Stacked Graphical Learning for Text Mining

Опубликовано 6 сентября 2016, 16:48

In the talk, I will introduce an efficient statistical machine learning approach for classifying networked data, stacked graphical learning. In stacked graphical learning, a base learner is first applied to the training data with local attributes only to make predictions in a cross validation-like technique. Then we expand the local feature vectors by generating new features based upon the predictions of relevant instances. Finally the base learner is applied to the expanded feature sets to make the final predictions. I have applied stacked graphical learning to many real problems including document classification, information extraction, and multi-task problems in an information extraction system. I also formally analyze an idealized version of the algorithm, provide proof of convergence of the idealized version of stacking, and discuss the conditions under which the algorithm of stacked graphical learning is nearly identical to the idealized stacked graphical learning. An online version of stacked graphical learning is studied to save training time and to handle large streaming datasets with minimal memory overhead.