICA with kNN has significantly higher accuracy across most of the experimental conditions.