FM: Factorization Machines解读

文献：

《Factorization Machines》，没法上传文献

原理解读：

FM主要是解决稀疏数据下的特征组合问题

普通的线性模型LR，我们都是将各个特征独立考虑的，并没有考虑到特征与特征之间的相互关系：

但实际上，特征之间可能具有一定的关联。以新闻推荐为例，一般男性用户看军事新闻多，而女性用户喜欢情感类新闻，那么可以看出性别与新闻的频道有一定的关联性，如果能找出这类的特征，是非常有意义的：

神经网络中的实现

出现了很多将fm与神经网络结合的模型，如deepfm、nfm等。下面讲解一下如何在代码中实现。
先讲一下难点在于二阶特征的形成，化简为和之积与积之和相减。这是推导过程：

deepfm的代码在keras中实现：
分为一阶、二阶、deep部分。
其中continuous_tensors代表连续型变量
的组合，embeddings_tensors代表离散型变量做embedding后的组合。

global deep_cols
    global embedding_cols
    global cont_cols
    with open("./data/embedding_number_v4.json",'r', encoding='UTF-8') as f:
        load_dict = json.load(f)
    reg = 1e-4
    v_factors=4
    unique_vals=load_dict
    embeddings_first_order_tensors=[]
    embeddings_second_order_tensors=[]
    input_layers=[]
    for ec in embedding_cols:
        layer_name = ec 
        vocabulary_size=unique_vals[ec]
        inp = Input(shape=(1,), dtype='int64', name=layer_name)
        input_layers.append(inp)
        # 1 order
        embeddings_first_order_tensor=Embedding(vocabulary_size, 1, input_length=1)(inp)
        embeddings_first_order_tensors.append(embeddings_first_order_tensor)
        # 2order
        embeddings_second_order_tensor=Embedding(vocabulary_size,v_factors,input_length=1)(inp)
        embeddings_second_order_tensors.append(embeddings_second_order_tensor)
        del embeddings_second_order_tensor

    continuous_first_order_tensors = []
    continuous_second_order_tensors=[]
    for cc in cont_cols:
        layer_name = cc
        inp = Input(shape=(1,), dtype='int64', name=layer_name)
        input_layers.append(inp)
        inp_2=Reshape((1, 1))(inp)
        inp_3=Lambda(int_to_float)(inp_2)
        # 1 order
        continuous_first_order_tensor=Dense(1)(inp_3)
        continuous_first_order_tensors.append(continuous_first_order_tensor)
        # 2 order
        #continuous_second_order_tensor=RepeatVector(1)(Dense(v_factors)(inp_3))
        continuous_second_order_tensor=Dense(v_factors)(inp_3)
        continuous_second_order_tensors.append(continuous_second_order_tensor)
        del continuous_second_order_tensor

# 1-order
    #first_order=Flatten()(Add()(continuous_first_order_tensors+embeddings_first_order_tensors))# 1
    first_order=concatenate(inputs=continuous_first_order_tensors+embeddings_first_order_tensors,axis=2)
    first_order=Dense(1,kernel_regularizer=l2(0.01))(first_order)
    print("first_order:",first_order)
    # 2-order
    second_order_inp=embeddings_second_order_tensors+continuous_second_order_tensors
    print("second_order",second_order_inp)

    # _________sum square part _____________
    summed_features_emb=Add()(second_order_inp) #input 列表
    print("summed_features_emb:",summed_features_emb)
    summed_features_emb_square=Lambda(lambda x: tf.square(x))(summed_features_emb)
    print("summed_features_emb_square:",summed_features_emb_square)

    # _________ square_sum part _____________
    squared_features_emb=Lambda(lambda x: tf.square(x))(second_order_inp)
    print("squared_features_emb:",squared_features_emb)
    squared_sum_features_emb=Lambda(lambda x: tf.reduce_sum(x,0))(squared_features_emb)
    print("squared_sum_features_emb:",squared_sum_features_emb)

    second_order=Lambda(lambda x: 0.5 * tf.subtract(x[0],x[1]))([summed_features_emb_square, squared_sum_features_emb])
    #second_order=Flatten()(second_order)

    # deep-part
    #d = Flatten()(concatenate(second_order_inp))
    d = concatenate(second_order_inp)
    d = Dense(128, activation='relu',kernel_regularizer=l2(0.01))(d)
    #d=Dropout(0.5)(d)
    d = Dense(64, activation='relu',kernel_regularizer=l2(0.01))(d)
    #d=Dropout(0.5)(d)
    d = concatenate([first_order,second_order,d])
    out = Dense(1, activation='sigmoid')(d)
    print("out:",out)
    out=Reshape([1])(out)
    deep = Model(input_layers, out)
    deep.summary()

具体实现见：https://github.com/zhangyingerjelly/recommendation/tree/master/%E5%8D%95%E7%9B%AE%E6%A0%87%E6%8E%A8%E8%8D%90%E6%A8%A1%E5%9E%8B