A propagation trend deduction method and system based on multi-agent cooperation

Abstract

The application relates to the technical field of artificial intelligence, and discloses a propagation trend deduction method and system based on multi-agent cooperation; the method comprises the following steps: constructing a social network environment with multiple agents, defining an image vector for the agents; generating a social adjacency matrix; the internal cognitive module of each agent is responsible for processing the internal state of the agent, and the external behavior module defines the parameterized action of the interaction between the agent and the social network environment and other agents; embedding vectors obtained based on the internal state and external information of the agent are taken as inputs of a decision model, a social context bias matrix is integrated into the attention mechanism of the decision encoder of the decision model to reflect the social local atmosphere; internal positions are generated based on the embedding vectors, and external actions are output in combination with the social environment to drive propagation simulation and trend deduction; the application has significant advantages in processing multi-modal information, depicting individual heterogeneity and realizing large-scale forward-looking prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a propagation trend deduction method and system based on multi-agent cooperation. BACKGROUND

[0002] With the popularity of social media, instant messaging tools and other network platforms, the speed, breadth and influence of information dissemination have never been stronger. Effective understanding and prediction of the propagation trend of specific information (such as commercial advertisements, emergencies, etc.) has important application value for enterprise decision-making and event monitoring.

[0003] Existing propagation trend analysis techniques mostly use statistical models or simplified dynamic models (such as SIR model, IC model, etc.) based on historical data. However, there are generally the following problems: first, the model assumptions are too idealized, making it difficult to depict the heterogeneity and complexity of individual behavior in real society; second, it is difficult to effectively handle modern information forms containing text, images and other modalities; third, it is mostly a post-analysis tool, with limited prediction ability for new events, making it difficult to conduct forward-looking sand table deduction. This is because information propagation is essentially a macroscopic phenomenon emerging from the interaction of a large number of heterogeneous individuals, and traditional methods are difficult to capture the complex mapping relationship from micro individual decision-making to macro group behavior. The deviation of the deduction result will lead to decision-making errors and reduce its application value in real scenarios. SUMMARY

[0004] To solve the above technical problems, the present application provides a propagation trend deduction method and system based on multi-agent cooperation to improve the accuracy and realism of the deduction result. First, the present application constructs a virtual society composed of a large number of agents with independent personality and decision-making ability, and uses a small-world network to construct its social topology to simulate the real social structure. Second, a cognitive architecture capable of processing multi-modal information and making complex decisions is designed for each agent, which integrates the agent's own internal attributes and the influence of the external social environment. Finally, by simulating the interaction behavior of agents around a specific information, the emerging macroscopic propagation law is observed and analyzed, thereby realizing accurate deduction of the propagation trend.

[0005] To solve the above technical problems, the present application adopts the following technical solutions:

[0006] In a first aspect, the present application provides a propagation trend deduction method based on multi-agent cooperation, comprising:

[0007] The individual in the social network is simulated by an agent, a social network environment with multiple agents is constructed, and an image vector composed of demographic characteristics, cognitive characteristics and behavior characteristics is defined for each agent; a small-world network generation algorithm is used to construct the social adjacency matrix between agents to simulate the network topology of the social network environment;

[0008] Each agent has an internal cognitive module and an external behavior module; the internal cognitive module is responsible for processing the internal state of the agent, including a memory system and a planning reflection unit, and is provided with a basic behavior memory bank for storing the pre-computed encoding results of basic behaviors for direct retrieval when the agent performs the basic behaviors; the external behavior module defines the parameterized actions of the agent interacting with the social network environment and other agents;

[0009] The embedding vector obtained based on the internal state of the agent and the external information received from the social network environment is used as the input of the decision model, and a social context bias matrix is integrated into the attention mechanism of the decision encoder of the decision model to reflect the social local atmosphere; a two-stage decision process is set, first an internal position is generated based on the embedding vector, and then an external action is output in combination with the internal position to drive the propagation simulation and trend deduction.

[0010] In one embodiment, the image vector composed of demographic characteristics, cognitive characteristics and behavior characteristics is defined for each agent, specifically including:

[0011] ;

[0012] wherein, the image vector, the demographic characteristic vector includes age, occupation and education level; the cognitive characteristic vector includes initial trust level of information in different fields, information critical thinking level and interest preference distribution; the behavior characteristic vector includes social activity, daily information contact frequency and sharing tendency.

[0013] In one embodiment, the small-world network generation algorithm is used to construct the social adjacency matrix between agents to simulate the network topology of the social network environment, specifically including:

[0014] The Watts-Strogatz small-world network generation algorithm is used to construct the social adjacency matrix between agents : arrange agent nodes on a ring lattice, each agent node is connected to the nearest neighbor nodes to form a regular network; traverse each edge in the regular network with a reconnection probability Randomly reconnect one end of the edge to another node in the regular network.

[0015] In one embodiment, the internal cognitive module is responsible for processing the internal state of the agent, including the memory system and the planning reflection unit, and is provided with a basic behavior memory bank for storing the pre-computed encoding results of the basic behaviors for direct retrieval and use when the agent executes the basic behaviors, specifically including:

[0016] The memory system has a fast memory mechanism, and a global basic behavior memory bank is set up to store the pre-computed encoding results of the basic behaviors in the form of key-value pairs:

[0017] ;

[0018] wherein, is the type of the jth basic behavior as the key; and are the embedding vector and the importance score of the jth basic behavior as the value, respectively.

[0019] In one embodiment, the embedding vectors based on the internal state of the agent and the external information received from the social network environment are used as inputs of the decision model, specifically including:

[0020] At the jth time step of the propagation simulation of the decision model, the input data of the decision model of the ith agent includes the internal state and the external information of the ith agent:

[0021] ;

[0022] denotes a classification label, denotes a separator label;

[0023] The input data is mapped to a high-dimensional vector by the embedding vector layer; the state embedding vector obtained after mapping the internal state , the multi-modal information embedding vector obtained after mapping the external information , and the static portrait embedding vector obtained after mapping the portrait vector are added element by element to obtain the final embedding vector :

[0024] ​​ ;

[0025] wherein, injecting the fundamental personality and preference of the agent into the decision model.

[0026] In one of the embodiments, the social context bias matrix is incorporated into the attention mechanism of the decision encoder of the decision model to reflect the social local atmosphere, specifically comprising:

[0027] The embedding vector obtained based on the internal state and external information of the agent is input into the decision encoder based on the multi-layer Transformer structure, and the social context bias matrix is introduced in the self-attention calculation process of the decision encoder ; The embedding vector is generated by aggregating the internal state or external action of the neighbor nodes of the current agent in the social adjacency matrix at the moment , reflecting the social local atmosphere of the current agent.

[0028] In one of the embodiments, the self-attention calculation process of the decision encoder is:

[0029] ;

[0030] wherein, are the query matrix, the key matrix and the value matrix respectively, is the attention mask, denotes the softmax function, denotes the transpose, denotes the square root function, denotes the dimension of the key matrix, denotes the final output of the attention mechanism calculation.

[0031] In one of the embodiments, a two-stage decision process is set, i.e., first generating an internal stance based on the embedding vector, and then outputting an external action in combination with the internal stance to drive the propagation simulation and trend deduction, specifically comprising:

[0032] Internal stance generation: the embedding vector is input into a classification head to output the internal stance of the agent ;

[0033] The output feature of the decision encoder is input into another classification head, and the internal stance is taken as an additional input to output the external action .

[0034] In one of the embodiments, a multi-task loss function is used to train the decision model:​

[0035] ;

[0036] wherein, is an internal position loss, is an external action loss, and all employ cross-entropy loss, is a hyperparameter.

[0037] In a second aspect, the present application provides a computer system comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method of any one of the embodiments of the first aspect when executing the computer program.

[0038] Compared with the prior art, the present application has the beneficial technical effects that:

[0039] The propagation trend deduction method based on multi-agent cooperation provided by the present application includes three parts of social network environment construction, agent core architecture, and propagation simulation and deduction. In order to make the deduced propagation trend closer to the real world and improve the prediction accuracy, the present application constructs a heterogeneous and multi-dimensional portrait for each agent, and uses a small-world network to simulate the real social structure; in order to improve the simulation efficiency and scale, a fast memory mechanism is designed; in order to ensure the rationality of the decision-making behavior, a self-consistent decision-making process is proposed by introducing social context and two-stage reasoning. The propagation trend deduction technology of the present application can more effectively capture the complex process from micro individual interaction to macro phenomenon emergence, has significant advantages in processing multi-modal information, depicting individual heterogeneity and realizing large-scale forward-looking prediction, and has good application prospects in event monitoring, market analysis, public health and safety warning scenes. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 is a method flowchart in the embodiments of the present application. DETAILED DESCRIPTION

[0041] A preferred embodiment of the present application will be described in detail below with reference to the accompanying drawings.

[0042] As Figure 1 shown, the propagation trend deduction method based on multi-agent cooperation in the present application includes the following steps:

[0043] S1, simulate individuals in a social network by an agent, construct a social network environment with multiple agents, define a portrait vector composed of demographic characteristics, cognitive characteristics and behavior characteristics for each agent, and use a small-world network generation algorithm to construct a social adjacency matrix between agents to simulate the network topology of the social network environment;

[0044] S2, each agent has an internal cognitive module and an external behavior module; the internal cognitive module is responsible for processing the agent's internal state, including a memory system and a planning and reflection unit, and has a basic behavior memory bank to store the pre-computed encoding results of basic behaviors, so that they can be directly retrieved and used when the agent performs basic behaviors; the external behavior module defines the parameterized actions of the agent in interacting with the social network environment and other agents.

[0045] S3 uses the embedding vectors obtained from the agent's internal state and external information as input to the decision model. It incorporates a social context bias matrix into the attention mechanism of the decision encoder to reflect the local social atmosphere. It sets up a two-stage decision process: first, it generates internal stances based on the embedding vectors, and then combines the internal stances to output external actions to drive propagation simulation and trend inference.

[0046] Propagation trend prediction technology is generally defined as follows: given a set of agents in a social network environment and a source message as input data, a computer algorithm automatically predicts the propagation trend, key nodes, and lifecycle of the message in the network based on the collaborative interactions between the agents. Propagation trend prediction technology aims to solve the following problem: given a set of agents in a social network environment and a source message as input data, a computer algorithm automatically predicts the propagation trend, key nodes, and lifecycle of the message in the network based on the collaborative interactions between the agents. A set of agents, each defined by a multi-dimensional profile vector, and these agents are linked by a social adjacency matrix. Interconnected. When a source message... When injected into this social network environment, the algorithm needs to be based on its internal state. and external information Simulates the collaborative interaction between intelligent agents, automatically deduces the propagation trend of information in the network, such as identifying key influencing nodes and predicting their propagation lifecycle, and finally outputs an analysis report containing indicators such as propagation rate and scope of influence.

[0047] This invention comprises three core parts: the construction of a multi-dimensional intelligent agent social network environment; the core architecture of a multi-modal intelligent agent; and the propagation simulation and trend inference model. Each part will be described in detail below.

[0048] 1. Construction of a multidimensional intelligent agent social network environment.

[0049] This section aims to construct a highly realistic virtual social network environment that serves as a carrier of information dissemination; its authenticity is the foundation for the accuracy of subsequent inferences.

[0050] (1) Construction of agent profile vectors:

[0051] To ensure the heterogeneity and realism of agent behavior, the first step is to define the heterogeneity of each agent in the system. Define a multi-dimensional image vector The portrait vector is composed of multiple quantifiable sub-vectors and is initialized by sampling from real-world social survey data. The formula is as follows:

[0052] ;

[0053] wherein, is a demographic feature vector, including age, occupation, education level, etc. is a cognitive feature vector, including initial trust in different fields of information, information critical thinking level, interest preference distribution, etc. is a behavior feature vector, including social activity, average daily information contact frequency, sharing tendency, etc.

[0054] (2) Social network topology construction:

[0055] In order to simulate the characteristics of close community structure and weak connections across communities coexisting in real society, the Watts-Strogatz small-world network generation algorithm is used to construct the social adjacency matrix between agents . The algorithm flow is as follows: first, arrange agent nodes on a ring lattice, each node is connected to its nearest neighbor nodes, forming a regular network. Then, traverse each edge in the regular network, and randomly reconnect one end of the edge to another node in the regular network with a small reconnection probability . This process can be represented as:

[0056] ;

[0057] wherein, represents the Watts-Strogatz small-world network generation algorithm. The constructed social adjacency matrix has a high clustering coefficient and a low average path length , which can truly reflect the propagation dynamics characteristics of information in real social networks.

[0058] 2. Multi-modal agent core architecture.

[0059] Each agent is designed as an autonomous decision-making entity based on a multi-modal large model. Its core architecture consists of an internal cognitive module and an external behavior module, which work together to drive the behavior of the agent.

[0060] (1) Internal cognitive module: This module is responsible for processing the internal state of the agent, and the core is the memory system and the planning reflection unit.

[0061] Memory system The human cognitive model is imitated and divided into sensor memory, short-term memory and long-term memory. In order to solve the problem of large-scale simulation computing overhead, a fast memory mechanism is designed. The core of the mechanism is a global basic behavior memory bank The bank stores the pre-computed encoding results of common low-complexity behaviors (such as "browsing information flow", "liking") in the form of key-value pairs.

[0062] .

[0063] Among them, is the basic behavior type, and are the corresponding embedding vectors and importance scores respectively. When the agent intends to perform a basic behavior in , the system directly retrieves the result from the library to update its short-term memory, without calling the complete multi-modal decision model, thereby significantly reducing the token consumption and time cost. For complex behaviors (such as "publishing long comments") not defined in , the complete decision model is called.

[0064] (2) External behavior module:

[0065] This module defines a series of actions that the agent can interact with the environment and other agents. The actions are parameterized to support rich interactions. For example, the forwarding action can be represented as , where is the information content, is the sharing target (such as a friend, group or public release), is the additional comment text.

[0066] 3. Propagation simulation and trend inference model.

[0067] This module is the core of the method, responsible for driving the entire simulation process and performing the final trend analysis.

[0068] (1) Input and embedding vector layer:

[0069] At each time step of the simulation, for the agent that receives external information , the input of its decision model is composed of its internal state and external information . The formula is as follows:

[0070] ;

[0071] denotes a classification token, usually placed at the beginning of the input sequence; denotes a separator token, used to explicitly separate internal states and external information . External information is the information content that the agent receives from the social network environment, including source information and other multi-modal data such as text, images, etc. propagated by other agents.

[0072] The input data passes through an embedding vector layer, which maps each part to a high-dimensional vector. The final embedding vector is the sum of the state embedding vector , the multi-modal information embedding vector , and the static portrait embedding vector . .

[0073] ;

[0074] where, is used to inject the agent's fundamental personality and preferences into the model, ensuring personalized decision-making; a multi-modal encoder such as CLIP is used as the embedding vector layer to uniformly encode the text and images in the external information to obtain .

[0075] (2) Decision encoder and self-consistent decision-making process:

[0076] The embedding vector is fed into a decision encoder with a multi-layer Transformer structure. To enable the agent's decision-making to consider its social environment, the invention introduces a social context bias matrix during self-attention calculation. By aggregating the states or behaviors of the neighbor nodes of the current agent in the social adjacency matrix at time , the social local atmosphere of the current agent is reflected. The modified self-attention calculation process is shown as follows:

[0077] ;

[0078] where, are the query matrix, key matrix, and value matrix, respectively, is the attention mask. In this way, social influence is directly integrated into the agent's deep understanding of information.

[0079] The output features of the encoder are then used in a two-stage self-consistent decision-making process:

[0080] Phase 1 (Internal stance generation): The final embedding vector input to the decision encoder is fed into a classification head, outputting an internal stance representing the agent’s “first reaction” (such as: support, oppose, neutral, skeptical).

[0081] Phase 2 (External action decision): The encoder’s full output (including social context) is fed into another classification head, with the previous phase’s output as additional input, finally outputting a specific external action .

[0082] (3) Model training and trend inference:

[0083] During training, a multi-task loss function is used to simultaneously optimize the accuracy of internal stance prediction and external action prediction. The formula is as follows:

[0084] ;

[0085] where and are cross-entropy losses, is a hyperparameter.

[0086] During prediction (inference), source information is injected from one or more source nodes , and iterated according to time steps . All agent action logs are recorded in real time . After the simulation ends, process , calculate and output a series of propagation trend indicators, including: propagation rate curve, impact range (node coverage), peak time, key propagation path, and a list of identified “super spreaders” nodes, ultimately forming a complete trend inference report.

[0087] The terminology used herein is for the purpose of describing specific embodiments only and is not intended to limit the present disclosure. As used herein, the terms “comprise”, “comprise”, and the like indicate the presence of the described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0088] It should be understood that although the steps in the flowcharts of the drawings are shown in sequential order, such that each step depends on the previous step in order to occur, the steps are not necessarily performed in this order. Unless the context clearly indicates otherwise, there is no strict order requirement for the performance of the steps, and the steps can be performed in other orders. Moreover, at least some of the steps in the flowcharts of the drawings can include multiple steps or stages, which are not necessarily performed at the same time, and which can be performed in different orders, in parallel or in alternating fashion with other steps or stages of other steps.

[0089] In one embodiment, a computer system is provided, which can be a server. The computer system comprises a processor, a memory and a network interface connected by a system bus. The processor of the computer system is configured to provide computing and control capabilities. The memory of the computer system comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer system is configured to store data used in the above method. The network interface of the computer system is configured to communicate with an external terminal through a network connection. The computer program is configured to be executed by the processor to implement the above method.

[0090] The technical features of the above embodiments can be combined in any manner. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not contradict each other, they should be considered as falling within the scope of the present disclosure.

[0091] It will be apparent to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and that the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Thus, the embodiments are to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the above description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein and their scope is to be accorded the broadest interpretation so as to encompass all meaningful novel techniques compatible with the basic features of the application. The reference numerals in the claims should not be construed as limiting the scope of the claims by their inclusion. Any reference signs in the claims should not be construed as limiting the scope of the claims by their inclusion. The use of any and all examples, or exemplary language (e.g., "such as" and "preferably"), in the specification is intended to convey that the feature so preceded is one example, and not the only example, of a feature that can be employed in accordance with specific embodiments of the application and thus the scope of the application should not be limited by such examples.

[0092] Furthermore, it should be understood that although the specification is described in terms of embodiments, not every embodiment includes only one independent technical solution, and the specification is described in this way only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that those skilled in the art can understand.

Claims

1. A propagation trend deduction method based on multi-agent cooperation, characterized in that, The application relates to a social network simulation method and system based on artificial intelligence. The method comprises the following steps: An individual in a social network is simulated by an intelligent agent, a social network environment with multiple intelligent agents is constructed, an image vector composed of a population characteristic, a cognitive characteristic and a behavior characteristic is defined for each intelligent agent, a social adjacent matrix between the intelligent agents is constructed by using a small-world network generation algorithm to simulate the network topology of the social network environment, each intelligent agent has an internal cognitive module and an external behavior module, the internal cognitive module is responsible for processing the internal state of the intelligent agent, including a memory system and a planning reflection unit, and a basic behavior memory bank is arranged for storing pre-computed encoding results of basic behaviors for direct retrieval when the intelligent agent executes the basic behaviors, the external behavior module defines parameterized actions of the intelligent agent interacting with the social network environment and other intelligent agents, the internal state includes a sensor memory, a short-term memory and a long-term memory, the image vector defined for each intelligent agent specifically comprises the following steps: The internal cognitive module is responsible for processing the internal state of the intelligent agent, including a memory system and a planning reflection unit, and a basic behavior memory bank is arranged for storing pre-computed encoding results of basic behaviors for direct retrieval when the intelligent agent executes the basic behaviors, the external behavior module defines parameterized actions of the intelligent agent interacting with the social network environment and other intelligent agents, the internal state includes a sensor memory, a short-term memory and a long-term memory, the image vector defined for each intelligent agent specifically comprises the following steps: Embedded vectors, obtained from the agent's internal state and external information received from the social network environment, are used as inputs to the decision model. A social context bias matrix is ​​incorporated into the attention mechanism of the decision encoder to reflect the local social atmosphere: the embedded vectors are input into the decision encoder based on a multi-layer Transformer structure, and the social context bias matrix is ​​introduced during the self-attention calculation process of the decision encoder. ; By aggregating the current agents in the social adjacency matrix Neighbor nodes in The generation of internal states or external actions at any given moment. This reflects the current social atmosphere of the intelligent agent; the self-attention calculation process of the decision encoder is as follows: ; These are the query matrix, key matrix, and value matrix, respectively. For attention masking, This represents the softmax function. For transpose, It is a square root function. Let be the dimension of the key matrix. The final output of the attention mechanism is calculated; external information is the information content that the agent receives from the social network environment. A two-stage decision-making process is set, first generating an internal stance based on the embedding vector, and then combining the internal stance to output external actions to drive the propagation simulation and trend deduction, specifically including: internal stance generation: send the embedding vector into a classification head to output the internal stance of the agent ​ The output features of the decision encoder Send to another sorting head, and use an internal position. As additional input, output external action .

2. The method of claim 1, wherein, The internal cognitive module is responsible for processing the internal state of the intelligent agent, including a memory system and a planning reflection unit, and a basic behavior memory bank is arranged for storing pre-computed encoding results of basic behaviors for direct retrieval when the intelligent agent executes the basic behaviors, the external behavior module defines parameterized actions of the intelligent agent interacting with the social network environment and other intelligent agents, the internal state includes a sensor memory, a short-term memory and a long-term memory, the image vector defined for each intelligent agent specifically comprises the following steps: ； wherein, is an image vector, is a demographic feature vector, including age, occupation, and education level; is a cognitive feature vector, including initial trust level on information in different domains, information critical thinking level, and interest preference distribution; is a behavioral feature vector, including social activity level, average daily information exposure frequency, and sharing tendency.

3. The method of claim 1, wherein the method comprises: The embedded vector obtained based on the internal state of the intelligent agent and external information received from the social network environment is taken as the input of the decision model, and the method specifically comprises the following steps: A Watts-Strogatz small-world network generation algorithm is used to construct a social adjacency matrix among the agents : One ring lattice is used to arrange the nodes of the agents, and each node of an agent is connected to the nearest nodes to form a regular network; each edge in the regular network is traversed to randomly reconnect one end of the edge to another node in the regular network with a reconnection probability .

4. The propagation trend deduction method based on multi-agent cooperation according to claim 1, characterized in that, The processor executes the computer program to realize the steps of the method in any one of claims 1 to 6. The memory system has a fast memory mechanism, and a global basic behavior memory library is set The pre-computed encoding result of the basic behavior is stored in the form of key-value pairs ； wherein, is a type of the jth base behavior as a key; and are an embedding vector and an importance score of the jth base behavior, respectively, as values.

5. The method of claim 1, wherein the method comprises: ​ At the first time step of the propagation simulation by the decision model, the input data of the decision model of the first agent comprises the internal state of the first agent and external information : ； a classification marker, a separator marker; Input data Each part is mapped to a high-dimensional vector through the embedding vector layer; the internal state State embedding vector obtained after mapping , external information Multimodal information embedding vector obtained after mapping and portrait vector Static portrait embedding vector obtained after mapping Element-wise addition, get the final embedding vector : ； wherein, for injecting the fundamental personality and preferences of an agent into a decision model.

6. The method of claim 1, wherein the method is based on multi-agent collaboration. Using a multi-task loss function Training the decision model: ； where, is the internal position loss, is the external action loss, and all use cross-entropy loss, is a hyperparameter.

7. A computer system comprising a memory and a processor, said memory storing a computer program, characterized in that, ​

Images