An information optimization model and device based on hybrid neural networks

Abstract

This invention relates to an information optimization model and apparatus based on a hybrid neural network. The invention constructs a user interaction graph and uses a graph neural network model to model user historical behavior, improving user representation capabilities while alleviating the problem of sparse interaction data. It also constructs an information optimization strategy initialization module that learns initial values ​​from user historical behavior based on the user-information interaction graph, mitigating the problem of zero-learning reducing user experience. For user-push information interaction behavior during use, a recurrent neural network is used to describe the interaction behavior, and information nodes are updated based on an information propagation algorithm to dynamically obtain the changing trends of user interests. Finally, it constructs a candidate set module for information optimization and designs an information optimization ranking module to solve the problem of a large information selection space, further improving information query and optimization performance.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and information retrieval technology, specifically relating to an information optimization model and device based on a hybrid neural network. Background Technology

[0002] With the advent of the big data era, the digital economy is gradually changing people's production and lifestyles, bringing great convenience to people's lives. However, whether it's searching for web pages in a search engine or finding files on a personal computer system, the essence of the business is to quickly select content and services that match the user's information needs from massive amounts of information, which is closely related to information retrieval and selection technologies.

[0003] The main goal of information retrieval and optimization is to find information that meets the user's information needs from large-scale heterogeneous datasets based on the user's query, and then optimize the information through ranking strategies. Essentially, this includes information retrieval and information ranking. Researchers both domestically and internationally have conducted extensive research on information retrieval models. Based on their development history, these models can be categorized into traditional rule-based information retrieval models, machine learning-based query ranking methods, and deep learning-based query ranking methods. Traditional information retrieval models use manually designed features and functions to build query models, making it difficult to integrate more parameters into the document scoring function. Machine learning-based query ranking methods utilize machine learning algorithms to learn ranking functions from large datasets and labels, enabling the integration of numerous features to build query models and thus improving query accuracy. Deep learning-based neural ranking methods use neural networks to build complex query models, learning the semantic information of text to improve the word mismatch problem in the first two methods. Currently, many search engine companies still largely rely on machine learning-based ranking learning models when providing search services. Although the accuracy of relevance search is improved compared to traditional information retrieval models, manually designed features are time-consuming and labor-intensive, and it is difficult to guarantee the completeness of the selected features. Deep learning, as a representation-based learning method, has the ability to discover hidden structures and features useful for prediction tasks from large-scale training data. Therefore, it can better alleviate the limitations of machine learning-based query methods.

[0004] In recent years, deep learning, with neural networks at its core, has been successfully applied to various tasks in speech recognition, computer vision, and natural language processing due to its powerful representation learning capabilities. Given the successful applications of deep learning in these fields, how to leverage it to improve the accuracy of relevance calculation in information retrieval has become a hot research topic. However, existing methods still have some shortcomings: ① User interaction data exhibits severe sparsity, affecting the representation of user and information features, thus reducing the effectiveness of information selection in information retrieval; ② It does not fully utilize comprehensive information such as users' historical behavior, preferences changing over time, and popularity information, which can influence user preferences; the expression of user information preference models needs further optimization; ③ The large scale and rapid update speed of information content result in a large space for information retrieval selection, affecting the efficiency and accuracy of information retrieval. Summary of the Invention

[0005] This patent proposes an information optimization model and device based on a hybrid neural network to solve problems such as sparse user feedback data, dynamic changes in user preferences, and excessive information space in information query and optimization.

[0006] To achieve the above technical objectives, the technical solution adopted by the present invention is: an information optimization model based on a hybrid neural network, the model comprising the following modules:

[0007] User behavior feature representation module: Based on graph convolutional neural network, feature learning and modeling of large-scale historical behavior of user and information interaction are performed to construct the initial interest information representation of users, which is used for the information optimization strategy initialization module;

[0008] Information Optimization Strategy Initialization Module: Based on the learning results of the user behavior feature expression module, the node representation of information and the initial value of information (the user's interest value in the information) are obtained using a fully connected network, and the information optimization initialization strategy is guided by the initial value of the information.

[0009] Dynamic update module: Based on recurrent neural network, it models the interaction behavior between users and push information, and is used to update the candidate set of the candidate set selection module, so that the model can continuously learn the dynamic changes in users' interests.

[0010] Candidate set selection module: Dynamic clustering is performed on the node representations obtained from the user behavior feature expression module and / or the dynamic update module to generate candidate sets.

[0011] Information Optimization and Sorting Module: Sorts the output results of the candidate set filtering module and generates the information optimization results.

[0012] Furthermore, the user behavior feature expression module is based on the knowledge graph G=(E,R) to model the historical behavior of user interaction with information and the attributes of information. By connecting the information nodes in the user interaction history with the user nodes and connecting the attribute nodes of the information with the information, a graph containing user behavior information is obtained, namely the behavior graph G'=(E',R').

[0013] We utilize behavioral graph G' and hybrid neural networks to learn feature representations of users and information; first, we initialize the vector representation i for each candidate piece of information. t ∈R d , where d represents the dimension of the vector representation, and then the node information in graph G' is propagated using a graph convolutional neural network (GCN) to obtain a better information feature representation.

[0014] Furthermore, the user behavior feature representation module utilizes a multi-layer graph convolutional neural network to learn the node representations in graph G'. The calculation steps for each layer are as follows: First, for each node, calculate its neighbor representations:

[0015]

[0016] In equation (1), N(h) represents the neighbors of node h. Let represent the vector of the i-th node in the k-th layer of the GCN network. Then, using the neighbor representation... Update the representation of the node itself:

[0017]

[0018] In equation (2), W k and B k These are the learnable network parameters, σ is the activation function, and after being calculated through multiple layers of graph convolutional neural networks, the node vector output by the last layer of the graph convolutional neural network is the node representation i of the information on graph G'. t The node representation of the user u t t represents time.

[0019] Furthermore, the activation function is the ReLU activation function, with the formula ReLU = max(0, x).

[0020] Furthermore, the information optimization strategy initialization module, based on the user behavior feature expression module, uses a fully connected network to learn the user's interest in information based on the user's historical behavior:

[0021] L=σ(W l concat(i t ,u t )+b l (3)

[0022] V=σ(W v L+b v (4)

[0023] In the formula, L is the output vector of the fully connected layer, and W... l The parameters represent the fully connected network, and the user representation node u is generated using the behavior graph representation module. t and the information representation node i t We obtain the initial value V of a certain piece of information and provide an initialization strategy for information selection and push based on this initial value.

[0024] The preferred candidate set screening module uses the node representation of the information obtained from the user behavior feature expression module and / or the user behavior feature expression module to perform dynamic clustering, including: every τ times of sorting and selection in the entire information selection process, the entire information space is re-clustered, and the initial cluster center of this clustering is the cluster center of the τ-1th clustering.

[0025] After clustering is completed, the representation of the cluster centers and the node representation of the users are input into the policy initialization module to obtain the user's interest value for each cluster, and the cluster with the highest interest value is selected as the user's basic candidate set.

[0026] Furthermore, the cluster with the highest interest value is selected as the user's basic candidate set. At the same time, a certain amount of information is randomly sampled from other clusters outside the basic candidate set to obtain the final candidate set. The information optimization result is generated based on the final candidate set.

[0027] Furthermore, the information optimization and sorting module uses a fully connected network to perform convolution calculations on the information node vectors in the candidate set, calculates the correlation between information and the correlation between information and user interests, sums the two correlations to obtain the information push score, sorts the information according to the push score from largest to smallest, and outputs the top K information, representing the top K information that the user is most interested in, where K is a system setting parameter.

[0028] Furthermore, after generating the initial K preferred pieces of information, the information is pushed to the user. The user will interact with the preferred information, indicating the user's preferences and changing trends regarding the pushed information. This is a serialized information. The dynamic update module uses a recurrent neural network to represent the information using the user's interactive behavior. The update of each cell unit of the recurrent neural network is shown below:

[0029] z t =σ g (W z i t +B z h t-1 +bz (5)

[0030] r t =σ g (W z i t +B z h t-1 +b r (6)

[0031] h' t =σ h (W h i t +B h (r t *h t-1 )+b h (7)

[0032] h t =(1-z) t )*h t-1 +z t h' t (8)

[0033] In the formula, i t The input information is represented by z. t and r t Let b represent the output vectors of the update gate and reset gate, b represent the bias parameters of the network layer, and W represent the output vectors of the reset gate and reset gate. (.) and B (.) Represents the learnable parameters of the network;

[0034] For the hidden layer representation h t The update uses a linear associative approach to update h obtained from the previous calculation. t-1 And the new hidden layer candidate h in this update t-1 Combination; hidden layer representation h t It is considered as a vector representation of the information at time t. This representation is an updated representation of the information based on user interaction behavior. This vector representation will be input into the preferred candidate set filtering module to update the candidate set and the result of the preferred information ranking.

[0035] The present invention also provides an information optimization device based on a hybrid neural network, characterized in that it includes the following modules:

[0036] User behavior feature representation module: Based on graph convolutional neural network, feature learning and modeling of large-scale historical behavior of user and information interaction are performed to construct the initial interest information representation of users;

[0037] Information Optimization Strategy Initialization Module: Based on the learning results of the user behavior feature expression module, the node representation of information and the initial value of information are obtained using a fully connected network, and the initial value of information guides the information optimization initialization strategy.

[0038] Dynamic update module: Models the interaction between users and push information based on recurrent neural networks;

[0039] The preferred candidate set selection module dynamically clusters node representations based on information obtained from the user behavior feature expression module and / or the dynamic update module to generate a candidate set.

[0040] Information Optimization and Sorting Module: Sorts the output results of the candidate set filtering module and generates the information optimization results.

[0041] This invention employs the aforementioned technical solutions. Compared to existing technologies, the user behavior feature expression module of this invention solves the problem of sparse feedback by constructing a behavior graph and using a graph convolutional neural network to model user representations. The information optimization strategy initialization module learns initial values ​​from user behavior history using node representations on the behavior graph, thus addressing the problem of learning from scratch harming user experience. For user interaction with push notifications during use, a recurrent neural network describes the interaction behavior, and a related propagation algorithm updates user information nodes, dynamically acquiring the user's current interests and dynamically grasping changes in user information. This invention also designs a candidate set selection module, using node representations of information for dynamic clustering to generate a candidate set with a smaller number of information items, thus solving the problem of a large information space. Finally, an information optimization and sorting module is designed to further improve information query and optimization performance. Attached Figure Description

[0042] Figure 1 Here is a flowchart of the information optimization model algorithm based on hybrid neural networks;

[0043] Figure 2 Flowchart of the algorithm for the user behavior feature representation module;

[0044] Figure 3 Flowchart of the information optimization strategy initialization module algorithm;

[0045] Figure 4 The flowchart of the algorithm for the candidate set selection module is shown below;

[0046] Figure 5 The algorithm flowchart for the dynamic update module; Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0048] This invention presents an information optimization model based on a hybrid neural network, comprising five key modules:

[0049] User behavior feature representation module: Based on graph convolutional neural network, feature learning and modeling of large-scale historical behavior of user and information interaction are performed to construct the initial interest information representation of users, which is used for the information optimization strategy initialization module;

[0050] Information Optimization Strategy Initialization Module: Based on the learning results of the User Behavior Feature Expression Module, the node representation of information and the initial value of information (the initial value represents the user's interest in the information) are obtained using a fully connected network, and the information optimization initialization strategy is guided based on the initial value of the information.

[0051] Dynamic update module: Based on recurrent neural network, it models the interaction behavior between users and push information, and is used to update the candidate set of the candidate set selection module, so that the model can continuously learn the dynamic changes in users' interests.

[0052] Candidate set selection module: Dynamic clustering is performed using node representations obtained from the user behavior feature expression module and / or the dynamic update module to generate candidate sets.

[0053] Information Optimization and Sorting Module: Sorts the output results of the candidate set filtering module and generates the information optimization results.

[0054] The specific algorithm flow of the above information optimization model is as follows:

[0055] The historical interaction data between users and information is represented by graphs to generate a user-information behavior graph, and a graph convolutional neural network is used to generate a serialized representation of users and a representation of information nodes.

[0056] Using graph neural networks to fit users' historical behavior, initial value is generated for information nodes;

[0057] Based on the initial value, the entire information space is dynamically clustered to generate a candidate set;

[0058] Based on the information in the candidate set, the push score of the information is calculated, and the final information optimization ranking result is generated.

[0059] Based on the user's interaction with the pushed information during the model usage process, a recurrent neural network is used to update the information node representation, and update the candidate set and information optimization ranking results.

[0060] (1) User behavior feature expression module

[0061] To address the sparsity problem of user-information interaction feedback, this model constructs a user behavior feature representation module, such as... Figure 2 As shown, the overall process consists of: ① modeling user historical behavior using graphs; ② learning features from the behavior graph using graph convolutional neural networks (GCNs).

[0062] ① First, construct the user behavior graph. This module is based on the knowledge graph G=(E,R) and models the historical behavior of user interaction with information and the attributes of information. By connecting the information nodes in the user interaction history with the user nodes and connecting the attribute nodes of the information with the information, a graph containing user behavior information is obtained, namely the behavior graph G'=(E',R').

[0063] ② Then, the behavioral graph and graph neural network are used to learn the feature representations of users and information. In the interactive information query and preferred recommendation system, the user's interaction behavior represents the user's preferences. This module learns the user's low-dimensional feature representation based on the graph neural network and the user's historical interaction behavior. First, the vector representation i is initialized for each candidate information. t ∈R d , where d represents the dimension of the vector representation. Then, a graph convolutional neural network is used to propagate the node information in the behavioral graph, thereby obtaining a better representation of information features.

[0064] This module utilizes a multi-layer graph convolutional neural network (GCN) to learn node representations in a graph. A GCN is a connection model that captures graph dependencies through message passing between nodes, and can represent information of arbitrary depth from a domain. The computation steps for each layer are as follows: First, for each node, this module computes its neighbor representations:

[0065]

[0066] In the formula, N(h) represents the neighbors of node h. This represents the vector of the i-th node in the k-th layer of the GCN network. Then, this module updates the node's own representation using the neighbor representation:

[0067]

[0068] Among them, W k and B kσ represents the learnable network parameters, and σ is the activation function. Here, this module chooses the ReLU activation function, with the formula ReLU = max(0, x). After computation through multiple layers of graph convolutional neural networks, the node vector output by the last layer of the GCN is the node representation i of the information in graph G'. t The node representation of the user u t .

[0069] The information representation generated by the user behavior feature expression module includes graph neighbor information, which effectively alleviates the problem of sparse feedback.

[0070] (2) Information Optimization Strategy Initialization Module

[0071] To address the system cold start problem, i.e., the severe damage to user experience caused by training from scratch in the early stages, this invention constructs an information optimization strategy initialization module, such as... Figure 3 As shown. This module utilizes users' historical interaction data to construct a strategy initialization module, generating initialization strategies to provide prior strategies for information query optimization, thereby guiding model training and ensuring good performance in the initial stages of information query push. This module uses a behavior graph G' to learn the initial value of user preferences for information by studying users' historical behavior, and uses this initial value to guide the information optimization and ranking module. Based on the behavior graph representation module, this module uses a fully connected network to learn users' interests in information based on their historical behavior:

[0072] L=σ(W l concat(i t ,u t )+b l (3)

[0073] V=σ(W v L+b v (4)

[0074] Where L is the output vector of the fully connected layer, and W l This represents the parameters of a fully connected network. This module utilizes a behavioral graph representation module to represent the user's representation node u. t and the information representation node i t The system obtains an initial value V for a given piece of information and provides an initialization strategy for information query optimization based on this initial value V. Through the information optimization strategy initialization module, the initial values ​​of candidate information are generated, thereby mitigating the problem of impaired user experience caused by the model learning from scratch.

[0075] (3) Optimized Candidate Set Screening Module

[0076] To address the issue of a large information space (information pool), this invention constructs a preferred candidate set selection module. This module selects node representation i from the information obtained in the user behavior feature expression module and employs the K-means clustering algorithm for dynamic clustering. Specifically, every τ iterations in the entire information selection and push process, the entire information space is re-clustered, with the initial cluster centers being the cluster centers from the τ-1th iteration. Dynamic clustering is used because, as the process progresses, the representation of information in the graph changes with increasing user interaction. Periodic re-clustering ensures that the candidate set selection module can adapt to these changes, better generating the optimal candidate set based on the characteristics of the information itself and user behavior.

[0077] After clustering, the cluster center representations and user node representations are input into the policy initialization module to obtain the user's interest value for each cluster. The cluster with the highest interest value is selected as the user's basic candidate set. Simultaneously, a certain amount of information is randomly sampled from other clusters and added to the basic candidate set to obtain the final candidate set. Through the candidate set filtering module, the information in the resulting candidate set contains the information that the user is most interested in. Furthermore, because the representations of this information are relatively similar, the model can learn more representative features between the information when learning from this candidate set, thereby further improving the information query optimization effect.

[0078] (4) Information Optimization and Sorting Module

[0079] After obtaining the user's representation, the final candidate set, and the user's initial interests, this invention also designs an information optimization and ranking module to combine this information with the training of the aforementioned module, further improving the performance of information query and optimized push. This module uses a fully connected network to perform convolution calculations on the information node vectors in the candidate set, calculating the correlation between information and the correlation between information and user interests. After summing the two correlations, the push score of the information is obtained. The information is then sorted from largest to smallest according to the push score, and the top K pieces of information are output, representing the top K pieces of information that the user is most interested in, where K is a system setting parameter.

[0080] (5) Dynamic update module

[0081] After generating the initial K preferred pieces of information, they are pushed to the user. The user will interact with the pushed preferred information, indicating the user's preferences and changing trends for the pushed information. This is a serialized information, which is represented by the user's interaction behavior through a recurrent neural network. The update of each cell unit of the recurrent neural network is shown below:

[0082] z t =σ g (W z it +B z h t-1 +b z (5)

[0083] r t =σ g (W z i t +B z h t-1 +b r (6)

[0084] h' t =σ h (W h i t +B h (r t *h t-1 )+b h (7)

[0085] h t =(1-z) t )*h t-1 +z t h' t (8)

[0086] In the formula, i t The input information is represented by z. t and r t Let b represent the output vectors of the update gate and reset gate, b represent the bias parameters of the network layer, and W represent the output vectors of the reset gate and reset gate. (.) and B (.) Represents the learnable parameters of the network;

[0087] For the hidden layer representation h t The update uses a linear associative approach to update h obtained from the previous calculation. t-1 And the new hidden layer candidate h in this update t-1 Combination; hidden layer representation h t It is regarded as a vector representation of the information at time t. This representation is an updated representation of the information based on user interaction behavior. This representation vector will be input into the preferred candidate set filtering module to update the candidate set and the result of the preferred information ranking.

Claims

1. An information optimization model based on a hybrid neural network, characterized in that, The model includes the following modules: User behavior feature representation module: Based on graph convolutional neural networks, feature learning and modeling of large-scale historical behavior of user interaction with information are performed to construct the initial interest information representation of users; Information optimization strategy initialization module: Based on the learning results of the user behavior feature expression module, the node representation of information and the initial value of information are obtained using a fully connected network, and the initial value of information guides the information optimization initialization strategy. Dynamic update module: Models the interaction between users and push information based on recurrent neural networks; The preferred candidate set selection module dynamically clusters node representations based on information obtained from the user behavior feature expression module and / or the dynamic update module to generate a candidate set. Information Optimization and Sorting Module: Sorts the output results of the candidate set filtering module and generates information optimization results; The user behavior feature expression module is based on graph G=(E,R) to model the historical behavior of user interaction with information and the attributes of information. By connecting information nodes in the user interaction history with user nodes and connecting attribute nodes of information with information, a behavior graph G'=(E',R') containing user behavior information is obtained. For each candidate information, initialize its vector representation i. t ∈R d , where d represents the dimension of the vector representation, and then, the node information in graph G' is propagated using a graph convolutional neural network to obtain the information feature representation; The preferred candidate set screening module performs dynamic clustering based on the node representation of the information obtained by the user behavior feature expression module and / or the dynamic update module, including: every τ sorting and selection in the entire information selection process, the entire information space is re-clustered, and the initial cluster center of this clustering is the cluster center of the τ-1th clustering. After clustering is completed, the representation of the cluster center and the node representation of the user are input into the policy initialization module to obtain the user's interest value for each cluster, and the cluster with the highest interest value is selected as the user's basic candidate set. The cluster with the highest interest value is selected as the user's basic candidate set. At the same time, a certain amount of information is randomly sampled from other clusters outside the basic candidate set to obtain the final candidate set. The information optimization and sorting module uses a fully connected network to perform convolution calculations on the information node vectors in the candidate set, calculates the correlation between information and the correlation between information and user interests, sums the two correlations to obtain the information push score, sorts the information according to the push score from largest to smallest, and outputs the top K information.

2. The information optimization model based on a hybrid neural network according to claim 1, characterized in that, The user behavior feature representation module uses a graph convolutional neural network to learn the node representations in graph G'. The calculation steps for each layer are as follows: For each node, calculate its neighbor representations: In the formula, N(h) represents the neighbors of node h. This represents the vector of the i-th node in the k-th layer GCN network; subsequently, it utilizes the neighbor representation. Update the representation of the node itself: In the formula, W k and B k These are the learnable network parameters, σ is the activation function, and after being calculated through multiple layers of graph convolutional neural networks, the node vector output by the last layer of the graph convolutional neural network is the node representation i of the information on graph G'. t The node representation of the user u t .

3. The information optimization model based on a hybrid neural network according to claim 2, characterized in that, The activation function is the ReLU activation function, and the formula is ReLU = max(0, x).

4. The information optimization model based on a hybrid neural network according to claim 2, characterized in that, The information optimization strategy initialization module, based on the user behavior feature expression module, uses a fully connected network to learn the user's interest in information based on the user's historical behavior. L=σ(W l concat(i t ,u t )+b l ) (3) V=σ(W v L+b v ) (4) In the formula, L is the output vector of the fully connected layer, and W... l The parameters represent the fully connected network, and the user representation node u is generated using the behavior graph representation module. t and the information representation node i t The initial value V of a certain piece of information is used to guide the selection of an initialization strategy.

5. The information optimization model based on a hybrid neural network according to claim 1, characterized in that, After generating the initial preferred information, it is pushed to the user. The user interacts with the preferred information, and the dynamic update module uses the user's interaction to represent the information through a recurrent neural network. The update of each cell unit of the recurrent neural network is shown below: z t =σ g (W z i t +B z h t-1 +b z ) (5) r t =σ g (W z i t +B z h t-1 +b r ) (6) h' t =σ h (W h i t +B h (r t *h t-1 )+b h ) (7) h t =(1-z t )*h t-1 +z t h' t (8) In the formula, i t The input information is represented by z. t and r t Let b represent the output vectors of the update gate and reset gate, b represent the bias parameters of the network layer, and W represent the output vectors of the reset gate and reset gate. (.) and B (.) Represents the learnable parameters of the network; Using a linear associative approach, the previously calculated h is used to... t-1 And the new hidden layer candidate h in this update t-1 Combined to update the hidden layer representation h t Hidden layer representation h t The vector representation of the information at time t is considered and input into the preferred candidate set filtering module.

6. An information optimization device based on a hybrid neural network, characterized in that, Includes the following modules: User behavior feature representation module: Based on graph convolutional neural network, feature learning and modeling of large-scale historical behavior of user and information interaction are performed to construct the initial interest information representation of users; Information optimization strategy initialization module: Based on the learning results of the user behavior feature expression module, the node representation of information and the initial value of information are obtained using a fully connected network, and the initial value of information guides the information optimization initialization strategy. Dynamic update module: Models the interaction between users and push information based on recurrent neural networks; The preferred candidate set selection module dynamically clusters node representations based on information obtained from the user behavior feature expression module and / or the dynamic update module to generate a candidate set. Information Optimization and Sorting Module: Sorts the output results of the candidate set filtering module and generates information optimization results; The user behavior feature expression module is based on graph G=(E,R) to model the historical behavior of user interaction with information and the attributes of information. By connecting information nodes in the user interaction history with user nodes and connecting attribute nodes of information with information, a behavior graph G'=(E',R') containing user behavior information is obtained. For each candidate information, initialize its vector representation i. t ∈R d , where d represents the dimension of the vector representation, and then, the node information in graph G' is propagated using a graph convolutional neural network to obtain the information feature representation; The preferred candidate set screening module performs dynamic clustering based on the node representation of the information obtained by the user behavior feature expression module and / or the dynamic update module, including: every τ sorting and selection in the entire information selection process, the entire information space is re-clustered, and the initial cluster center of this clustering is the cluster center of the τ-1th clustering. After clustering is completed, the representation of the cluster center and the node representation of the user are input into the policy initialization module to obtain the user's interest value for each cluster, and the cluster with the highest interest value is selected as the user's basic candidate set. The cluster with the highest interest value is selected as the user's basic candidate set. At the same time, a certain amount of information is randomly sampled from other clusters outside the basic candidate set to obtain the final candidate set. The information optimization and sorting module uses a fully connected network to perform convolution calculations on the information node vectors in the candidate set, calculates the correlation between information and the correlation between information and user interests, sums the two correlations to obtain the information push score, sorts the information according to the push score from largest to smallest, and outputs the top K information.

Images