accuracy decays with more topics about bayespa HOT 2 CLOSED

even for binary classification

pamedlda = medlda.OnlineGibbsMedLDA(num_topic=80, labels=2, words=61188)
pamedlda.train_with_gml('../data/binary_train.gml', batchsize=32)
(pred, ind, acc) = pamedlda.infer_with_gml('../data/binary_test.gml', num_sample=10)

acc = 0.56 for 80 topics, while acc = 0.80 for 20 topics.

from bayespa.

added a parameter called "stepsize", which adjust the weights for each datapoint. This parameter is set to be dataset size / batch size in our ICML paper.

a rough explanation for this phenomenon:

Assume that K is large (say 80). after first few mini-batches, the bayesian posterior is going to be multi-modal due to uncertainty (too many parameters compared to data). therefore, for the latent samples of initial mini-batches as well as the variational approximate to be accurate, we'll set J (#latent samples per data point) to be large.

So any of the three following solutions can mitigate the accuracy denegeration:

• increase J.
• more sweeps over dataset (more than 1).
• set stepsize to be large.

Revisiting the binary classification experiment, the solutions lead to the following results:

• set J = 10, test accuracy = 0.76.
• set pass = 3, test accuracy = 0.81.
• set stepsize = 25, test accuracy = 0.81.

The third solution seems to be most computationally efficient one.

from bayespa.