Same setup as the Gaussian version but with binary features: likelihoods[c][j] is $P(\text{feature } j = 1 \mid c)$. Return the unnormalised log score for each class, not a probability.

$$\text{score}_c = \log \pi_c + \sum_j \bigl[x_j \log \theta_{cj} + (1-x_j)\log(1-\theta_{cj})\bigr]$$

Classifiers usually stop here: to pick a class you only need the argmax, and normalising costs an exponential you do not need.

Input

priors = [0.5, 0.5]
likelihoods = [[0.9, 0.1], [0.2, 0.8]]
features = [1, 0]

Output

[-0.9038682118755978, -3.912023005428146]