Neuralcf neural collaborative filtering network

One 、 summary

CF（ Collaborative filtering ） Combined with deep learning

Review matrix decomposition

Background of the model

The model structure of matrix decomposition is relatively simple , It is obtained by inner product of user hidden vector and item hidden vector “ Similarity degree ”, Here, similarity refers to the prediction of the score . In particular, the output layer cannot effectively fit the optimization objectives , The model is easy to be in the state of under fitting （ Under fitting means that the degree of model fitting is not high ）. Therefore, on this basis NeuralCF Model .

---

Two 、NeuralCF The structure of the original model

NeuralCF use “ Multilayer neural network + Output layer ” Instead of the simple inner product operation in the matrix decomposition model
Advantage is ：
1、 Let the user vector and the item vector intersect more fully , Get more valuable feature combination information
2、 Introduce more nonlinear features , Make the model more expressive

---

3、 ... and 、 Introduction of hybrid model

GMF- Generalized matrix factorization

Here we explain the user hidden vector $P_u=P^T{v}_u^U$ , Item implicit vector isomorphism

class GMF(nn.Module):
    def __init__(self,user_num, item_num, factor_num):
        super(GMF,self).__init__()
        self.embed_user_GMF = nn.Embedding(user_num,factor_num)
        self.embed_item_GMF = nn.Embedding(item_num,factor_num)
        self.predict_layer = nn.Linear(factor_num,1)
        self._init_weight_()

    def _init_weight_(self):
        nn.init.normal_(self.embed_item_GMF.weight,std=0.01)
        nn.init.normal_(self.embed_user_GMF.weight,std=0.01)

    def forward(self,user,item):
        embed_user_GMF = self.embed_user_GMF(user)
        embed_item_GMF = self.embed_item_GMF(item)
        output_GMF = embed_user_GMF*embed_item_GMF
        prediction = self.predict_layer(output_GMF)
        return prediction.view(-1)

---

MLP( Multilayer perceptron )

Usually , Each layer is fully connected to the next layer , The output of each artificial neuron on one layer becomes the input of several artificial neurons on the next layer .MLP There are at least three layers of artificial neurons , As shown in the figure below .

The formula is as follows ：

---

Four 、Neural CF hybrid model

In order to make the fusion model more flexible , We allow GMF and MLP Learning independent embeddedness , And combine the two models by connecting their final hidden layer output .

The formula is as follows ：