Fig. 3.

We update the key model ${\tilde{M}}_k$ periodically to prevent feature drift. The new key model ${\tilde{M}}_{k+1}$ is selected by choosing a model from the set ${\left\{{\widehat{M}}_i\right\}}_{i=1}^N$ that has the lowest error e when compared to the previous key model ${\tilde{M}}_k$