A Distance Between Text Documents based on Topic Models and Ground Metric Learning

抄録

We propose a new distance between text documents that builds upon two techniques. We first represent each document in a database as a histogram of topics using the Latent Dirichlet Allocation (LDA) topic model. We then compare two documents by computing the earth mover's distance between their respective topic histograms. The Earth Mover's Distance parameter, which is in that case a metric matrix between topics, is estimated using Ground Metric Learning. We carry out experiments on different text databases that illustrate the interest of our approach.