DescriptionModern statistical machine learning techniques often rely on the assumption that data instances are independent and identically distributed (IID). However, recent work in statistical relational learning has demonstrated the utility of violating the independence assumption. Specifically, the research has shown the value of leveraging relationships between data instances based on higher-order paths. In this thesis, I present a novel Higher Order Collective Classifier (HOCC), a statistical relational machine learning technique that leverages latent information present in higher-order co-occurrences of items across data instances. A general framework is presented in which HOCC can be applied to event detection in time series data. Given the importance of cyber-security, HOCC is applied to two different data sets in the cyber-security domain: first, a Border Gateway Protocol (BGP) dataset, for detection and classification of anomalies, and second, a Network File System dataset for building models of user activity for masquerade detection. Performance of HOCC compares favorably against first-order models that do not leverage higher-order information, achieving separation of classes that heretofore were difficult to separate.