笔记：协同过滤

推荐算法主要分为3类：基于内容的推荐，协同过滤和混合推荐。
协同过滤主要利用user和item之间的互动信息。分为三种类型：user-based， item-based和model-based。

数据集

根据数据集的性质，可以分为显式反馈（评分，星级）和隐式反馈（浏览，点击，加入购物车）。
movielens数据集中，用户对电影的打分在1-5，是显式反馈。movielens-100k中有943个用户和1682个电影的交互信息。

import pandas as pd
import numpy as np
filename = "../Data/ml-100k/u.data"
separator = "\t"
col_names = ['user_id', 'item_id', 'rating', 'timestamp']
df = pd.read_csv(filename, sep=separator, names=col_names)
n_users = df.user_id.unique().shape[0]
n_items = df.item_id.unique().shape[0]
print(str(n_users) + ' users') #943 users
print(str(n_items) + ' items') #1682 items

将其转为user-item矩阵。注意虽然把缺失的评分项置为0，但不代表真实的评分为0，只是空项。接着需要将数据划分为训练集和测试集，对于每个用户，从用户的所有评分中选出10个作为测试集。

ratings = np.zeros((n_users, n_items))
for row in df.itertuples():
    ratings[row[1]-1, row[2]-1] = row[3]
sparsity = float(len(ratings.nonzero()[0]))
sparsity /= (ratings.shape[0] * ratings.shape[1])
sparsity *= 100
print('Sparsity: {:4.2f}%'.format(sparsity)) 
# Sparsity: 6.30%

def train_test_split(ratings):
    test = np.zeros(ratings.shape)
    train = ratings.copy()
    for user in range(ratings.shape[0]):
        test_ratings = np.random.choice(ratings[user, :].nonzero()[0],
                                        size=10,
                                        replace=False)
        train[user, test_ratings] = 0.
        test[user, test_ratings] = ratings[user, test_ratings]

    # Test and training are truly disjoint
    assert (np.all((train * test) == 0))
    return train, test
train, test = train_test_split(ratings)

user-based & item-based

在user-based和item-based协同过滤中，需要建立一个相似矩阵。用户相似矩阵中是每对用户间的相似度，项相似矩阵中是每对项的相似度。一种常用的计算相似度的方法是cosine。

def fast_similarity(ratings, kind='user', epsilon=1e-9):
    # epsilon -> small number for handling dived-by-zero errors
    if kind == 'user':
        sim = ratings.dot(ratings.T) + epsilon
    elif kind == 'item':
        sim = ratings.T.dot(ratings) + epsilon
    norms = np.array([np.sqrt(np.diagonal(sim))])
    return (sim / norms / norms.T)
user_similarity = fast_similarity(train, kind='user')
item_similarity = fast_similarity(train, kind='item')

得到相似度矩阵后，可以进行预测。对于user-based推荐，用户对项的评分被预测为其他用户对该项评分的加权和，权值为其他用户与该用户的相似度，用其他用户与该用户的相似度的和进行规范。

def predict_fast_simple(ratings, similarity, kind='user'):
    if kind == 'user':
        return similarity.dot(ratings) / np.array([np.abs(similarity).sum(axis=1)]).T
    elif kind == 'item':
        return ratings.dot(similarity) / np.array([np.abs(similarity).sum(axis=1)])

使用mean squared error对模型进行评价。mse的计算公式如下：

from sklearn.metrics import mean_squared_error
def get_mse(pred, actual):
    # Ignore nonzero terms.
    pred = pred[actual.nonzero()].flatten()
    actual = actual[actual.nonzero()].flatten()
    return mean_squared_error(pred, actual)
item_prediction = predict_fast_simple(train, item_similarity, kind='item')
user_prediction = predict_fast_simple(train, user_similarity, kind='user')
user_mse = get_mse(user_prediction, test)
item_mse = get_mse(item_prediction, test)
print('User-based CF MSE: ' + str(user_mse))
print('Item-based CF MSE: ' + str(item_mse))

top-k推荐

与上一种方法不同的是，top-k推荐只考虑与输入用户最相似的k个其他用户。调整参数k的取值获得最优解。

def predict_topk(ratings, similarity, kind='user', k=40):
    pred = np.zeros(ratings.shape)
    if kind == 'user':
        for i in range(ratings.shape[0]):
            top_k_users = [np.argsort(similarity[:,i])[:-k-1:-1]]
            for j in range(ratings.shape[1]):
                pred[i, j] = similarity[i, :][top_k_users].dot(ratings[:, j][top_k_users]) 
                pred[i, j] /= np.sum(np.abs(similarity[i, :][top_k_users]))
    if kind == 'item':
        for j in range(ratings.shape[1]):
            top_k_items = [np.argsort(similarity[:,j])[:-k-1:-1]]
            for i in range(ratings.shape[0]):
                pred[i, j] = similarity[j, :][top_k_items].dot(ratings[i, :][top_k_items].T) 
                pred[i, j] /= np.sum(np.abs(similarity[j, :][top_k_items]))        

    return pred

pred = predict_topk(train, user_similarity, kind='user', k=40)
print('Top-k User-based CF MSE: ' + str(get_mse(pred, test)))

pred = predict_topk(train, item_similarity, kind='item', k=40)
print('Top-k Item-based CF MSE: ' + str(get_mse(pred, test)))
# Top-k User-based CF MSE: 6.47059807493
# Top-k Item-based CF MSE: 7.75559095568

bias-subtracted推荐

该方法的思想是：一些用户对电影进行评分时倾向于打高分或低分。因此在对相似用户的评分进行加和时，会减去每个用户的平均评分，最后再加回平均评分。

def predict_nobias(ratings, similarity, kind='user'):
    if kind == 'user':
        user_bias = ratings.mean(axis=1)
        ratings = (ratings - user_bias[:, np.newaxis]).copy()
        pred = similarity.dot(ratings) / np.array([np.abs(similarity).sum(axis=1)]).T
        pred += user_bias[:, np.newaxis]
    elif kind == 'item':
        item_bias = ratings.mean(axis=0)
        ratings = (ratings - item_bias[np.newaxis, :]).copy()
        pred = ratings.dot(similarity) / np.array([np.abs(similarity).sum(axis=1)])
        pred += item_bias[np.newaxis, :]  
    return pred
user_pred = predict_nobias(train, user_similarity, kind='user')
print('Bias-subtracted User-based CF MSE: ' + str(get_mse(user_pred, test)))

item_pred = predict_nobias(train, item_similarity, kind='item')
print('Bias-subtracted Item-based CF MSE: ' + str(get_mse(item_pred, test)))
# Bias-subtracted User-based CF MSE: 8.67647634245
# Bias-subtracted Item-based CF MSE: 9.71148412222

matrix factorization

传统的奇异值分解要求矩阵不能有缺失数据，必须是稠密的。而用户物品评分矩阵是一个很典型的稀疏矩阵，直接使用传统的SVD进行协同过滤是比较复杂的。
推荐中的矩阵分解算法有：FunkSVD，BiasSVD，SVD++。这些算法和传统SVD的最大区别是不再将矩阵分解为$U \Sigma V^T$，而是两个矩阵$P^TQ$。

ALS SGD

。。。 。。。

参考文献：
http://www.cnblogs.com/pinard/p/6349233.html
http://blog.ethanrosenthal.com/2015/11/02/intro-to-collaborative-filtering/