1: \begin{abstract}

2: Missing responses is a missing data format in which outcomes are not always observed.

3: In this work we develop kernel machines that can handle missing responses. First, we propose a kernel machine family that uses mainly the complete cases. For the quadratic loss, we then propose a family of doubly-robust kernel machines. The proposed kernel-machine estimators

4: can be applied to both regression and classification problems. We prove oracle inequalities

5: for the finite-sample differences between the kernel machine risk and Bayes risk.

6: We use these oracle inequalities to prove consistency and to calculate convergence rates.

7: We demonstrate the performance of the two proposed kernel-machine families using both

8: a simulation study and a real-world data analysis.

9: \end{abstract}

10: