【GCN-RS】Learning Explicit User Interest Boundary for Recommendation (WWW‘22)

Learning Explicit User Interest Boundary for Recommendation (WWW’22)

Pointwise loss:
$$\mathcal{L}=\sum _{(u,x)\in \mathcal{T}}\psi (s(u,x),l(u,x))$$
Pairwise loss:
$$\mathcal{L}=\sum _{(u,p)\in I}\sum _{(u,n)\notin I}\varphi (s(u,n)-s(u,p))$$
Mixed loss ：
$$\mathcal{L}=\underset{L_p}{\underbrace{\sum _{(u,p)\in \mathcal{I}}\varphi \left(b_u-s(u,p)\right)}}+\underset{L_n}{\underbrace{\alpha \sum _{(u,n)\notin I}\varphi \left(s(u,n)-b_u\right)}}$$

$$b_u=W^T{P}_u,\varphi :MarginLossorLnSigmoid$$

It can also be regarded as a kind of debias Methods ,$b_u$ It's a scalar , And the only user embedding of , So for popularity bias Big users ,s Often very large , At this time, set a higher limit for him margin value , It is equivalent to setting a dynamic margin loss.

There's another possibility , Optimize positive samples loss $L_p$, Will increase the positive sample score $s(u,p)$, Reduce the boundary fraction $b_u$, therefore $b_u$ stay loss It can be seen as a kind of high-frequency user Penalty regularization of .