#em-algorithm

Question 1: Single choice

mediumAI-checkedProbabilistic & Graphical Models

A naive Bayes model for binary keyword features $\phi_1, \dots, \phi_N$ and a class $y$ is trained by EM on documents whose labels are all missing; $\phi^i$ is the feature vector of document $i$ . Writing $\theta^{(t)}$ for the parameters after $t$ iterations, which quantity is guaranteed never to decrease from one iteration to the next?

Every parameter $P(\phi_k = 1 \mid y, \theta^{(t)})$ 1
The sum of the posteriors $\sum_i P(y_i = 1 \mid \phi^i, \theta^{(t)})$ 2
The observed-feature likelihood $\prod_i P(\phi^i \mid \theta^{(t)})$ 3
The likelihood with each label set to its most probable class4

Timed test