# DPP 的构造

DPP通过最大后验概率估计，找到商品集中相关性和多样性最大的子集，从而作为推荐给用户的商品集。

$\inline P(Y) \propto det(L_{Y})=Vol^{2}(B_{i}),i\in Y$

• $\small r_{i}$为item i与user之间的相关性，且$\small r_{i}\geq 0$
• $\small s_{ij}=$为item i 与 item j之间的相似度度量，且$\small \left \| f_{i} \right \|_{2}=1$

$\small L_{ij}===r_{i}r_{j}s_{ij}$

$\dpi{100} \small P(Y) \propto det(L_{Y})=\prod_{i\in Y}r_{i}^{2}\cdot det(S_{Y})$

$\dpi{120} Y_{map}=\mathit{argmax}\begin{Bmatrix} det(L_{Y}) \end{Bmatrix}$

$\dpi{120} f(Y)=log det(L_{Y})$

$\dpi{100} \small f(X\cup {i})-f(X)\geq f(Y\cup {i})-f(Y)$

$\dpi{120} y=\mathit{argmax}\begin{Bmatrix} f(Y\cup {i})-f(Y) \end{Bmatrix}=\mathit{argmax}\begin{Bmatrix} log(detL_{Y\cup {i}})- log(detL_{Y})\end{Bmatrix}$,

# DPP模型求解

$\dpi{100} \small L_{Y_{g}}=VV^{T}$

$\dpi{100} \small L_{Y_{g}\cup {i}}=\begin{bmatrix} L_{Y_{g}}& L_{Y_{g},i} \\ L_{i,Y_{g}}& L_{ii}\end{bmatrix}=\begin{bmatrix} V & 0 \\ c_{i} & d_{i} \end{bmatrix}\begin{bmatrix} V & 0 \\ c_{i} & d_{i} \end{bmatrix}^{T}=\begin{bmatrix} VV^{T} & Vc_{i}^{T}\\ c_{i}V^{T} & d_{i}^{2}+c_{i}^{2} \end{bmatrix}$

$\dpi{100} \small Vc_{i}^{T}=L_{Y,i}$， $\dpi{100} \small d_{i}^{2}=L_{i,i}-\left \| c_{i} \right \|_{2}^{2}$

$\dpi{100} \small log(detL_{Y_{g}\cup {i}})=log(det(L_{Y_{g}}))+log(d_{i}^{2})$

$\dpi{100} \small Vc_{i}^{T}=L_{Y,i}$

$\dpi{100} \small \left\{\begin{matrix} Vc_{i}^{T}=L_{Y_g,i} \\ d_{i}^{2}=L_{i,i}-\left \| c_{i} \right \|_{2}^{2} \end{matrix}\right. \Rightarrow \left\{\begin{matrix}\begin{bmatrix} V& 0\\c_{j} & d_{j} \end{bmatrix}c_{i}^{'}^{T}=L_{Y_{g}\cup {j},i}=\begin{bmatrix} L_{Y_{g},{i}}\\ L_{ji}} \end{bmatrix} \\ \end{matrix}\right.$， $\dpi{100} \small d_{i}^{'}^{2}=L_{i,i}-\left \| c_{i}^{'} \right \|_{2}^{2}$

$\dpi{100} \small Vc_{i}^{T}=L_{Y_g,i}$带入上式中，推导可得：

$\dpi{100} \small c_{i}^{'}=\begin{bmatrix} c_{i} & (L_{ji}-/d_{j}) \end{bmatrix}=\begin{bmatrix} c_{i}&e_{i} \end{bmatrix}$

$\dpi{100} \small d_{i}^{'}^{2}=L_{i,i}-\left \| c_{i}^{'} \right \|_{2}^{2}=L_{i,i}-\left \| c_{i} \right \|_{2}^{2}-e_{i}^{2}=d_{i}^{2}-e_{i}^{2}$

# 滑动窗口式多样性

$\dpi{120} j=\mathit{argmax}\begin{Bmatrix} log(detL_{Y_{g}^{w}\cup {i}})- log(detL_{Y_{g}^{w}})\end{Bmatrix}$

# 总结

©️2019 CSDN 皮肤主题: 大白 设计师: CSDN官方博客