.
Date read: 2021-04-13
Paper: https://arxiv.org/pdf/2006.11011.pdf
Aims to separate out user ‘conformity’ from interest. The motivation being that recommendation models suffer popularity bias, which you can account for (with weightings etc.), but that doesn’t separate out who is consuming something based on conformity or because of genuine interest.
DICE - Disentangling Interest and Conformity with Causal Embedding.
Term to remember: IID - Independent and identically distributed random variables.
I haven’t read the paper in detail and whilst the methodology is complicated, from my understanding the approach comes down to one key feature, which is splitting the data out into conformity and interest driven clicks.
For example, consider you have item A and B shown to a user. A is more popular in general than B, then if:
Ca + Ia > Cb + Ib
as well as Ca > Cb
Cb + Ib > Ca + Ia
, whilst Ca > Cb
still remains and we also now know that Ib > Ia
C
- conformity
I
- interest
The data is split out based on the above, then trained separately and the combined loss is optimised together. Based on their testing at least, it seems to work very well, but I think the main interest seems to be it teasing out actual interest which makes it more generalisable and interesting for post recommendation model analysis.