How can this information be fed into the clustering algorithm to make it more precise?
An exception would have to be entered manually, I would assume.
This might be one of the challenges that is impeding the development of the banning software, since I'm pretty sure you'd be feeding transaction chains into a neural network, and the only way to take this information into account would be by adjusting weights, but that's assuming the clustering is using NNs in the first place.
The most reliable heuristics in order are:
-Address reuse
-Common input consolidation
-Matching input script type to change script type
-Other matching change fingerprints (like version/nlocktime value, RBF flag, fees paid in sat/vbyte or in total sats) when all output scripts are the same
The less reliable ones would be:
-Round payment amount
-1 output transfers, which could be a self spend or a payment
-Matching origins/destinations