Statistical relational learning through structural analogy and probabilistic generalization
Kenneth D. Forbus, Daniel T. Halstead · 2011
My primary research motivation is the development of a truly generic Machine Learning engine. Towards this, I am exploring the interplay between feature-based representations of data, for which there are powerful statistical machine learning algorithms, and structured representations, which are useful for reasoning and are capable of representing a broader spectrum of information. This places my work in the emergent field of Statistical Relational Learning. I combine the two approaches to representation by using analogy to translate back and forth from a relational space to a reduced feature space. Analogy allows us to narrow the search space by singling out structural likenesses in the data (which become the features) rather than relations, and also gives us a similarity metric for doing unsupervised learning. In the process, we gain several insights about the nature of analogy, and the relationship between similarity and probability.