Tracking the Relevance of Academic Conferences

An academic conference typically span a wide range of topics which changes with time. It is known that topics of conferences in different areas may overlap. Constantly, researchers are interested in the trend of a conference like, what the most popular topic/method/problem in this year’s papers is. The relevance of two different conferences reflects such a change as it is an indicator of interdisciplinary research. For example, The wide application of Neural Networks in Computer Vision may lead to increasing relevance of CVPR and NIPS. In this project, I want to propose a measure of relevance and give a quantitative demonstration of how the relevance changes with time. A paper is simply a document while a conference consisting of a number of papers can be regarded as a corpus. It is interesting how we can find features or representations of documents and further calculate the relevance between them. TF-IDF is a kind of word frequency feature that has been widely used for document representation and similarity matching[1], [2]. Recently with the success of word embedding, there are attempts to map a document direct into vector space(doc2vec[3]). All papers of a conference form a rich corpus which is suitable for Topic Modeling. Among all possible solutions, Latent Dirichlet allocation(LDA) has been proven an efficient method to estimate the latent topics of a corpus[4]. A number of variants were further proposed, to enable modeling with time[5], authors[6] or even multiple corpora[7]. LDA-based methods give possible relevance metrics based on latent variables or probability distribution. Recently, Word Embedding methods like word2vec[8] also provide possible distance metrics for different topics. The goal of this project includes the following two aspects: (I) to model or to represent different conferences over different years; (II) to measure the relevance of conferences. In this project we propose four approaches to this problem, and evaluate them both quantitatively and qualitatively.