$$ x'_{t,i} = x_{t,i} * M_{t, i} + b_{t,i} * (1 - M_{t, i}) $$

where $b_{t,i}$ is usually generated based on the data distribution.

$$ L(M) = \sum_{c=1}^{C} f(x) \log f(x') $$

References

Crabbé, J., & Van Der Schaar, M. (2021, July). Explaining time series predictions with dynamic masks. In International Conference on Machine Learning (pp. 2166-2177). PMLR.