jimbocoin on Nostr: It’s OK, I probably don’t need it. I’m researching pairwise comparison ...
It’s OK, I probably don’t need it.
I’m researching pairwise comparison aggregation methods for a personal project, and this one keeps coming up.
The main difference between my project and the literature is sparseness. I have a dataset of ~40k items. Users see random pairs and choose between them.
So, by aggregating pairwise rankings, I’d like to assign scores to items. From what I can tell, The Bradley-Terry model assigns one point to whichever item the user preferred and zero to the other.
The problem with this, IMO, is that it throws away half of the data. If a user prefers A to B, and we only give a point to A, then we’ve learned nothing about B (relative to unevaluated options like, say, C).
So I’m considering giving a negative score to items voted against. In the ABC case, after a user votes A over B, the scores would be: A=1, B=-1, C=0.
I was hoping to find existing literature about sparse pairwise ranking aggregation, but haven’t found much yet.
I’m researching pairwise comparison aggregation methods for a personal project, and this one keeps coming up.
The main difference between my project and the literature is sparseness. I have a dataset of ~40k items. Users see random pairs and choose between them.
So, by aggregating pairwise rankings, I’d like to assign scores to items. From what I can tell, The Bradley-Terry model assigns one point to whichever item the user preferred and zero to the other.
The problem with this, IMO, is that it throws away half of the data. If a user prefers A to B, and we only give a point to A, then we’ve learned nothing about B (relative to unevaluated options like, say, C).
So I’m considering giving a negative score to items voted against. In the ABC case, after a user votes A over B, the scores would be: A=1, B=-1, C=0.
I was hoping to find existing literature about sparse pairwise ranking aggregation, but haven’t found much yet.