本周学习汇报
1.推荐系统实战 Tensorflow实现双塔DNN排序模型
2.论文阅读《Neural Collaborative Filtering》
3.论文《Neural Collaborative Filtering》代码理解
一、 Tensorflow实现双塔DNN排序模型
1.数据集选取 采用的是Movielens数据集ml-10m
数据集从官网下载后,导入数据集:
# 导入数据集
df_user = pd.read_csv("users.dat", sep="::", header=None, engine="python",
names="UserID::Gender::Age::Occupation::Zip-code".split("::"))
df_movies = pd.read_csv("movies.dat", sep="::", header=None, engine="python",
names="MovieID::Title::Genres".split("::"))
df_ratings = pd.read_csv("ratings.dat", sep="::", header=None, engine="python",
names="UserID::MovieID::Rating::Timestamp".split("::"))
# 计算电影中每个题材的数目
genre_count = collections.defaultdict(int)
for genres in df_movies["Genres"].str.split("|"):
for genre in genres:
genre_count[genre] += 1
print(genre_count, end="/n")
# 只保留具有代表性的体裁
def get_highrate_genre(x):
sub_values = {}
for genre in x.split("|"):
sub_values[genre] = genre_count[genre]
return sorted(sub_values.items(), key=lambda x:x[1], reverse=True)[0][0]
df_movies["Genres"] = df_movies["Genres"].map(get_highrate_genre)
df_movies.sample(frac=1).head(3)
print(df_movies.sample(frac=1).head(3))
# 给列新增数字索引 目的是防止embedding过大
def add_index_column(param_df, column_name):
values = list(param_df[column_name].unique())
value_index_dict = {value: idx for idx, value in enumerate(values)}
param_df[f"{column_name}_idx"]=param_df[column_name].map(value_index_dict)
add_index_column(df_user, "UserID")
add_index_column(df_user, "Gender")
add_index_column(df_user, "Age")
add_index_column(df_user, "Occupation")
add_index_column(df_movies, "MovieID")
add_index_column(df_movies, "Genres")
print(df_user.head())
print(df_movies.head())
# 合并成一个df
df = pd.merge(pd.merge(df_ratings, df_user), df_movies)
df.drop(columns=["Timestamp", "Zip-code", "Title"], inplace=True)
print(df.sample(frac=1).head(3))
# 算出词表的大小
num_users = df["UserID_idx"].max()+1
num_movies = df["MovieID_idx"].max()+1
num_genders = df["Gender_idx"].max()+1
num_ages = df["Age_idx"].max()+1
num_occupations = df["Occupation_idx"].max()+1
num_genres = df["Genres_idx"].max()+1
print(num_users, num_movies, num_genders, num_ages, num_occupations, num_genres)
# 评分的归一化
min_rating = df["Rating"].min()
max_rating = df["Rating"].max()
df["Rating"] = df["Rating"].map(lambda x:(x-min_rating)/(max_rating-min_rating))
print(df.sample(frac=1).head(3))
# 构建训练数据集 为了快捷只采用了采样10%的数据 输出是rating
df_sample = df.sample(frac=0.1)
X = df_sample[["UserID_idx", "Gender_idx", "Age_idx", "Occupation_idx", "MovieID_idx", "Genres_idx"]]
y = df_sample.pop("Rating")
# 搭建双塔模型并训练
def get_model():
# 函数式API搭建双塔DNN模型
# 输入
user_id = keras.layers.Input(shape=(1,), name="user_id")
gender = keras.layers.Input(shape=(1,), name="gender")
age = keras.layers.Input(shape=(1,), name="age")
occupation = keras.layers.Input(shape=(1,), name="occupation")
movie_id = keras.layers.Input(shape=(1,), name="movie_id")
genre1 = keras.layers.Input(shape=(1,), name="genre1")
# user塔
user_vector = tf.keras.layers.concatenate([
layers.Embedding(num_users, 100)(user_id),
layers.Embedding(num_genders, 2)(gender),
layers.Embedding(num_ages, 2)(age),
layers.Embedding(num_occupations, 2)(occupation)
])
# 全连接
user_vector = layers.Dense(32, activation="relu")(user_vector)
user_vector = layers.Dense(8, activation='relu',
name="user_embedding", kernel_regularizer='12')(user_vector)
# movie塔
movie_vector = tf.keras.layers.concatenate([
layers.Embedding(num_movies, 100)(movie_id),
layers.Embedding(num_genres, 2)(genre1)
])
# 全连接
movie_vector = layers.Dense(32, activation='relu')(movie_vector)
movie_vector = layers.Dense(8, activation='relu',
name="movie_embedding", kernel_regularizer='12')(movie_vector)
# 每个用户的embedding和item的embedding作点积
dot_user_movie = tf.reduce_sum(user_vector*movie_vector, axis=1)
dot_user_movie = tf.expand_dims(dot_user_movie, 1)
output = layers.Dense(1, activation='sigmoid')(dot_user_movie)
return keras.models.Model(inputs=[user_id, gender, age, occupation, movie_id, genre1], outputs=[output])
model = get_model()
model.compile(loss=tf.keras.losses.MeanSquaredError(),
optimizer=keras.optimizers.RMSprop())
- 论文阅读《Neural Collaborative Filtering》
论文贡献:
1、提出了一种神经网络结构来模拟用户和项目的潜在特征,并设计了基于神经网络的协同过滤的通用框架NCF。
2、表明MF可以被解释为NCF的特例,并利用多层感知器来赋予NCF高水平的非线性建模能力。
3、对两个真实数据集进行广泛的实验,以证明NCF方法的有效性和对使用深度学习进行协作过滤的承诺。
代码是使用keras来实现的深度学习,其中GMF.py是传统的Matrix Factorization算法,关键代码分为两部分:
def get_model(num_users, num_items, latent_dim, regs=[0,0]):
# Input variables
user_input = Input(shape=(1,), dtype='int32', name = 'user_input')
item_input = Input(shape=(1,), dtype='int32', name = 'item_input')
MF_Embedding_User = Embedding(input_dim = num_users, output_dim = latent_dim, name = 'user_embedding',
init = init_normal, W_regularizer = l2(regs[0]), input_length=1)
MF_Embedding_Item = Embedding(input_dim = num_items, output_dim = latent_dim, name = 'item_embedding',
init = init_normal, W_regularizer = l2(regs[1]), input_length=1)
# Crucial to flatten an embedding vector!
user_latent = Flatten()(MF_Embedding_User(user_input))
item_latent = Flatten()(MF_Embedding_Item(item_input))
# Element-wise product of user and item embeddings
predict_vector = merge([user_latent, item_latent], mode = 'mul')
# Final prediction layer
#prediction = Lambda(lambda x: K.sigmoid(K.sum(x)), output_shape=(1,))(predict_vector)
prediction = Dense(1, activation='sigmoid', init='lecun_uniform', name = 'prediction')(predict_vector)
model = Model(input=[user_input, item_input],
output=prediction)
return model
上述代码是构建模型结构,首先定义Input为一维多列的数据,然后是Embedding层,Embedding主要是为了降维,就是起到了look up的作用,然后是Merge层,将用户和物品的张量进行了内积相乘(latent_dim 表示两者的潜在降维的维度是相同的,因此可以做内积),紧接着是一个全连接层,激活函数为sigmoid。
下面是MLP.py的源码:
def get_model(num_users, num_items, layers = [20,10], reg_layers=[0,0]):
assert len(layers) == len(reg_layers)
num_layer = len(layers) #Number of layers in the MLP
# Input variables
user_input = Input(shape=(1,), dtype='int32', name = 'user_input')
item_input = Input(shape=(1,), dtype='int32', name = 'item_input')
MLP_Embedding_User = Embedding(input_dim = num_users, output_dim = layers[0]/2, name = 'user_embedding',
init = init_normal, W_regularizer = l2(reg_layers[0]), input_length=1)
MLP_Embedding_Item = Embedding(input_dim = num_items, output_dim = layers[0]/2, name = 'item_embedding',
init = init_normal, W_regularizer = l2(reg_layers[0]), input_length=1)
# Crucial to flatten an embedding vector!
user_latent = Flatten()(MLP_Embedding_User(user_input))
item_latent = Flatten()(MLP_Embedding_Item(item_input))
# The 0-th layer is the concatenation of embedding layers
vector = merge([user_latent, item_latent], mode = 'concat')
# MLP layers
for idx in xrange(1, num_layer):
layer = Dense(layers[idx], W_regularizer= l2(reg_layers[idx]), activation='relu', name = 'layer%d' %idx)
vector = layer(vector)
# Final prediction layer
prediction = Dense(1, activation='sigmoid', init='lecun_uniform', name = 'prediction')(vector)
model = Model(input=[user_input, item_input],
output=prediction)
return model
最重要的也是构建模型的部分,与GMF不同的有两个部分,首先是user_latent和item_latent的merge的部分,不再采用内积的形式,而是contract拼接的方式;再者就是for循环构建深层全连接神经网络,内部Layer的激活函数是relu,最后一层的激活函数仍然是sigmoid。
接下来是NeuMF.py,将MLP和GMF进行了融合,模型构建代码如下
def get_model(num_users, num_items, mf_dim=10, layers=[10], reg_layers=[0], reg_mf=0):
assert len(layers) == len(reg_layers)
num_layer = len(layers) #Number of layers in the MLP
# Input variables
user_input = Input(shape=(1,), dtype='int32', name = 'user_input')
item_input = Input(shape=(1,), dtype='int32', name = 'item_input')
# Embedding layer
MF_Embedding_User = Embedding(input_dim = num_users, output_dim = mf_dim, name = 'mf_embedding_user',
init = init_normal, W_regularizer = l2(reg_mf), input_length=1)
MF_Embedding_Item = Embedding(input_dim = num_items, output_dim = mf_dim, name = 'mf_embedding_item',
init = init_normal, W_regularizer = l2(reg_mf), input_length=1)
MLP_Embedding_User = Embedding(input_dim = num_users, output_dim = layers[0]/2, name = "mlp_embedding_user",
init = init_normal, W_regularizer = l2(reg_layers[0]), input_length=1)
MLP_Embedding_Item = Embedding(input_dim = num_items, output_dim = layers[0]/2, name = 'mlp_embedding_item',
init = init_normal, W_regularizer = l2(reg_layers[0]), input_length=1)
# MF part
mf_user_latent = Flatten()(MF_Embedding_User(user_input))
mf_item_latent = Flatten()(MF_Embedding_Item(item_input))
mf_vector = merge([mf_user_latent, mf_item_latent], mode = 'mul') # element-wise multiply
# MLP part
mlp_user_latent = Flatten()(MLP_Embedding_User(user_input))
mlp_item_latent = Flatten()(MLP_Embedding_Item(item_input))
mlp_vector = merge([mlp_user_latent, mlp_item_latent], mode = 'concat')
for idx in xrange(1, num_layer):
layer = Dense(layers[idx], W_regularizer= l2(reg_layers[idx]), activation='relu', name="layer%d" %idx)
mlp_vector = layer(mlp_vector)
# Concatenate MF and MLP parts
#mf_vector = Lambda(lambda x: x * alpha)(mf_vector)
#mlp_vector = Lambda(lambda x : x * (1-alpha))(mlp_vector)
predict_vector = merge([mf_vector, mlp_vector], mode = 'concat')
# Final prediction layer
prediction = Dense(1, activation='sigmoid', init='lecun_uniform', name = "prediction")(predict_vector)
model = Model(input=[user_input, item_input],
output=prediction)
return model
代码的前半部分分别是GMF和MLP的内部layer构建过程,在 predict_vector = merge([mf_vector, mlp_vector], mode = 'concat')这一行开始对两者的输出进行了merge,方式为concat。最后包了一层的sigmoid。
看完了构建模型的代码,有几个细节值得关注:
- 训练样本的正负比例如何设定?
def get_train_instances(train, num_negatives):
user_input, item_input, labels = [],[],[]
num_users = train.shape[0]
for (u, i) in train.keys():
# positive instance user_input.append(u)
item_input.append(i)
labels.append(1)
# negative instances
for t in xrange(num_negatives):
j = np.random.randint(num_items)
while train.has_key((u, j)):
j = np.random.randint(num_items)
user_input.append(u)
item_input.append(j)
labels.append(0)
return user_input, item_input, labels
该函数是获取用户和物品的训练数据,其中num_negatives控制着正负样本的比例,负样本的获取方法也简单粗暴,直接随机选取用户没有选择的其余的物品。
2.保存了训练的模型,该怎么对数据进行预测?
从evalute.py中的源码中可以得到答案
def eval_one_rating(idx):
rating = _testRatings[idx]
items = _testNegatives[idx]
u = rating[0]
gtItem = rating[1]
items.append(gtItem)
# Get prediction scores
map_item_score = {}
users = np.full(len(items), u, dtype = 'int32')
predictions = _model.predict([users, np.array(items)],
batch_size=100, verbose=0)
for i in xrange(len(items)):
item = items[i]
map_item_score[item] = predictions[i]
items.pop()
# Evaluate top rank list
ranklist = heapq.nlargest(_K, map_item_score, key=map_item_score.get)
hr = getHitRatio(ranklist, gtItem)
ndcg = getNDCG(ranklist, gtItem)
return (hr, ndcg)
输入只要保证和训练的时候的格式一样即可,这里作者事先构建了negative的数据,也就是说对negative的物品和测试集合中的某一个物品进行了预测,最终选取topK的,来评测是否在其中(注getHitRatio函数不是最终结果,只是0/1) eval_one_rating 函数只是对测试集合中的某个用户的某个物品,以及和事先划分好的负样本组合在一起进行预测,最终输出该测试物品是否在topK中。
3.Embedding 层的物品的latent_dim和用户的latent_dim是一致的,如果不一致是否可以?
在实际中未必两者的维度是一致的,这里受限于keras的merge函数的参数要求,输入的数据的shape必须是一致的,所以必须是一致的。以及Merge中的mode参数,至于什么时候选择contact,什么时候选择mul,我觉得依赖于模型效果,在实际工程中选择使得最优的方式。
4 python MLP.py --dataset ml-1m --epochs 20 --batch_size 256 --layers [64,32,16,8] 理解
这是运行MLP的参数,layers的参数在逐渐减小,这也是深度神经网络的潜在设置,一般意义上越深的layer是对前面的更高层次的抽象。