Subject continuity

The toolbox features a subject continuity loss. This may be useful if the model is trained on a dataset where the assigned subjects are expected to change smoothly and continuosly over time. This could be the case when modeling slowly changing, non-stationary data.

This can be configured using the dsr_continuous_subject_groups (list[list[int]]) parameter. Each entry in the outside list is a list of subject indices, which is regularized such that the subject vectors are continuosly changing:

$$\mathcal{L}_n = \sum_{i=1}^{N_{n} - 1}||p^{(i)} - p^{(i+1)}||^2,$$

where $n$ is the index of the entry and $p$ is a subject vector. Note that the index $i$ goes over the given list at the $n$th entry, thus the order of continuity does not have to match the order of the subject indices.

The regularization can be weighted using the alpha_subject_continuity parameter, which is part of the value_scheduler.