Prediction for Bernoulli model with covariates seems incorrect

[jchiquet]

In reference to GrossSBM/sbm#3,

if the following couple of lines are correct,

Then it seems that the prediction function in the case of Bernoulli for LBM/SBM with covariates is unexact, isn't it ?

    B <- matrix(0, nrow(adj), ncol(adj))
    for (k in 1:length(covariates)) {
        B <- B + model_parameters[[Q]]$beta[k] * covariates[[k]]
    }
    if (membership_name == "LBM") {
        return(sigmo(memberships[[Q]]$Z1 %*% model_parameters[[Q]]$m %*% 
            t(memberships[[Q]]$Z2) + B))
    }
    else {
        return(sigmo(memberships[[Q]]$Z %*% model_parameters[[Q]]$m %*% 
            t(memberships[[Q]]$Z) + B))

It would be correct only of the Z are indicator matrices but they are posterior probabilities.

Am I missing something?

[jb-leger]

You are right. I will fix it. Thanks.

[jchiquet]

What an efficient, straight to the point answer!

[jb-leger]

I need some time to write a vectorized implementation. The problem is easy to understand, I wrote this with indicator matrices, and I didn't use my brain to obtain the prediction.

[jchiquet]

Same problem on my side ;-) Especially since most of the time, at the end of the day, posterior probabilities are almost indicators...