[Udemy] Recommender Systems and Deep Learning in Python

1. Welcome

主要讲四部分内容：

non-personized systems

　　popularity: 基于流行度或者最大利益化的推荐。缺点也明显：你可能在特殊地方有些特殊需求，或者你本来就是大多数人不一样

　　Association: 找出订单里一起下单的物品的相关性，一般有Aproiri, FP 等算法

collaborative filtering

matrix factorization (and its variant like probablistic matrix factorization), also known as SVD

Deep learning

2. Simple recommentation systems

基于popularity 的推荐要考虑时效性，比如一则新闻虽然曾经是爆炸性的阅读量很多，但是不合适出现新闻的推荐中，这就需要在popularity 和 age(时间老化) 之间做平衡.

具体地，Hacker News 网站用的公式为：也叫 rank formula

另一个具体的例子是Reddit:

如果平均值一样，那么需要考虑rating 个数，可以参考下面公式：

google 的 PageRank 算法是基于 Markov 模型的. Markov 模型就是NLP里面的unigram, bigram 的概念，基于前面的条件算出后面结果出现的概率.

怎么评估 Rank 结果

3. Collaborative Filtering

user-user CF - based on user-user similarity

item-item CF - based on item-item similarity

参考资料[2] 里面有具体的代码，不过不是矩阵实现，最好看我翻译的另一篇文章 Comprehensive Guide to build a Recommendation Engine from scratch (in Python) / 从0开始搭建推荐系统，这里有矩阵实现, 更快而且我觉得更明白.

4. Matrix Factorization & Deep Leanring

4.1 Matrix Factorization

1. 先来个最basic 版的matrix factorization 的公式，就是把矩阵X分解成 X=WU. metric 用 Sum Squere Error.

2. 再考虑 bias, user bias 和 item bias

3. formulas: with regularization

4. MF with SVD, 确定是不能有missing data (MF 和SVD 很相似，MF有两个矩阵U和V, SVD 有3个 U,S,V)

5. Probabilistic Matrix Factorization, 没看懂，只能拷些图片在这里了

6. Bayesian Matrix Factorization, 也没看懂，讲者说是optional 就不看了

7. matrix factorization in Keres

# keras model
u = Input(shape=(1,))
m = Input(shape=(1,))
u_embedding = Embedding(N, K, embeddings_regularizer=l2(reg))(u) # (N, 1, K)
m_embedding = Embedding(M, K, embeddings_regularizer=l2(reg))(m) # (N, 1, K)

# subsubmodel = Model([u, m], [u_embedding, m_embedding])
# user_ids = df_train.userId.values[0:5]
# movie_ids = df_train.movie_idx.values[0:5]
# print("user_ids.shape", user_ids.shape)
# p = subsubmodel.predict([user_ids, movie_ids])
# print("p[0].shape:", p[0].shape)
# print("p[1].shape:", p[1].shape)
# exit()


u_bias = Embedding(N, 1, embeddings_regularizer=l2(reg))(u) # (N, 1, 1)
m_bias = Embedding(M, 1, embeddings_regularizer=l2(reg))(m) # (N, 1, 1)
x = Dot(axes=2)([u_embedding, m_embedding]) # (N, 1, 1)

# submodel = Model([u, m], x)
# user_ids = df_train.userId.values[0:5]
# movie_ids = df_train.movie_idx.values[0:5]
# p = submodel.predict([user_ids, movie_ids])
# print("p.shape:", p.shape)
# exit()


x = Add()([x, u_bias, m_bias])
x = Flatten()(x) # (N, 1)

model = Model(inputs=[u, m], outputs=x)
model.compile(
loss='mse',
# optimizer='adam',
# optimizer=Adam(lr=0.01),
optimizer=SGD(lr=0.08, momentum=0.9),
metrics=['mse'],
)

4. 2 Deep Learning

1. 在matrix factorization 上加了deep network, 形成的 deep learning. 前面讲的 matrix factorization 只是linear model, deep learning 会提供 non-linear 能力，所以理论上比只有前面棕色部分（matrix factorization部分）好