Clustering and topic discovery in scientific literature

Discovery of latent semantic groupings and identification of intrinsic structures in data collections is a crucial task for many data analysis needs. Various unsupervised machine learning algorithms have been devised to accomplish this task for a vast number of applications, including topic identification in text databases, clustering similar images in web search, disease identification in medical fields to name a few. Most algorithms, however, have been designed for homogeneous data where the algorithm works on a uniform set of attributes that represent the data objects. Real-world datasets, on the other hand, are richer in structure and contain multiple levels of connectivity among the data objects, such as hyperlinks in web pages, textual annotations or text contexts surrounding images on the web and citations and authorship information of scientific literature. Utilizing only a single information source provides a narrow focus of view into the real nature of the relationships between data objects. Each additional dimension of connectivity increases our understanding of the semantic characteristics of the collection and improves our ability to detect distinct groups of objects where the objects in each group exhibit similar properties. This thesis presents three algorithms that merge multiple sources of information for clustering and topic discovery in collections of academic papers. The first algorithm combines textual content of academic papers with the information extracted from the citation relationships between the papers for finding scientific topic clusters in the data collection. The second algorithm integrates authorship information of documents into the text-based clustering process to yield improved clustering solutions. Based on the validation from these two algorithms that it is possible to improve topic discovery by utilizing additional dimensions of similarity among data objects, we provide a generative model that merges citation relationships, authorship information, user queries, user tags and the timestamps of documents for discovering scientific topics in collections of academic papers. Further, the generative process can effectively model the evolutionary characteristics of document collections and can discover the change of the popularity of scientific topics over time.