Add a "neighbor" method to `interpolate_bads` that works without exact sensor positions

[sappelhoff]

Describe the new feature or enhancement

interpolate_bads has several methods implemented, like the most commonly used spherical spline approach, but also "mne", and "nearest" (for fnirs, although it is not documented what this entails).

All of these methods require 3D sensor positions to work, however, sometimes these may not be available.

As an alternative, it would be great to pass an adjacency matrix (see e.g., find_ch_adjacency), and have the interpolation be done as a (weighted) average of the neighbors.

This is something that is already supported in FieldTrip: https://github.com/fieldtrip/fieldtrip/blob/55e8f26facac8eaf16ba7222433eba3080803b5c/ft_channelrepair.m#L3-L7

Describe your proposed implementation

1. There should be a function to easily create an "mne adjacency matrix" from scratch, for example by passing a dict of neighbors:

mydict = {
    ch1: [ch2, ch3],  # mapping a channel to its neighbors
    ...
}
adjacency = mne.channels.make_adjacency(mydict)

2. .interpolate_bads method should have a new method argument called neighbors and a new parameter adjacency, which gets used for neighbor interpolation.

Describe possible alternatives

I am not aware of an alternative for channel interpolation when no sensor positions are available.

Additional context

This proposed functionality would have to exhibit some "smart" behavior for cases where too few "good" neighbors are available. Potentially with an additional parameter n_required_good_neighbors (or similar), and if that threshold is not met, no interpolation happens for the channel in question

[larsoner]

I think idea (2) is fine.

For idea (1) I think I'd prefer to help people understand a bit how to interpret and construct an adjacency matrix. At the end of the day it's an ndarray of shape (n_ch, n_ch) and you put 1's where channels are adjacent. So the hard part is really constructing a suitable mydict. Once you have that the helper you're talking about is very short and I think not too hard to understand:

adj = np.zeros((len(ch_names), len(ch_names))
for fro, tos in mydict.items():
    ii = ch_names.index(fro)
    for to in tos:
        jj = ch_names.index(to)
        adj[ii, jj] = adj[jj, ii] = 1

So is probably better suited to an example/tutorial that talks about what an adjacency matrix is.

[drammock]

+1 for option (2). I prefer the name "custom" rather than "neighbors" for the method where users pass in a custom adjacency matrix.