In the Dynamic Key-Value Memory Network (DKVMN), after obtaining attention weights, the second step is making a prediction. The latent knowledge state in the value memory $\mathbf{M}_t^v$ is read to create a read vector:

$r_t = \sum_{i=1}^N w_{ti} (\mathbf{M}_{ti}^v)^T$

The read vector and the knowledge component embedding $k_t$ are concatenated to generate a feature vector $f_t$. This is used to calculate the probability $p_t$ of the student answering correctly:

$f_t = \tanh(\mathbf{W}_f [r_t, k_t] + \mathbf{b}_f)$ $p_t = \sigma(\mathbf{W}_p f_t + \mathbf{b}_p)$

These functions are applied element-wise, where $\mathbf{W}$ and $\mathbf{b}$ represent weight matrices and bias vectors.