Thesis Proposal Stacked Graphical Learning

Traditional machine learning methods assume that instances are independent while in reality there are many relational datasets, such as hyperlinked web pages, scientific literatures with dependencies among citations, social networks, and more. Recent work on graphical models has demonstrated performance improvement on relational data. In my thesis I plan to study a meta-learning scheme called stacked graphical learning. Given a relational template, the stacked graphical model augments a base learner by expanding one instance’s features with predictions on other related instances. The stacked graphical learning is efficient, capable of capturing dependencies easily, and can be constructed based on any kind of base learner. The thesis proposal describes the algorithm for stacked graphical models, evaluates the approach on some real world data, and compares the performance to other methods. For my thesis I plan to explore more of the strengths and weaknesses of the approach, and apply it to tasks in the SLIF system.