A1299
Title: Time-varying weighted latent Dirichlet allocation
Authors: Louisa Kontoghiorghes - Kings College London (United Kingdom) [presenting]
George Kapetanios - Kings College London (United Kingdom)
Abstract: A time-varying weighted latent Dirichlet allocation (wLDA), a generative probabilistic topic modeling method that can track the evolution of topics in a series of documents, is introduced. This approach combines the latent Dirichlet allocation (LDA) with time-varying weights, estimating the model's corpus-dependent parameters with the weighted log-likelihood. In the time-varying wLDA, the estimation of time-varying topics is done using the most recent documents according to the time-varying weights in each time index, the so-called rolling window estimation. This approach accounts for the importance of documents, giving terms in more influential documents greater contribution to the topic distribution estimation. In addition, this methodology addresses topic estimation in the presence of an imbalanced number of documents in each set within the corpus. The methodology is applied to assess the topic evolution of the abstracts of a scientific conference.
