小试牛刀——协同滤波（SVD、Item-based）代码实现

参考论文：Advances in Collaborative Filtering
参考资料：http://blog.csdn.net/dark_scope/article/details/17228643

---

数据集：movielens100k
这里我只选用了u1.base, u1.test 两个部分分别作为测试集和训练集。

Baseline Predictor

$b_{ui} \rightarrow$ A baseline prediction for an unknown rating.
$b_{u} \rightarrow$ the observed deviations of user u from the average.
$b_{i} \rightarrow$ the observed deviations of item i from the average.
$\mu \rightarrow$ the overall average rating.

我们可以通过优化下面的目标函数来估计 $b_u$ 和 $b_i$ 的值：

有一个更加简便但是会损失精确度的方法可以用来估计 $b_u$ 和 $b_i$ 的值：

评估结果

我们使用RMSE（Root Mean Squared Error）的值来评估结果的好坏：

|TestSet| ：表示测试集的数目。

SVD

概念总结

用 $q_i^T p_u$ 可以用来用户 $u$ 与项目 $i$ 之间的交互，那么预测的评分为：

优化的目标函数（$b_i、b_u、q_i、p_u$）：

采用随机梯度下降进行优化：

算法实现

from __future__ import division
import numpy as np
import scipy as sp
from numpy.random import random
class SVD_CF:
    def __init__(self,train_X,k=20):
        self.train_X=np.array(train_X)
        self.k=k#size of latent vector
        self.num_train=self.train_X.shape[0]#size of train dataset
        self.bu={}#deviations of user u
        self.bi={}#deviations of item i
        self.mu=np.mean(self.train_X[:,2])
        self.qi={}#latent vector of item
        self.pu={}#latent vector of user
        self.movie_user={}
        self.user_movie={}
        for i in range(self.num_train):
            uid=self.train_X[i][0]
            mid=self.train_X[i][1]
            rat=self.train_X[i][2]
            self.movie_user.setdefault(mid,{})
            self.movie_user[mid][uid]=rat
            self.user_movie.setdefault(uid,{})
            self.user_movie[uid][mid]=rat
            self.qi.setdefault(mid, random((self.k, 1)) / 10 * (np.sqrt(self.k)))
            self.pu.setdefault(uid, random((self.k, 1)) / 10 * (np.sqrt(self.k)))#latent vector初始化为一串随机数
    def pred(self,uid,mid):
        self.bi.setdefault(mid, 0)
        self.bu.setdefault(uid, 0)
        self.qi.setdefault(mid, np.zeros((self.k, 1)))#对于训练集中没有的item，latent vector 设为0
        self.pu.setdefault(uid, np.zeros((self.k, 1)))#对于训练集中没有的user，latent vector 设为0
        # if (self.qi[mid] == None):
        #     self.qi[mid] = np.zeros((self.k, 1))
        # if (self.pu[uid] == None):
        #     self.pu[uid] = np.zeros((self.k, 1))
        ans = self.mu + self.bi[mid] + self.bu[uid] + np.sum(self.qi[mid] * self.pu[uid])
        if ans > 5:#大于5的评分输出5
            return 5
        elif ans < 1:
            return 1#小于1的评分输出1
        return ans

    def train(self,steps=20,gamma=0.04,lda=0.15):
        print("Train on SVD:")
        for step in range(steps):
            print('the ',step,'-th  step is running')
            rmse_sum=0.0
            kk = np.random.permutation(self.num_train)#随机打乱顺序
            for j in range(self.num_train):
                i=kk[j]
                uid = self.train_X[i][0]
                mid = self.train_X[i][1]
                rat = self.train_X[i][2]
                rui = self.pred(uid,mid)
                eui = rat - rui
                rmse_sum+=eui**2
                self.bu[uid]+=gamma*(eui-lda*self.bu[uid])
                self.bi[mid]+=gamma*(eui-lda*self.bi[mid])
                self.qi[mid]+=gamma*(eui*self.pu[uid]-lda*self.qi[mid])
                self.pu[uid]+=gamma*(eui*self.qi[mid]-lda*self.pu[uid])
            gamma=gamma*0.93
            rmse=np.sqrt(rmse_sum/self.num_train)
            print("the rmse of this step on train data is ",rmse)
    def test(self,test_X):
        test_X = np.array(test_X)
        output = []
        sums = 0
        print("Test on SVD:")
        print("the test data size is ", test_X.shape)
        for i in range(test_X.shape[0]):
            pre = self.pred(test_X[i][0], test_X[i][1])
            output.append(pre)
            sums += (pre - test_X[i][2]) ** 2
        rmse = np.sqrt(sums / test_X.shape[0])
        print("the rmse on test data is ", rmse)
        return output

提示：在测试的时候由于忘记 import scipy as sp 一直报错

---

Item-based

概念总结

项目 $i$ 和项目 $j$ 之间的相似性用皮尔森相关系数（Pearson correlation coefficient）$ρ_{ij }$来表示：

$b_{ui}$ 为均值，$r_{ui}$ 为预测值。数据集$U(i,j)$ 包含所有同时给项目$i$ 和项目$j$打分的用户。

根据上面的相似性计算公式，我们在同一个用户 $u$ 评分的情况下为项目 $i$选出 $k$ 个与其最为相似的项目，作为它的邻居，表示为 $S^k (i;u)$ 。预测的评分值可以看作是与其相似的 k <script type="math/tex" id="MathJax-Element-25">k</script> 个项目评分的加权平均值。（在下面的实现中，我们并没有选取k个邻居，而是使用所有的）

算法实现

item_based.py 实现算法：

from __future__ import division
import numpy as np
import scipy as sp

class Item_based_CF:
    def __init__(self,train_X):
        self.train_X=np.array(train_X)
        self.movie_user={}#用来找U(i,j)
        self.user_movie={}#用来找S^k(i;u)
        self.ave = np.mean(self.train_X[:, 2])#均值
        for i in range(self.train_X.shape[0]):
            uid=self.train_X[i][0]
            mid=self.train_X[i][1]
            rat=self.train_X[i][2]
            self.movie_user.setdefault(mid,{})#dict.setdefault(key, default=None)default -- 键不存在时，设置的默认键值。
            self.movie_user[mid][uid]=rat
            self.user_movie.setdefault(uid, {})
            self.user_movie[uid][mid] = rat
        self.movie_sim={}
    def sim_cal(self,m1,m2):
        self.movie_sim.setdefault(m1, {})
        self.movie_sim.setdefault(m2, {})
        self.movie_user.setdefault(m1, {})#可能出现训练集中没有打过分的电影
        self.movie_user.setdefault(m2, {})
        self.movie_sim[m1].setdefault(m2, -1)
        self.movie_sim[m2].setdefault(m1, -1)

        if self.movie_sim[m1][m2] != -1:#直接输出已经计算过的sim
            return self.movie_sim[m1][m2]
        si = {}#存储U(i,j)
        for user in self.movie_user[m1]:
            if user in self.movie_user[m2]:
                si[user] = 1
        n = len(si)
        if (n == 0):#U(i,j)为空
            self.movie_sim[m1][m2] = 1
            self.movie_sim[m2][m1] = 1
            return 1
        #求皮尔森系数
        s1 = np.array([self.movie_user[m1][u] for u in si])
        s2 = np.array([self.movie_user[m2][u] for u in si])
        sum1 = np.sum(s1)
        sum2 = np.sum(s2)
        sum1Sq = np.sum(s1 ** 2)
        sum2Sq = np.sum(s2 ** 2)
        pSum = np.sum(s1 * s2)
        num = pSum - (sum1 * sum2 / n)
        den = np.sqrt((sum1Sq - sum1 ** 2 / n) * (sum2Sq - sum2 ** 2 / n))
        if den == 0:#只有当den不等于0才有意义
            self.movie_sim[m1][m2] = 0
            self.movie_sim[m2][m1] = 0
            return 0
        self.movie_sim[m1][m2] = num / den
        self.movie_sim[m2][m1] = num / den
        return num / den
    def pred(self,uid,mid):
        sim_accumulate = 0.0
        rat_acc = 0.0
        for item in self.user_movie[uid]:
            sim = self.sim_cal(item, mid)
            if sim < 0: continue
            rat_acc += sim * self.user_movie[uid][item]
            sim_accumulate += sim
        if sim_accumulate == 0:  # no same user rated,return average rates of the data
            return self.ave
        return rat_acc / sim_accumulate

    def test(self, test_X):
        test_X = np.array(test_X)
        output = []
        sums = 0
        print("Test on Item-Based:")
        print("the test data size is ", test_X.shape)
        for i in range(test_X.shape[0]):
            pre = self.pred(test_X[i][0], test_X[i][1])
            output.append(pre)
            sums += (pre - test_X[i][2]) ** 2
        rmse = np.sqrt(sums / test_X.shape[0])
        print("the rmse on test data is ", rmse)
        return output

test.py 实现读取数据和测试：

import numpy as np
from item_based import Item_based_CF
from svd_mf import SVD_CF
def data_read():
    test = 'D:/MachineLearning/Notes/Recommendation/Projects/test01_CF/data/ml-100k/u1.test'
    train = 'D:/MachineLearning/Notes/Recommendation/Projects/test01_CF/data/ml-100k/u1.base'
    train_X = []
    test_X = []
    with open(test, 'r') as f:
        lines = f.readlines()
        for each in lines:
            p = each.split()
            new = [int(p[0]), int(p[1]), int(p[2])]
            test_X.append(new)

    with open(train, 'r') as f:
        lines = f.readlines()
        for each in lines:
            p = each.split()
            new = [int(p[0]), int(p[1]), int(p[2])]
            train_X.append(new)

    return train_X,test_X
#
train_X,test_X=data_read()
a=Item_based_CF(train_X)
a.test(test_X)
b=SVD_CF(train_X,30)
b.train()
b.test(test_X)

OUTPUT：Test on Item-Based:
the test data size is (20000, 3)
the rmse on test data is 1.04164917355
Train on SVD:
the 0 -th step is running
the rmse of this step on train data is 1.06147863504
the 1 -th step is running
the rmse of this step on train data is 0.933068598689
the 2 -th step is running
the rmse of this step on train data is 0.917020530118
the 3 -th step is running
the rmse of this step on train data is 0.908024566318
the 4 -th step is running
the rmse of this step on train data is 0.899151009875
the 5 -th step is running
the rmse of this step on train data is 0.891401965185
the 6 -th step is running
the rmse of this step on train data is 0.884340062176
the 7 -th step is running
the rmse of this step on train data is 0.877877195046
the 8 -th step is running
the rmse of this step on train data is 0.872217832148
the 9 -th step is running
the rmse of this step on train data is 0.867243365719
the 10 -th step is running
the rmse of this step on train data is 0.862927174872
the 11 -th step is running
the rmse of this step on train data is 0.859048442347
the 12 -th step is running
the rmse of this step on train data is 0.85523094384
the 13 -th step is running
the rmse of this step on train data is 0.852608188393
the 14 -th step is running
the rmse of this step on train data is 0.849679360252
the 15 -th step is running
the rmse of this step on train data is 0.847087016118
the 16 -th step is running
the rmse of this step on train data is 0.84498687518
the 17 -th step is running
the rmse of this step on train data is 0.842603466917
the 18 -th step is running
the rmse of this step on train data is 0.840825090125
the 19 -th step is running
the rmse of this step on train data is 0.838784925391
Test on SVD:
the test data size is (20000, 3)
the rmse on test data is 0.929369246698