机器学习（九） — 协同滤波与基于内容的滤波

Collaborative and Content-based filter

1 collaborative filter

1. definition

recommend items based on rating of users who gave similar rating

2. cost function

1. learn $w^{(1)},b^{(1)},\cdots ,w^{(n_u)},b^{(n_u)}$

$$J=\frac{1}{2}\sum _{j=1}^{n_u}\sum _{i:r(i,j)=1}(w^{(j)}\cdot x^{(i)}-b^{(j)}-y^{(i,j)}{)}^2+\frac{\lambda }{2}\sum _{j=1}^{n_u}\sum _{k=1}^n(w_k^{(j)}{)}^2$$

2. learn $x^{(1)},\cdots ,x^{(n_m)}$

$$J=\frac{1}{2}\sum _{j=1}^{n_u}\sum _{i:r(i,j)=1}(w^{(j)}\cdot x^{(i)}-b^{(j)}-y^{(i,j)}{)}^2+\frac{\lambda }{2}\sum _{j=1}^{n_u}\sum _{k=1}^n(x_k^{(i)}{)}^2$$

3. collaborative filter

$$J(w,b,x)=\frac{1}{2}\sum _{j=1}^{n_u}\sum _{i:r(i,j)=1}(w^{(j)}\cdot x^{(i)}-b^{(j)}-y^{(i,j)}{)}^2+\frac{\lambda }{2}\sum _{j=1}^{n_u}\sum _{k=1}^n(w_k^{(j)}{)}^2+\sum _{j=1}^{n_u}\sum _{k=1}^n(x_k^{(i)}{)}^2$$

4. for binary labels

$$\begin{gathered} g(x)=\frac{1}{1+e^{-z}} \\ f(x)=g(w^{(j)}\cdot x^{(i)}-b^{(j))}) \\ L=-y^{(i,j)}log(f(x))-(1-y^{(i,j)})log(1-f(x)) \\ j(w,b,x)=\sum L \end{gathered}$$

2 Content-based filter

1. definition

recommand items based on fearures of user and item to find good match

2. cost function
$$J=\sum _{(i,j):r(i,j)=1}(v_u^{(j)}{v}_m^{(j)}-y^{(i,j)}{)}^2+NN(regularization-term)$$