R/batchLDA.R

In kasparmartens/oxwaspLDA: Topic modelling with LDA

### Without smoothing

#' @export
LDA = function(W, n_topics = 3, max_iter = 100, max_iter_E_step = 10, convergence_threshold=1e-6){
  k = n_topics
  V = ncol(W)
  M = nrow(W)
  alpha <- rep(1, k)
  beta <- rdirichlet(k, rep(1, V))
  gamma <- matrix((alpha + V/k), M, k)
  phi <- array(0, dim = c(V, k, M))

  likelihood = rep(NA, max_iter)
  for(i in 1:max_iter){
    # E step
    obj = E_step(gamma, phi, alpha, beta, W, max_iter_E_step, convergence_threshold)
    likelihood[i] = obj$likelihood
    phi = obj$phi
    gamma = obj$gamma
    # M step
    beta = update_beta(phi, W, k)
    alpha = update_alpha(alpha, gamma, M)
    if(i > 5){
      alpha = update_alpha_vec(alpha, gamma, M)
    }
    cat(sprintf("Iteration %d of EM completed. Likelihood: %1.3f \n", i, likelihood[i]))
    if(check_convergence(likelihood, i, convergence_threshold)) break
  }
  return(list(likelihood = likelihood[1:i], phi = obj$phi, gamma = obj$gamma, alpha = alpha, beta=beta))
}