The Penn Discourse Treebank (PDTB) is a data set that promotes the advancement of dis- course analysis tools. Improving discourse analysis is useful for other areas of Natural Language Processing, but it still proves to be a challenging task. This research focuses on improving sense relation classification in the PDTB for implicit relations at all three levels. The features selected for classification are motivated by prior research and statistical testing in this research. The goal is not only to provide features that improve classification in PDTB, but also to select features which are broad enough to be effective beyond the scope of the PDTB. Moreover, these features are derived from a variety of categories such as Semantics, Syntax and Entity in order to ensure stronger results. Using these features, of which many are new for PDTB sense classification, and Naive Bayes, Maximum Entropy and SVM, this research shows improvement in a number of senses that prior research has had difficulty improving, mainly Comparison and Contingency.