Traditional machine learning approaches make the assumption that instances are independent and identically distributed (IID). We term models constructed under the IID assumption first-order because in general they only leverage relationships between attributes within instances (e.g., co-occurrence relationships). Thus classification of a single instance (of previously unseen data) is possible because no additional context is needed to infer class membership. Such a context-free approach, however, does not exploit valuable information about relationships between instances in the dataset.
In our research we are developing a novel framework for learning that, unlike approaches that assume instances are IID, leverages implicit co-occurrence relationships between attributes in different instances. We term these implicit co-occurrence relationships higher-order paths. Attributes ( e.g., words in documents in text collections) are richly connected by such higher-order paths, and the model builts by our higher order learners exploit this rich connectivity pattern. In our work to-date we have developed both supervised and unsupervised learning approaches including Higher Order Naive Bayes, Higher Order SVM, Higher Order Classification Based ARM and Distributed Higher Order ARM. We are also have a framework under development that leverages human-computer interaction entitled Distributed Interactive Higher Order Privacy Enhancing Knowledge Discovery (DI HOPE KD).
Week 1: Became associated with the problem and began to look at several previous studies in depth, focusing mainly on one produced by Microsoft.