【RS-Defence】Detection of Shilling Attack Based on Bayesian Model and User Embedding (ICTAI‘2018)

Detection of Shilling Attack Based on Bayesian Model and User Embedding (ICTAI’2018)

直接将user embedding输入到分类器，因此需要强化一下user embedding，把更多的信息加进去。

三个loss训练user embedding。

loss1：Matrix factorization

矩阵分解正常loss，分解user-item评分矩阵$R$
$$L=\sum _{u\in m,i\in n}\left(r_{u,i}-{\hat{r}}_{u,i}\right)+\lambda \left(\sum _u{\Vert p_u\Vert }^2+\sum _i{\Vert q_i\Vert }^2\right)$$

loss2：User embedding

大致思路是构建一个user-user矩阵$M$，用user embedding分解矩阵$M$。

构建user-user 矩阵$M$

构建user-user SPPMI (Shifted Positive Point Mutual Information)，构建方式：
P M I ( u , v ) = log ⁡ # ( u , v ) ⋅ ∣ D ∣ # u ⋅ # v S P P M I ( u , v ) = max ⁡ { P M I ( u , v ) − log ⁡ s , 0 } \begin{gathered} P M I(u, v)=\log \frac{\#(u, v) \cdot|D|}{\# u \cdot \# v} \\ S P P M I(u, v)=\max \{P M I(u, v)-\log s, 0\} \end{gathered} PMI(u,v)=log#u⋅#v#(u,v)⋅∣D∣​SPPMI(u,v)=max{ PMI(u,v)−logs,0}​
where $\#(u,v)$ denotes the number of items that have been jointly rated by both user $u$ and user $v$, $\#(u)=<$