Multi-modal network fusion method and system of HIV infection source tracking and intervention strategy

Abstract

The invention discloses a multi-modal network fusion method and system for an HIV infection source tracking and intervention strategy. The method comprises the steps that an index immediate network adversarial generative graph model is utilized to construct a multi-modal network for HIV transmission; identifying key nodes in the virus transmission process; and based on the key nodes of the multi-modal network, implementing an intervention strategy and evaluating an intervention effect. According to the method, the unified multi-modal complex network is constructed by fusing the online social network, the offline behavior network and the HIV molecular network, so that the problem of insufficient data integration in the prior art is solved. A multi-modal node importance sorting algorithm based on deep learning is introduced, and key propagation nodes in the network are accurately identified in combination with a reinforcement learning technology. And a partition nested random network search and data driven deep reinforcement learning algorithm is adopted, so that the calculation complexity of key node identification and intervention strategy optimization is remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of public health and AIDS prevention and control, and particularly to a multi-modal network fusion method for HIV source tracking and intervention strategies. Background Art

[0002] There are multiple limitations in the existing technology for HIV transmission research, which are mainly reflected in network data integration, key node identification, intervention strategy optimization, and computational complexity. First of all, current research is usually limited to a single type of network analysis, such as online social networks or molecular networks, and it is difficult to reveal the multi-modal characteristics of virus transmission, especially in complex MSM groups. Due to the lack of organic integration of online / offline social networks and molecular networks, the characterization of transmission paths often lacks comprehensiveness and accuracy. In addition, traditional epidemiological questionnaire survey methods are limited by recall bias and incomplete data, and the identification of key nodes is mostly based on simple network centrality indicators, such as degree, betweenness centrality, etc.

[0003] There are also deficiencies in intervention strategy optimization. Existing evaluation methods based on the SIR model usually rely on uniform assumptions, ignoring node heterogeneity and unevenness of transmission paths in complex networks. This limitation makes the intervention strategy lack pertinence and it is difficult to give full play to the role of key nodes in controlling HIV transmission. At the same time, the task of identifying key nodes in multi-layer complex networks belongs to the NP-hard problem with high computational complexity. Existing technologies lack effective optimization methods when dealing with these problems, such as the combination of partitioning algorithms and deep reinforcement learning, resulting in low analysis efficiency for large-scale complex networks and difficulty in implementing intervention optimization.

[0004] Theoretical models such as the SIR model are based on uniform assumptions and fail to reflect node heterogeneity and complex network coupling characteristics in real networks. Limitations of computing resources and technical tools further exacerbate these problems. The analysis of multi-layer complex networks requires high computing power, and existing methods lack optimization means when dealing with NP-hard problems and it is difficult to effectively process large-scale networks. Summary of the Invention

[0005] The present invention provides a multi-modal network fusion method and system for HIV source tracking and intervention strategies. In view of the above problems, innovative technical solutions are proposed to improve the accuracy and efficiency of HIV transmission research and intervention.

[0006] In a first aspect, the present invention provides a multi-modal network fusion method for HIV source tracking and intervention strategies, which includes:

[0007] S1 - Using an exponential random network adversarial generation graph model to construct a multi-modal network for the transmission of the human immunodeficiency virus HIV;

[0008] S2 - Identify the key nodes in the process of virus transmission;

[0009] S3 - Based on the key nodes of the multimodal network, implement intervention strategies and evaluate the intervention effects.

[0010] In a second aspect, a multimodal network fusion system for HIV source tracking and intervention strategies, comprising:

[0011] A multimodal network construction module, configured to construct a multimodal network for the transmission of human immunodeficiency virus (HIV) by using an exponential random network adversarial generation graph model;

[0012] A key node identification module, configured to identify the key nodes in the process of virus transmission;

[0013] A network intervention evaluation module, based on the key nodes of the multimodal network, implements intervention strategies and evaluates the intervention effects.

[0014] Beneficial effects

[0015] 1) By integrating online social networks, offline behavior networks, and HIV molecular networks, the present invention constructs a unified multimodal complex network, solving the problem of insufficient data integration in traditional research. In the process of network construction, an exponential random network generation model (ERGM) is used to capture the dynamic characteristics of social networks, and a deep adversarial generation network (GAN) is used to generate multimodal interaction data, thereby comprehensively depicting the HIV transmission path. Such a multimodal network can accurately reflect the dynamic characteristics of virus transmission, providing a scientific basis for in-depth analysis of transmission patterns.

[0016] 2) In terms of key node identification, the present invention introduces a multimodal node importance ranking algorithm based on deep learning and combines reinforcement learning techniques to accurately identify the key transmission nodes in the network. Considering the network characteristics, transmission dynamics attributes, and data diversity of nodes, this method significantly improves the accuracy and efficiency of key node identification.

[0017] 3) In terms of optimizing intervention strategies, the present invention combines transmission dynamics modeling with a reinforcement learning optimization algorithm, extends the SIR model, and uses Monte Carlo simulation to simulate the effects of different intervention strategies to ensure the scientific nature and pertinence of the strategies. At the same time, to break through the computational bottleneck in complex network optimization, the present invention adopts a partitioned nested random network search and data-driven deep reinforcement learning algorithm, significantly reducing the computational complexity of key node identification and intervention strategy optimization, providing an efficient solution for practical applications.

[0018] 4) The core objective of the present invention is to achieve the source tracking and transmission tracing of HIV virus based on a multi-modal network, and improve the efficiency of public health prevention and control through intelligent intervention strategies. This innovative technology breaks through the limitations of traditional HIV prevention and control technologies, provides an important reference for the application of complex network analysis in the field of disease transmission, and opens up a new path for the precise prevention and control of HIV transmission, with important theoretical significance and broad practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a schematic flow chart of a multi-modal network fusion method for HIV source tracking and intervention strategy provided in Embodiment 1 of the present invention;

[0020] Figure 2 is a schematic diagram of an adversarial generative network deep learning model of a random walk type function provided by the present invention;

[0021] Figure 3 is a schematic diagram of the Graph-ACISD reinforcement learning optimization algorithm provided by the present invention;

[0022] Figure 4 is a schematic structural diagram of a multi-modal network fusion system for HIV source tracking and intervention strategy provided in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0024] See Figure 1 , which is a schematic flow chart of a multi-modal network fusion method for HIV source tracking and intervention strategy provided in Embodiment 1 of the present invention. The method includes the following steps:

[0025] Step 1, use an exponential random network adversarial generative graph model to construct a multi-modal network for the transmission of the human immunodeficiency virus (HIV).

[0026] In this step, in this embodiment, an adversarial generative network deep learning model based on an exponential random walk type function is applied to integrate and construct the multi-modal network. The schematic diagram of the model is as shown in Figure 1As shown in the figure. A graph representation model GraphSAGE (Graph Sample and Aggregated) and a graph generation model DiffusionGGM (Diffusion-based Graph Generative Model) are used to construct a multi-level complex network for men who have sex with men.

[0027] First, define the network data source for each layer as S i , where i represents the i-th layer of the network, and construct the corresponding graph G i . Convert each graph G i into an adjacency matrix and a feature matrix, denoted as A i and X i respectively. Use GraphSAGE to initialize an initial graph representation, denoted as For each data source S i , perform T iterations, where the hyperparameter T represents the number of iterations, that is:

[0028]

[0029] Concatenate the final representations i obtained for each data source S together to form a multi-level complex network representation:

[0030]

[0031] Finally, use the graph generation model DiffusionGGM to learn the latent representation of the fused multi-level complex network:

[0032] Graph autoencoder: Encoder: Z = DiffusionGGM_Encoder(H multi )

[0033] Graph decoder: Decoder: A recon = DiffusionGGM_Encoder(Z)

[0034] Step 2, identify the key nodes in the virus transmission process.

[0035] In this step, this embodiment proposes an algorithm for ranking the importance of development nodes. An extended model of the memory SIR (Susceptible-Infected-Recovered) virus transmission dynamics equation is adopted, and Monte Carlo simulation is used to simulate the transmission mode of the AIDS virus and different intervention measures in the constructed multi-modal network, and identify the key nodes in the virus transmission process. Here, the key nodes refer to the infected or potential infected individuals with a wide range of HIV transmission capabilities in the multi-modal network and the nodes that play a key bridging role in the transmission process. This embodiment designs a graph-based deep learning algorithm (Deep Graph Key Node Identification, DGKNI) to identify the key nodes of complex networks and takes into account the coupling problem of multi-level networks. The present invention is based on the Graph Autoencoder (GAE) and simultaneously introduces a graph attention mechanism to handle the coupling of multi-level networks.

[0036] First, use the graph autoencoder GAE to learn the low-dimensional representation of the nodes. Denote X as the node feature matrix, W0 and W1 as the weight matrices, b0 and b1 as the bias vectors, and σ as the activation function.

[0037] h i = σ(W1·σ(W0·X + b0) + b1)

[0038] Introduce the graph attention mechanism so that the network can pay attention to the contributions of nodes at different levels to the key node identification task. W2, W3, W4, a, and b2 are all parameters, and N i is the neighbor of node i.

[0039] α ij = softmax j (LeakyReLU(a T [W2h i , W3h j ))

[0040] h' i = ∑ j∈Ni α ij h j

[0041] The score of the key node using the prediction layer is:

[0042] y i = σ(W4h' i + b2)

[0043] The loss function uses the binary cross-entropy loss function. In this algorithm, the graph autoencoder is used to learn the low-dimensional representation of nodes, the graph attention mechanism is used to handle the coupling of multi-level networks, and the prediction of key nodes is completed through the prediction layer. The algorithm optimizes the model by minimizing the loss function so that it can accurately identify key nodes in complex networks.

[0044] Step 3: Based on the key nodes of the multi-modal network, implement an intervention strategy and evaluate the intervention effect.

[0045] In this step, based on the key nodes identified in Step 2, the present invention proposes a data-driven deep learning reinforcement algorithm for Markov decision processes, which combines the partitioned nested stochastic network search proposed in Step 1 and the memory priority algorithm in Step 2 to implement the above intervention strategy to achieve the effectiveness of precise prevention and control and precise intervention. Step 3 of this embodiment is based on the graph reinforcement learning algorithm Graph-ACISD (Graph driven and Actor-Critic based Intervention Strategy Designer) of the classic Actor-Critic framework. The ACISD model includes three modules: the graph embedding module (Graph Embedding Module), the actor module (Actor module), and the critic module (Critic module), and its structure is as Figure 2 shown.

[0046] S3-1 Graph Embedding Module

[0047] This step uses the graph embedding method to learn the spatial features of nodes in the network. Given the network:

[0048] G=(V,ε)

[0049] They respectively represent the set V of all nodes and the set ε of all edges in the network G.

[0050] The graph embedding module maps each node in the network into a low-dimensional dense vector (embedding vector) to express the spatial features of the node, as follows

[0051]

[0052] Let |V| denote the number of nodes in the network, and d1 denote the dimension of the node embedding vector. Currently, common graph embedding algorithms can be divided into three categories: graph factorization-based embedding algorithms (such as LINE, HOPE), random walk-based embedding algorithms (such as DeepWalk, Node2vec), and graph neural network-based embedding algorithms (such as GraphSAGE, Graph Convolution Networks). The ACISD model will select the optimal algorithm from the above classical algorithms as the final GraphEmb function.

[0053] S3-2 Actor Module

[0054] The Actor module needs to give the optimal intervention strategy based on the node features learned by the graph embedding module, that is, the set of key nodes that should receive the intervention. As shown in the figure above, after receiving the embedding vectors of all nodes in the network, the multi-layer perceptron (MLP) first maps them to a non-linear hidden layer, and then generates a score vector by the linear output layer:

[0055] SV ∈ R |V|

[0056] Each element value in the score vector represents the possibility that the corresponding node is a key node. The higher the score vector value of each node, the more likely it is that the node is a key node. Finally, the Actor module gives the set of key nodes that should receive the intervention based on the score vector. The screening result is represented by the binary vector shown below:

[0057] x ∈ R |V|

[0058] x i ∈ {0, 1}

[0059] A value of 1 indicates that the node is selected as a key node, and a value of 0 indicates that the node is not selected as a key node. The formula shown below represents the maximum number of nodes K that can be intervened under a given intervention budget:

[0060] Σ i x i = K

[0061] The most direct way to screen key nodes is to select the nodes corresponding to the largest K element values in the score vector as key nodes. The process is shown in the following formula:

[0062]

[0063] However, since this process is not differentiable, the Gumbel Softmax sampling method is used to propagate the gradient at discrete nodes. Finally, the score vector processed by the Gumbel samples is connected to the Critic module via the Straight-Through Estimator. The Straight-Through Estimator uses the argKlargest function to output a binary vector to represent the positions of key nodes during forward propagation; and uses the Softmax function to propagate the gradient to achieve gradient propagation during backward propagation.

[0064] The loss function of the Actor module is shown as follows:

[0065] J A = -Critic(G, Actor(G))

[0066] S3-3 Critic Module

[0067] The Critic module needs to evaluate the effectiveness of the strategy based on the multi-layer complex network information of MSM population and the intervention strategy output by the Actor module. To capture the overall information of the multi-layer complex network of MSM population, the Critic uses a Filter-Aggregator composed of multi-layer convolutional layers to filter out the key information in the initial embedding vector of nodes and aggregate the feature information of neighbor nodes. The node feature vector processed by the Filter-Aggregator is shown as follows:

[0068] J C = [Siudtr(G, x) - C]

[0069] C represents a constant or reference value used to adjust the range of the output J C The processed node feature vector is directly fed into the pooling layer for further compression on the one hand, and the compression result is the network feature vector P, which summarizes the overall information of the multi-layer complex network of the target MSM population; on the other hand, it is combined with the intervention strategy output by the Actor module, and after being compressed by another pooling layer, the overall vector Q of key nodes is output, which summarizes the main information of all key nodes.

[0070] The present invention uses the Kronecker product of the network feature vector P and the overall vector Q of key nodes to explore the mutual relationship between network information and key node information. The product result is mapped into a scalar r through multiple non-linear layers (ReLu, Sigmoid), and this scalar is the score finally generated by the Critic module to evaluate the efficiency of the current intervention strategy. The Critic module will update its own score generation system (module parameters) according to the simulation results of the current intervention strategy based on the epidemic transmission dynamics model, so that the score generated by the Critic module is closer to the simulation results of the transmission dynamics model.

[0071] The loss function of the Critic module is:

[0072] J C =[Siudtr(G,x)-Critic(G,x)] 2

[0073] In this embodiment, by cooperating with relevant dating platforms, desensitized data of online social network features is obtained, and data of the HIV molecular network is used to obtain relatively complete offline social network data. The multi-modal network constructed based on the online / offline social network and the HIV molecular network more completely and accurately displays the network features of HIV transmission among key populations. The present invention uses an adversarial generative network deep learning model based on an exponential random walk type function to integrate and construct the multi-modal network. Within the scope of a city, a multi-modal network based on the online / offline social network and the molecular network is constructed for a type of key population (such as MSM) and key nodes are identified, and then the implementation and effect evaluation of precise intervention are guided. The present invention combines a partitioned nested random network search memory priority algorithm and a data-driven Markov decision process deep reinforcement learning algorithm to solve the NP-Hard technical difficulties.

[0074] See Figure 1 , which is a multi-modal network fusion system for HIV source tracing and intervention strategy provided by Embodiment 2 of the present invention, and includes a multi-modal network construction module, a key node identification module, and a network intervention evaluation module.

[0075] The multi-modal network construction module is used to construct a multi-modal network for the transmission of the human immunodeficiency virus (HIV) by using an exponential random network adversarial generation graph model.

[0076] The key node identification module is used to identify key nodes in the process of virus transmission.

[0077] The network intervention evaluation module implements an intervention strategy and evaluates the intervention effect based on the key nodes of the multi-modal network.

[0078] The construction steps of the multimodal network construction module are as follows:

[0079] In this embodiment, an adversarial generative network deep learning model based on an exponential random walk type function is applied to integrate and construct the multimodal network. The schematic diagram of the model is as shown in Figure 1 . A graph representation model GraphSAGE (Graph Sample and Aggregated) and a graph generation model DiffusionGGM (Diffusion-based Graph Generative Model) are used to construct a multi-level complex network for men who have sex with men.

[0080] First, define the network data source of each layer as S i , where i represents the i-th layer of the network, and construct the corresponding graph G i . Convert each graph G i into an adjacency matrix and a feature matrix, denoted as A i and X i respectively. Use GraphSAGE to initialize an initial graph representation, denoted as For each data source S i , perform T iterations, where the hyperparameter T represents the number of iterations, that is:

[0081]

[0082] Concatenate the final representations i obtained for each data source S together to form a multi-level complex network representation:

[0083] H multi = Concatenate(H1 (T) , H2 (T) ,..., H k (T) )

[0084] Finally, use the graph generation model DiffusionGGM to learn the latent representation of the fused multi-level complex network:

[0085] Graph autoencoder: Encoder: Z = DiffusionGGM_Encoder(H multi )

[0086] Graph decoder: Decoder: A recon = DiffusionGGM_Encoder(Z)

[0087] The recognition steps of the key node recognition module are as follows:

[0088] This embodiment proposes an algorithm for ranking the importance of development nodes. An extended model that memorizes the SIR (Susceptible Infected Recovered) virus transmission dynamics equation is used, and Monte Carlo simulation is adopted to simulate the transmission pattern of the AIDS virus and different intervention measures in the constructed multimodal network, identifying the key nodes in the virus transmission process. Here, the key nodes refer to the infected or potentially infected individuals with a wide range of HIV transmission capabilities in the multimodal network and the nodes that play a key bridging role in the transmission process. This embodiment designs a graph-based deep learning algorithm (Deep Graph Key Node Identification, DGKNI) to identify the key nodes of complex networks and takes into account the coupling problem of multi-level networks. The present invention is based on the Graph Autoencoder (GAE) and simultaneously introduces the graph attention mechanism to handle the coupling of multi-level networks.

[0089] First, use the graph autoencoder GAE to learn the low-dimensional representation of nodes. Denote X as the node feature matrix, W0 and W1 as the weight matrices, b0 and b1 as the bias vectors, and σ as the activation function.

[0090] h i = σ(W1·σ(W0·X + b0) + b1)

[0091] Introduce the graph attention mechanism so that the network can focus on the contributions of nodes at different levels to the key node identification task. W2, W3, W4, and a, b2 are all parameters, and N i is the neighbor of node i.

[0092] α ij = softmax j (LeakyReLU(a T [W2h i , W3h j ))

[0093]

[0094] The score of the key node using the prediction layer is:

[0095] y i = σ(W4h' i + b2)

[0096] The loss function uses the binary cross-entropy loss function. In this algorithm, the graph autoencoder is used to learn the low-dimensional representation of nodes, the graph attention mechanism is used to handle the coupling of multi-level networks, and the prediction of key nodes is completed through the prediction layer. The algorithm optimizes the model by minimizing the loss function so that it can accurately identify the key nodes in complex networks.

[0097] The network intervention evaluation module is based on the graph reinforcement learning algorithm Graph-ACISD (Graph driven and Actor-Critic based Intervention Strategy Designer) of the classical Actor-Critic framework. The ACISD model includes three modules: a graph embedding module (Graph Embedding Module), an actor module (Actor module), and a critic module (Critic module). Its structure is as Figure 2 shown.

[0098] The graph embedding module is used to learn the spatial features of nodes in the network by using graph embedding methods. Specifically, given a network:

[0099] G = (V, ε)

[0100] represent the set V of all nodes and the set ε of all edges in the network G, respectively.

[0101] The graph embedding module maps each node in the network into a low-dimensional dense vector (embedding vector) to express the spatial features of the node, as follows

[0102]

[0103] |V| represents the number of nodes in the network, and d1 represents the dimension of the node embedding vector. Currently, common graph embedding algorithms can be divided into three categories: embedding algorithms based on graph factorization (such as LINE, HOPE), embedding algorithms based on random walk (such as DeepWalk, Node2vec), and embedding algorithms based on graph neural network (such as GraphSAGE, Graph Convolution Networks). The ACISD model will select the optimal algorithm from the above classical algorithms as the final GraphEmb function.

[0104] The Actor module needs to give the optimal intervention strategy based on the node features learned by the graph embedding module, that is, the set of key nodes that should be intervened. As shown in the above figure, after the multi-layer perceptron (MLP) receives the embedding vectors of all nodes in the network, it first maps them to a non-linear hidden layer, and then generates a score vector (Score Vector) by the linear output layer:

[0105] SV ∈ R |V|

[0106] Each element value in the score vector represents the possibility that the corresponding node is a key node. The higher the score vector value of each node, the more likely the node is to be a key node. Finally, the Actor module gives a set of key nodes that should receive interventions based on the score vector. The screening result is represented by the binary vector shown below:

[0107] x ∈ R |V|

[0108] x i ∈ {0, 1}

[0109] A value of 1 indicates that the node is selected as a key node, and a value of 0 indicates that the node is not selected as a key node. The formula shown below represents the maximum number of nodes K that can be intervened under a given intervention budget:

[0110] ∑ i x i = K

[0111] The most direct key node screening method is to select the nodes corresponding to the largest K element values in the score vector as key nodes, and its process is shown by the following formula:

[0112]

[0113] However, since this process is not differentiable, the Gumbel Softmax sampling method is used to propagate the gradient at discrete nodes. Finally, the score vector processed by Gumbel samples is connected to the Critic module via a Straight-Through Estimator. The Straight-Through Estimator uses the argKlargest function to output a binary vector to represent the key node positions during forward propagation; and uses the Softmax function to propagate the gradient during backpropagation to achieve gradient propagation.

[0114] The loss function of the Actor module is shown below:

[0115] J A = -Critic(G, Actor(G))

[0116] The Critic module needs to evaluate the effectiveness of the intervention strategy based on the multi-layer complex network information of men who have sex with men and the output of the Actor module. To capture the overall information of the multi-layer complex network of men who have sex with men, the Critic uses a Filter-Aggregator composed of multi-layer convolutional layers to filter out the key information in the initial embedding vector of the nodes and aggregate the feature information of neighbor nodes. The node feature vector processed by the Filter-Aggregator is shown below:

[0117] J C = [Siudtr(G, x) - C]

[0118] C represents a constant or reference value used to adjust the output J C The range of, M represents the feature vector of the node, and d2 represents the dimension of the processed node feature vector. On the one hand, the processed node feature vector is directly fed into the pooling layer for further compression, and the compression result is the network feature vector P, which summarizes the overall information of the multi-layer complex network of the target MSM population; on the other hand, it is combined with the intervention strategy output by the Actor module, and after being compressed by another pooling layer, the overall key node vector Q is output, which summarizes the main information of all key nodes.

[0119] The present invention uses the Kronecker product of the network feature vector P and the overall key node vector Q to explore the relationship between network information and key node information. The product result is mapped to a scalar r through multiple non-linear layers (ReLu, Sigmoid), and this scalar is the score finally generated by the Critic module to evaluate the efficiency of the current intervention strategy. The Critic module will update its own score generation system (module parameters) according to the simulation results of the current intervention strategy based on the infectious disease transmission dynamics model, so that the score generated by the Critic module is closer to the simulation results of the transmission dynamics model.

[0120] The loss function of the Critic module is:

[0121] J C = [Siudtr(G, x) - Critic(G, x)] 2

[0122] In this embodiment, by cooperating with relevant dating platforms, desensitized data of online social network features is obtained, and using the data of the HIV molecular network, relatively complete offline social network data is obtained. This multi-modal network constructed based on online / offline social networks and the HIV molecular network more completely and accurately shows the network characteristics of HIV transmission among key populations. The present invention uses an adversarial generative network deep learning model based on an exponential random walk type function to integrate and construct the multi-modal network. Within the scope of a city, a multi-modal network based on online / offline social networks and molecular networks is constructed for a type of key population (such as MSM) and key node identification is carried out, and then the implementation and effect evaluation of precise intervention are guided. The present invention combines the partition nested random network search memory priority algorithm and the data-driven Markov decision process deep reinforcement learning algorithm to solve the NP-Hard technical difficulties.

[0123] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above various methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0124] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the technical field, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. Multimodal network fusion method for tracking and intervention strategies of HIV sources of infection, characterized in that Including: S1 - Using an exponential random network adversarial generative graph model to construct a multi-modal network for the transmission of the human immunodeficiency virus (HIV); S2 - Identifying key nodes in the process of virus transmission; S3 - Based on the key nodes of the multi-modal network, implementing intervention strategies and evaluating the intervention effects.

2. The method according to claim 1, wherein The specific steps of step S1 are as follows: Applying a deep learning model of an adversarial generative network with an exponential random walk function to integrate and construct a multi-modal network; Defining the network data sources for each layer, constructing corresponding graphs, and converting each graph into an adjacency matrix and a feature matrix; Initializing an initial graph vector using a graph model; for each data source, performing T iterations, where the hyperparameter T represents the number of iterations, and concatenating the final representation graph vectors obtained from each data source to form a multi-level complex network; Using a graph generation model to learn the latent representation of the fused multi-level complex network.

3. The method according to claim 1, characterized in that The specific steps of step S2 are as follows: Using the extended model of the memory SIR virus transmission dynamics equation and Monte Carlo simulation to simulate the transmission mode of the AIDS virus and different intervention measures in the constructed multi-modal network, and identifying key nodes in the process of virus transmission. Here, key nodes refer to infected or potentially infected individuals with a wide range of HIV transmission capabilities in the multi-modal network and nodes with a key transmission bridge role in the transmission process; Using a graph autoencoder (GAE) to learn the low-dimensional representation of nodes; Introducing a graph attention mechanism to enable the network to focus on the contributions of nodes at different levels to the key node identification task, so as to handle the coupling of multi-level networks; Completing the prediction of key nodes through a prediction layer; The algorithm optimizes the model by minimizing the loss function to enable it to accurately identify key nodes in a complex network.

4. The method according to claim 1, characterized in that The specific steps of step S3 are as follows: Through a data-driven Markov decision process deep learning reinforcement algorithm combined with a partitioned nested random network search and memory priority algorithm, implementing the above intervention strategies to achieve the effectiveness of precise prevention and control and precise intervention; Using a graph embedding algorithm to learn the spatial characteristics of nodes in the network, and mapping each node in the network to a low-dimensional dense vector to express the spatial characteristics of the node; After a multi-layer perceptron (MLP) receives the embedding vectors of all nodes in the network, it first maps them to a non-linear hidden layer, and then generates a score vector by a linear output layer; selects the nodes corresponding to the largest several element values in the score vector as key nodes, and gives a set of key nodes to be intervened based on the score vector; the score vector after Gumbel sampling is connected to the critic module through a straight-through estimator. The straight-through estimator uses the top-K maximum value index function to output a binary vector to represent the position of the key node during forward propagation; during backpropagation, it uses a differentiable discrete distribution sampling function to propagate the gradient to achieve gradient propagation; Using a filter-aggregator composed of multiple convolutional layers to filter out the key information in the initial embedding vector of the node and aggregate the feature information of neighbor nodes, and outputting the node feature vector M; The node feature vector M is directly transported to the pooling layer for further compression, and the compression result is the network feature P, which summarizes the overall information of the multi-layer complex network of the target behavior population; On the other hand, the node feature vector M is combined with the intervention strategy output by the actor module. After being compressed by another pooling layer, the overall vector Q of the key nodes is output. The Kronecker product of the network feature vector P and the overall vector Q of the key nodes is used to explore the relationship between the network information and the key node information. The product result is mapped to a scalar r through multiple non-linear layers. This scalar is the score finally generated by the critic module to evaluate the efficiency of the current intervention strategy; The critic module will update its own score generation system according to the simulation results of the current intervention strategy based on the epidemic transmission dynamics model.

5. A multimodal network fusion system for HIV source tracking and intervention strategies, characterized in that Including: A multi-modal network construction module for constructing a multi-modal network of human immunodeficiency virus (HIV) transmission using an exponential random network adversarial generation graph model; A key node identification module for identifying the key nodes in the virus transmission process; A network intervention evaluation module that, based on the key nodes of the multi-modal network, implements an intervention strategy and evaluates the intervention effect.

6. The system according to claim 5, wherein The construction steps of the multi-modal network construction module are: Integrate and construct the multi-modal network using an adversarial generation network deep learning model of the exponential random walk type function; Define the network data sources for each layer, construct the corresponding graph, and convert each graph into an adjacency matrix and a feature matrix; Initialize an initial graph vector using the graph model; perform T iterations for each data source, where the hyperparameter T represents the number of iterations. Concatenate the final representation graph vectors obtained from each data source to form a multi-level complex network; Use the graph generation model to learn the latent representation of the fused multi-level complex network.

7. The system according to claim 5, wherein The identification steps of the key node identification module are: Expand the model that memorizes the SIR virus transmission dynamics equation and use Monte Carlo simulation to simulate the transmission pattern of the AIDS virus and different intervention measures in the constructed multi-modal network to identify the key nodes in the virus transmission process. Here, the key nodes refer to the infected or potentially infected individuals with a wide range of HIV transmission capabilities in the multi-modal network and the nodes that play a key transmission bridge role in the transmission process; Use the graph autoencoder GAE to learn the low-dimensional representation of the nodes; Introduce the graph attention mechanism so that the network can focus on the contributions of nodes at different levels to the key node identification task to handle the coupling of the multi-level network; Complete the prediction of the key nodes through the prediction layer; The algorithm optimizes the model by minimizing the loss function so that it can accurately identify the key nodes in the complex network.

8. The system according to claim 5, wherein The network intervention evaluation module further includes: A graph embedding module for learning the spatial features of the nodes in the network using the graph embedding algorithm, mapping each node in the network to a low-dimensional dense vector to express the spatial features of the nodes; The actor module is used for the multi-layer perceptron (MLP) to receive the embedding vectors of all nodes in the network, first map them to the nonlinear hidden layer, and then generate the score vector by the linear output layer; select the nodes corresponding to the largest number of element values in the score vector as the key nodes, and give the set of key nodes that should be intervened based on the score vector; the score vector after the Gobel sampling process is connected to the critic module through the direct estimator, and the direct estimator uses the first K large value index function to output a binary vector to represent the position of the key node during forward propagation; and adopts a differentiable discrete distribution sampling function to propagate the gradient during back propagation to achieve gradient propagation; The critic module is used to use a filter-aggregator composed of multiple convolutional layers to filter out the key information in the node's initial embedding vector and aggregate the feature information of neighboring nodes, and output the node feature vector M; on the one hand, the node feature vector M is directly transmitted to the pooling layer for further compression, and the compression result is the network feature P, which summarizes the overall information of the multi-layer complex network of the target behavior population; on the other hand, the node feature vector M is combined with the intervention strategy output by the actor module, and after compression by another pooling layer, the key node overall vector Q is output; the Kronecker product of the network feature vector P and the key node overall vector Q is used to mine the relationship between network information and key node information, and the product result is mapped to a scalar r through multiple nonlinear layers, which is the score finally generated by the critic module to evaluate the efficiency of the current intervention strategy; the critic module will update its own score generation system according to the simulation results of the current intervention strategy based on the transmission dynamics model of infectious diseases.

9. The system according to claim 8, characterized in that The graph embedding algorithm of the graph embedding module is an embedded algorithm based on graph decomposition, an embedded algorithm based on random walk, or an embedded algorithm based on graph neural network.

Images