#gaussian-process

Question 1: Single choice

mediumAI-checkedNearest Neighbors & Kernel Methods

Each gene's expression curve is modelled as a sample from a Gaussian process: the measurements $\mathbf{y}$ at time points $t_1, \dots, t_r$ have distribution $\mathcal{N}(\mathbf{0}, K)$ , where $K_{kl} = k(t_k, t_l) + \sigma_n^2\delta_{kl}$ . For some genes, a few of the $r$ measurements are missing. How should the likelihood of such a gene be computed?

Fill the missing entries with 0, the prior mean, and use the full $r \times r$ matrix $K$ 1
Drop every gene that has a missing value from the analysis2
Use only the observed entries of $\mathbf{y}$ and the matching rows and columns of $K$ 3
Set the rows and columns of $K$ for the missing times to zero and keep its size4

Timed test