A conversation recommendation method based on graph convolutional recurrent neural network with second-order aggregation

Abstract

The present invention discloses a conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation, comprising the following steps: step S0, constructing a directed conversation graph according to a conversation sequence, wherein each node in the conversation graph represents a click item; step S1, encoding each click item to obtain a low-dimensional embedding vector of each click item; step S2, inputting the low-dimensional embedding vector of each click item into a graph convolutional network and a recurrent neural network that fuses second-order collaborative signals to obtain a latent vector of each click item; step S3, learning to obtain a latent vector of the conversation based on the latent vector of each click item; step S4, multiplying the latent vector of each click item by the latent vector of the conversation, and inputting the result into a softmax layer to obtain a predicted click item. The conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation of the present invention has strong noise resistance and high accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and more specifically, relates to a conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation. Background Art

[0002] Recommendation systems have been widely used to solve the problem of information overload and recommend content that users may be interested in. However, sometimes it is impossible to obtain users' explicit preferences or interactive information such as ratings, and even the user's personal identity may be unknown. Only the user's historical click behavior during the current session is available, which brings about a cold start problem. Therefore, it is important to model limited behaviors in a session and make corresponding content recommendations.

[0003] Markov chain has achieved certain results in predicting the user's next behavior based on the user's previous behavior for session recommendation, but it only combines information independently for recommendation, and the dependency between session data cannot be processed.

[0004] With the introduction of deep learning, the recurrent neural network (RNN) method uses the input of user session click sequences to generate recommendation results, which can handle the dependencies between session data. However, the dependency of click item relationship conversion between sessions needs to be strengthened, and problems such as the processing of session graph node distribution structure cannot be solved.

[0005] In recent years, graph convolutional neural networks have been gradually applied to recommendation systems, using topological graphs to propagate aggregated node information. They have great advantages in learning recommendations based on graph structures, and their application in conversational recommendations needs further study.

[0006] Graph convolutional neural networks perform convolution and pooling based on the properties of graphs and are widely used in processing graph data to obtain the distribution structure information of session graph nodes. However, most current models only use the signals of first-order neighbors to update node embeddings in the graph convolution layer, which is easily affected by noise and the accuracy of recommendations needs to be improved. Summary of the invention

[0007] In order to solve the above problems, the present invention provides a conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation, which has strong noise resistance and high accuracy.

[0008] In order to achieve the above purpose, a technical solution adopted by the present invention is as follows:

[0009] A conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation includes the following steps:

[0010] Step S0, constructing a directed session graph according to the session sequence, wherein each node in the session graph represents a click item;

[0011] Step S1, encode each click item to obtain a low-dimensional embedding vector of each click item;

[0012] Step S2, inputting the low-dimensional embedding vector of each click item into the graph convolutional network and recurrent neural network that fuse the second-order collaborative signal to obtain the latent vector of each click item;

[0013] Step S3, obtaining a latent vector of the session based on the latent vector learning of each click item;

[0014] In step S4, the latent vector of each click item is multiplied by the latent vector of the session and input into the softmax layer to obtain the predicted click item.

[0015] Furthermore, the process of constructing the potential vector of the clicked item is as follows:

[0016] The low-dimensional embedding vector of each click item is input into the graph convolutional network that integrates the second-order collaborative signal to obtain the embedding vector of each click item, and then the embedding vector of each click item is input into the recurrent neural network to obtain the latent vector of the click item.

[0017] Furthermore, the embedding vector construction process of the clicked item is as follows:

[0018] Step S20: stack L graph convolutional layers to obtain the lth layer second-order collaborative signal of the clicked item

[0019]

[0020]

[0021] Where N is the click item E i The set of all 2nd-order adjacent click items of , Click item E j Passing adjacent click item E k To click item E i The lth layer second-order cooperative signal, P i，k，j is the second-order collaborative signal strength coefficient of the graph convolutional network, Respectively represent click items E i 、E j The l-1th layer embedding vector of

[0022] Step S21, using the second-order collaborative signal of the first layer of the click item and the l-1 layer embedding vector Embedding vector for clicked item To update:

[0023]

[0024] Among them, F l Represents the l-th layer embedding vector of all clicked items The constructed matrix, F l-1 The l-1th layer embedding vector representing all clicked items The constructed matrix; ReLU(·) is the activation function; operator Represents the matrix connection operation; S l (1) Represents the first-order coordination signal of the lth layer of all click items The constructed matrix; S l(2) Represents the lth-order second-order coordination signal of all click items The constructed matrix; Represents the weight matrix to be learned in the lth layer.

[0025] Furthermore, the potential vector h of the clicked item i The build operation is as follows:

[0026]

[0027]

[0028]

[0029]

[0030] Among them, Z i , R i are the reset gate and the update gate, representing the information to be discarded and retained; σ(·) represents the sigmoid function; M z 、M r 、M h ∈R 2d×d ; U z , U r , U h ∈R d×d is the parameter obtained after learning the model; h′ i is a candidate hidden state, tan h is the activation function, ⊙ is the element-wise multiplication operator, i∈{1, 2, ..., n}, and n is the total number of all clicked items in a single session.

[0031] Furthermore, step S3 comprises the following steps:

[0032] Step S30: Based on the potential vector h of each clicked item i Add to get the global vector h g :

[0033] Step S31: transform the global vector h gand the local vector h n Concatenate and then perform linear transformation to obtain the latent vector h of the session c :h c =W c [h g ;h n ];

[0034] Among them, W c is the parameter matrix obtained through model learning.

[0035] Furthermore, the number of layers of the graph convolution is L=3.

[0036] The above technical solution of the present invention has the following advantages compared with the prior art:

[0037] 1. Introducing second-order collaborative signals in the graph convolution layer and stacking multiple graph convolution layers are used to reduce the impact of noise when embedding click items and improve the accuracy of learning click item embedding; the accuracy is highest when L=3.

[0038] 2. The recurrent neural network layer obtains the temporal information of user sessions and enhances the dependencies between session data. It introduces global and local vectors to remove some noise in the original embedding vector space. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The present invention is a flowchart of a conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation in one embodiment of the present invention. Specific embodiments

[0040] The following will be combined with the accompanying drawings of the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] like Figure 1 As shown in the figure, a conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation includes the following steps:

[0042] Step S0: construct a directed session graph according to the session sequence, wherein each node in the session graph represents a click item; at this time, the click item set E = {E1, E2, ...E n};

[0043] Step S1: Encode each click item to obtain a low-dimensional embedding vector e for each click item. i :

[0044] e i=Enc(X,θ) (1)

[0045] In formula (1), Enc(·) is the encoding function, θ is the parameter to be learned of the encoding function Enc(·), and X represents the entity feature of the clicked item.

[0046] Step S2, inputting the low-dimensional embedding vector of each click item into the graph convolutional network and recurrent neural network that fuse the second-order collaborative signal to obtain the latent vector of each click item;

[0047] Specifically, the low-dimensional embedding vector of each click item is input into the graph convolutional network that integrates the second-order collaborative signal to obtain the embedding vector of each click item, and then the embedding vector of each click item is input into the recurrent neural network to obtain the latent vector of the click item.

[0048] The embedding vector construction process of the clicked item is as follows:

[0049] Step S20: stack L graph convolutional layers to obtain the lth layer second-order collaborative signal of the clicked item

[0050]

[0051]

[0052] In formula (2) and formula (3), N is the click item E i The set of all 2nd-order adjacent click items of , Click item E j Passing adjacent click item E k To click item E i The lth layer second-order cooperative signal, P i，k，j is the second-order collaborative signal strength coefficient of the graph convolutional network, Respectively represent click items E i 、E j The l-1th layer embedding vector of

[0053] Step S21, using the second-order collaborative signal of the first layer of the click item and the l-1 layer embedding vector Embedding vector for clicked item To update:

[0054]

[0055] In formula (4), F l The l-th layer embedding vector representing all clicked items The constructed matrix, F l-1 The l-1th layer embedding vector representing all clicked items The constructed matrix; ReLU(·) is the activation function; operator Represents the matrix connection operation; S l(1) Represents the first-order coordination signal of the lth layer of all click items The constructed matrix; S l(2) Represents the lth-order second-order coordination signal of all click items The constructed matrix; Represents the weight matrix to be learned in the lth layer. l(1) Specifically

[0056] In formula (5), F i←i Used to retain the feature information of the original node; Q represents the click item E i The set of all first-order neighbor click items in the lth layer; Indicates the adjacent click item E k To click item E i LeakyReLU(·) is the activation function.

[0057] The latent vector h of the clicked item i The build operation is as follows:

[0058]

[0059]

[0060]

[0061]

[0062] In formulas (6)-(9), Z i , R i are the reset gate and the update gate, representing the information to be discarded and retained; σ(·) represents the sigmoid function; M z 、M r 、M h ∈R 2d×d ; U z , U r , U h ∈R d×d is the parameter obtained after learning the model; h′ i is a candidate hidden state, tan h is the activation function, ⊙ is the element-wise multiplication operator, i∈{1, 2, ..., n}, and n is the total number of all clicked items in a single session.

[0063] Step S3, obtaining a latent vector of the session based on the latent vector learning of each click item;

[0064] Specifically, step S3 includes the following steps:

[0065] Step S30: Based on the potential vector h of each click item i Add to get the global vector h g :

[0066] Step S31: transform the global vector h g and the local vector h n Concatenate and then perform linear transformation to obtain the latent vector h of the session c :

[0067] h c =W c [h g ;h n ] (10)

[0068] In formula (10), W c is the parameter matrix obtained through model learning.

[0069] Step S4: transform the potential vector h of each click item i and the latent vector h of the session c Multiply them and input them into the softmax layer to get the predicted click items, where the potential vector h of each click item is i and the latent vector h of the session c The multiplication is The model corresponding to the above parameters or parameter matrices is a model constructed by graph convolutional network, recurrent neural network and linear transformation. The model uses cross entropy loss function to reversely update parameters for training, and the optimizer uses Adam optimizer. The entropy loss function is specifically:

[0070]

[0071] In formula (11), y i One-hot encoding vector representing the real clicked item, y′ i It is the encoding vector corresponding to the click item with the highest score predicted by softmax.

[0072] Preferably, the number of layers of the graph convolution is L=3.

[0073] The advantages of the conversation recommendation method of the present invention are described below with specific embodiments:

[0074] The main parameter configuration of the model is: the embedding dimension of the clicked item is set to 64, the number of graph convolution layers L is set to 3, the optimizer uses the Adam optimizer, and the learning rate is set to 10 -3 , the maximum number of iterations is 1000.

[0075] Tests were conducted on three data sets (Diginetica, Yoochoose1 / 64, Yoochoose1 / 4). In order to evaluate the effect of the method of the present invention, the methods compared with the method proposed in the present invention when training the model include: NeuCF, GCMC, NGCF, GraphRec, GCRNN, these five mainstream algorithm models. The specific meanings of the above five algorithm models are as follows:

[0076] NeuCF: A recommendation model based on matrix decomposition that uses deep neural networks to replace the inner product operation of embedded vectors.

[0077] GCMC: Uses one graph convolutional layer to aggregate the collaborative signals of first-order neighbors.

[0078] NGCF: Implicitly propagating aggregated collaborative signals by stacking multiple graph convolutional layers.

[0079] GraphRec: A collaborative filtering recommendation model based on GCN.

[0080] GCRNN: Conversational embedding nodes are obtained by aggregating the collaborative signals of first-order neighbors and recurrent neural network layers.

[0081] This embodiment uses two evaluation indicators P@20 and MRR@20, whose specific meanings are as follows:

[0082] P@20 (precision): indicates the proportion of accurate predictions among the top 20 recommended products (click items), and is a widely used method for evaluating prediction accuracy.

[0083] MRR@20 (average reciprocal ranking): If the ranking is over 20, the reciprocal ranking is set to 0. The larger the MRR value, the higher the accurately recommended product (click item) is in the recommended ranking list.

[0084] The experimental results of the conversation recommendation method of the present invention and the other five methods on the data set are shown in Table 1:

[0085] Table 1: Experimental results of the conversation recommendation method of the present invention and other five methods on the dataset

[0086]

[0087] As can be seen from Table 1, the prediction accuracy of the four recommendation models based on graph convolutional neural networks is significantly better than that of the NeuCF model based on matrix decomposition, which shows that the graph convolutional neural network recommendation model that integrates the entity features of the click item and the graph structure features on the topological graph can more accurately learn the embedded representation of the click item, thereby obtaining better prediction results; the prediction accuracy of the NGCF and GraphRec models using high-order collaborative signals is better than that of the GCMC model using first-order aggregation, which shows that aggregating high-order collaborative signals can make the recommendation results more optimized; the GCRNN model introduces a recurrent neural network while aggregating the first-order collaborative signal, so that the temporal information and dependency between sessions are further enhanced. According to the training results of the data set, the recommendation results of the GCRNN model are more accurate than those of the GCMC model, which shows that the recurrent neural network can capture the potential information between user sessions and help improve the performance of the recommendation model. The excellence of the present invention lies in the introduction of second-order aggregation and multi-layer stacking in the graph convolution layer, which increases the utilization of higher-order collaborative signals, and introduces a recurrent neural network to effectively process the propagation of information between sessions, capturing more accurate potential information between user sessions.

[0088] The above embodiments only express several implementation methods of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the present application. For those skilled in the art, the present invention can have various changes and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation, characterized in that: The following steps are involved: Step S0, constructing a directed session graph according to the session sequence, wherein each node in the session graph represents a click item; Step S1, encode each click item to obtain a low-dimensional embedding vector of each click item; Step S2, inputting the low-dimensional embedding vector of each click item into the graph convolutional network and recurrent neural network that fuse the second-order collaborative signal to obtain the latent vector of each click item; Step S3, obtaining a latent vector of the session based on the latent vector learning of each click item; Step S4, multiplying the latent vector of each click item with the latent vector of the session, and inputting the result into the softmax layer to obtain the predicted click item; The process of constructing the low-dimensional embedding vector of the click item is as follows: Step S20: stack L graph convolutional layers to obtain the lth layer second-order collaborative signal of the click item Where N is the click item E i The set of all 2nd-order adjacent click items of , Click item E j Passing adjacent click item E k To click item E i The lth layer second-order cooperative signal, P i,k,j is the second-order collaborative signal strength coefficient of the graph convolutional network, Respectively represent click item E i 、E j The l-1th layer embedding vector of Step S21, using the second-order collaborative signal of the first layer of the click item and the l-1 layer embedding vector Embedding vector for clicked item To update: Among them, F l The l-th layer embedding vector representing all clicked items The constructed matrix, F l-1 The l-1th layer embedding vector representing all clicked items The constructed matrix; ReLU(·) is the activation function; operator Represents the matrix connection operation; S l(1) Represents the first-order coordination signal of the lth layer of all click items The constructed matrix; S l(2) Represents the lth-order second-order coordination signal of all click items The constructed matrix; Represents the weight matrix to be learned in the lth layer.

2. According to claim 1, a conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation is characterized in that: The process of constructing the potential vector of the click item is as follows: The low-dimensional embedding vector of each click item is input into the graph convolutional network that integrates the second-order collaborative signal to obtain the embedding vector of each click item, and then the embedding vector of each click item is input into the recurrent neural network to obtain the latent vector of the click item.

3. According to claim 2, a conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation, characterized in that: The latent vector h of the clicked item i The build operation is as follows: Among them, Z i , R i are the reset gate and the update gate, representing the information to be discarded and retained; σ(·) represents the sigmoid function; M z 、M r 、M h ∈R 2d×d ; U z , U r , U h ∈R d×d is the parameter obtained after learning the model; h' i is a candidate hidden state, tanh is the activation function, ⊙ is the element-wise multiplication operator, i∈{1,2,…,n}, and n is the total number of all clicked items in a single session.

4. According to claim 3, a conversation recommendation method based on a graph convolutional recurrent neural network with second-order aggregation, characterized in that: Step S3 includes the following steps: Step S30: Based on the potential vector h of each click item i Add to get the global vector h g : Step S31: The global vector h g and the local vector h n Concatenate and then perform linear transformation to obtain the latent vector h of the session c :h c =W c [h g ;h n ]; Among them, W c is the parameter matrix obtained through model learning.

5. A conversation recommendation method based on a graph convolutional recurrent neural network based on second-order aggregation according to any one of claims 3-4, characterized in that: The number of layers of the graph convolution is L=3.

Images