Classification of Scientific Papers Using Machine Learning

The project aims to develop a domain-independent and adaptive approach for scientific document classification using both information fromdocument contents and citation links. We evaluate several content-based classification methods including K-nearest neighbours, nearest centroid, naive Bayes and decision trees and find that the naive Bayes outperform other when training set is sufficiently large. Using phrases in addition to words and a good feature selection strategy such as information gain is found to improve system accuracy in comparison with using words only. To combine citation links for classification, the project proposes two methods, linear labelling update and probabilistic labelling update. The two methods iteratively update the labellings of classified documents using categories information from neighbouring documents. Our experiments on the two methods show that, combining contents and citations significantly improves the system performance.