Marketing risk rehearsal method and system based on dynamic simulation

Abstract

The invention discloses a marketing risk rehearsal method and system based on dynamic simulation, and the method comprises the steps: building a multi-agent interaction environment model based on historical market environment data and multi-source social network data, and configuring a parameterized decision model for a competitor entity, a user circle layer entity and a media opinion entity; through a Monte Carlo simulation and multi-agent reinforcement learning hybrid deduction engine, performing multiple times of dynamic game simulation on an initial action sequence of a local entity in a multi-agent interaction environment model obtained through marketing scheme data conversion, and generating simulation trajectory data; and aggregating and analyzing a plurality of pieces of simulated trajectory data to identify and quantify a competitive vulnerability mode and a word-of-mouth propagation detonation point mode, generating a rehearsal insight report, and iteratively optimizing a marketing scheme through anti-factual reasoning according to the rehearsal insight report until a strategy robustness condition is met. According to the method, prospective quantitative evaluation and active optimization of potential linkage risks of a marketing scheme in a dynamic game environment are realized.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a marketing risk prediction method and system based on dynamic simulation. Background Technology

[0002] In a highly competitive and rapidly evolving digital environment, risk assessment and strategy simulation before launching major marketing campaigns have become crucial for businesses to improve the scientific nature of decision-making and mitigate potential losses. Traditional risk assessments often rely on statistical analysis of historical data, expert judgment, or static model predictions based on single market assumptions. However, with advancements in computing technology, existing techniques can now construct more complex simulation systems. For example, by integrating market trend analysis, consumer behavior modeling, and Monte Carlo simulations, it is possible to simulate the potential effects of marketing activities under different market conditions and optimize resource allocation, thereby providing data support for strategy development.

[0003] However, existing simulation methods in this regard still primarily focus on predicting the macro-level responses of the "market as a whole" or "consumer groups." Their simulation environments are typically pre-defined and relatively static, making it difficult to realistically depict the market as a complex dynamic system composed of multiple stakeholders. In the real market, the launch of a marketing strategy triggers a series of chain reactions, including targeted responses from competitors, the fermentation and variation of word-of-mouth on user social networks, and the amplification or shift in media opinion. Because existing technologies lack the ability to model and extrapolate this dynamic game process based on strategic interactions among multiple stakeholders, they struggle to systematically identify and quantify potential structural risks in marketing plans beforehand, such as strategic weaknesses exposed by competitive confrontation. This results in assessments and optimization recommendations generated based on existing technologies lacking foresight and strategic robustness when facing real and complex market games. Summary of the Invention

[0004] In view of the above problems, the present invention provides a marketing risk prediction method and system based on dynamic simulation. By constructing and simulating the dynamic game interaction process of multiple market players, it can realize the forward-looking identification and quantitative assessment of potential chain risks and structural weaknesses of marketing plans.

[0005] To achieve the above objectives, in a first aspect, this application provides a marketing risk prediction method based on dynamic simulation, comprising:

[0006] Receive marketing campaign data to be evaluated, historical market environment data, and multi-source social network data;

[0007] Based on historical market environment data and multi-source social network data, a multi-agent interaction environment model is constructed, which includes competitor entities, user circle entities, and media opinion entities. Parameterized decision models are configured for competitor entities, user circle entities, and media opinion entities respectively.

[0008] The marketing plan data to be evaluated is transformed into the initial action sequence of the party's entity in the multi-agent interaction environment model. Based on the multi-agent interaction environment model and the parameterized decision model, multiple dynamic game simulations are executed in parallel through a hybrid inference engine of Monte Carlo simulation and multi-agent reinforcement learning to generate multiple simulated trajectory data containing the evolution of the environment state and the action sequence of the entity.

[0009] By aggregating and analyzing multiple simulated trajectory data, we can identify and quantify the competitive vulnerability patterns and word-of-mouth tipping point patterns triggered by the marketing plan data to be evaluated in the dynamic game simulation, and generate a pre-simulation insight report containing competitive vulnerability maps and word-of-mouth tipping point probability cloud maps.

[0010] Based on the pre-simulation insight report, the key parameters in the marketing plan data to be evaluated are iteratively optimized through counterfactual reasoning to generate optimized marketing plan data. The optimized marketing plan data is then re-input into the hybrid inference engine for dynamic game simulation and evaluation until the predetermined strategy robustness conditions are met, and the final risk assessment and strategy optimization suggestions are output.

[0011] In some embodiments, based on historical market environment data and multi-source social network data, a multi-agent interaction environment model is constructed, including competitor entities, user circle entities, and media opinion entities, including:

[0012] Entity relationships are extracted and quantified from historical market environment data to generate a competitive relationship intensity matrix and market resource constraint parameters;

[0013] Graph structure modeling and community discovery are performed on multi-source social network data to generate user social network graphs and cross-platform influence network graphs. In the user social network graphs, nodes are labeled with multiple user circle labels divided by clustering algorithms.

[0014] Based on the competitive relationship intensity matrix, a competitive ecosystem graph is constructed with competitor entities as nodes, and a resource constraint model including budget, production capacity and channel scheduling is configured for competitor entities.

[0015] Based on user social network graphs and multiple user circle tags, a group behavior model is constructed with user circle entities as agents. The group behavior model integrates social network propagation dynamics and behavioral economics theory to simulate the process of individual decision-making being influenced by social factors and cognitive biases.

[0016] Based on a cross-platform influence network graph, an influence propagation model for media opinion entities is constructed. The influence propagation model encodes the characteristic differences of different information dissemination platforms and the position preference parameters of media opinion entities.

[0017] By coupling the competitive ecosystem map, the group behavior model, and the influence propagation model, a multi-agent interaction environment model is formed. This model defines the state perception and interaction rules between competitor entities, user circle entities, and media opinion entities.

[0018] In some embodiments, entity relationships are extracted and quantified from historical market environment data to generate a competitive relationship strength matrix and market resource constraint parameters, including:

[0019] Extract historical marketing campaign sequences, historical market share time-series data, and historical brand co-occurrence data from historical market environment data;

[0020] Based on historical marketing campaign sequences, the mutual response strength of marketing actions between different brand entities is calculated using a causal inference algorithm, serving as the primary quantitative indicator of competitive relationship strength.

[0021] Based on historical market share time series data, the negative correlation between changes in market share of different brands is calculated through time series correlation analysis, which serves as a second quantitative indicator of the intensity of competitive relationship.

[0022] Based on historical brand co-occurrence data, the intensity of opposing sentiment associations between different brand entities in the public opinion field is calculated through sentiment analysis in natural language processing, serving as a third quantitative indicator of the intensity of competitive relationships.

[0023] The first, second, and third quantitative indicators of competitive relationship strength are combined and weighted to generate a competitive relationship strength matrix.

[0024] By analyzing historical resource input data of each brand entity from historical market environment data, market resource constraint parameters are obtained through statistical regression fitting. These parameters characterize the constraint relationship between resource input and marketing effectiveness output.

[0025] In some embodiments, parameterized decision models are configured for competitor entities, user community entities, and media opinion entities, including:

[0026] Configure a first parameterized decision model for the competitor entity. The first parameterized decision model includes a policy network built based on a deep reinforcement learning algorithm. The input of the policy network is an environmental state vector containing the entity's own state, the competitor's state, and the overall market state. The output of the policy network is the probability distribution of the competitor entity's actions or deterministic action values ​​in the action space. The objective function of the first parameterized decision model is to maximize the long-term cumulative reward. The long-term cumulative reward is calculated by the changes in market share, changes in brand equity, and action costs.

[0027] A second parameterized decision-making model is configured for user circle entities. The second parameterized decision-making model includes a decision function constructed based on the intelligent agent modeling framework and behavioral economics theory. The input of the decision function is an individual attribute vector, a local social network state vector, and a received information stimulus vector. The decision function simulates individual decision-making by calculating the comprehensive utility value and introducing random noise. The comprehensive utility value integrates personal preference utility, social conformity utility, and risk perception utility under the prospect theory framework.

[0028] A third parametric decision-making model is configured for media opinion entities. The third parametric decision-making model includes a content selection function built based on the utility maximization theory. The input of the content selection function is a vector of candidate content attributes, a vector of audience matching degree, and a vector of potential commercial incentives. The content selection function calculates the expected influence return and commercial return by weighting, and selects the content with the highest comprehensive return to perform the dissemination action.

[0029] In some embodiments, graph structure modeling and community discovery are performed on multi-source social network data to generate user social network graphs and cross-platform influence network graphs, including:

[0030] Clean and align user interaction behavior data and content association data in multi-source social network data to identify unique user nodes and content nodes;

[0031] Based on user interaction behavior data, an initial social network graph is constructed with user nodes as vertices and follow relationships, forwarding relationships, or co-occurrence relationships as edges;

[0032] Applying a graph embedding algorithm to the initial social network graph yields a low-dimensional vector representation of user nodes;

[0033] Based on the low-dimensional vector representation of user nodes, a community detection algorithm is used to cluster user nodes, the clustering results are labeled as multiple user circle labels, and the multiple user circle labels are mapped back to the initial social network graph to generate a user social network graph with circle labels.

[0034] By integrating cross-platform account association data and content dissemination path data from multi-source social network data, media opinion leader nodes and their influence range on different information dissemination platforms can be identified.

[0035] Based on content dissemination path data, a bipartite graph is constructed with media opinion leader nodes and platforms as vertices and cross-platform content migration relationships as edges;

[0036] Projecting and weighting the bipartite graph generates a cross-platform influence network graph. The edge weights in the cross-platform influence network graph represent the probability or intensity of information migration between platforms.

[0037] In some embodiments, the marketing campaign data to be evaluated is transformed into an initial sequence of actions for the local entity in a multi-agent interaction environment model, including:

[0038] Analyze the marketing plan data to be evaluated, and extract the core marketing action elements, resource allocation plans, and execution timelines contained in the marketing plan data;

[0039] The core marketing action elements are mapped to standard action codes in the action space that the entity can execute in the multi-agent interaction environment model. The standard action codes include promotion action codes, advertising action codes, channel strategy action codes, and public opinion guidance action codes.

[0040] Based on the resource allocation plan, allocate corresponding resource quantification parameters to the mapped standard action codes;

[0041] Based on the execution timeline, the standard action codes for allocating resource quantification parameters are arranged and combined in chronological order to generate an initial action sequence.

[0042] In some embodiments, based on a multi-agent interaction environment model and a parameterized decision model, a hybrid inference engine combining Monte Carlo simulation and multi-agent reinforcement learning is used to execute multiple dynamic game simulations in parallel, generating multiple simulated trajectory data containing environmental state evolution and entity action sequences, including:

[0043] Initialize the hybrid simulation engine, and set the total number of parallel simulation instances and the maximum simulation time step for each simulation instance;

[0044] Generate an independent random seed for each parallel simulation instance, and initialize the initial state of other entities in the multi-agent interaction environment model, excluding the agent's own entity, based on the independent random seed;

[0045] Within each time step of each simulation instance, perform the following steps:

[0046] Invoke the first parameterized decision model of the competitor entity and generate competitive actions of the competitor entity based on the current environmental state;

[0047] The second parameterized decision model of the user circle entity is invoked to generate the group behavior actions of the user circle entity based on the current social network state and the intensity of information stimuli.

[0048] The third parameterized decision-making model of the media opinion entity is invoked to generate the dissemination actions of the media opinion entity based on the current content ecosystem and commercial incentives;

[0049] The actions of our own entity, competitive actions, group behavior actions, and propagation actions are input into the multi-agent interaction environment model. Based on the state transition rules defined in the multi-agent interaction environment model, the environment state of the next time step is calculated and updated.

[0050] During the environment state update process, check whether the preset key event triggering conditions are met. If they are met, activate the corresponding key event rule and modify the subsequent state transition rules.

[0051] Record the environmental state, the actions of all entities, and the key events triggered at the current time step to form a trajectory segment at the current time step;

[0052] When a simulation instance reaches the maximum time step or meets the simulation termination condition, the simulation instance ends, and all trajectory segments connected in chronological order in the simulation instance are summarized into complete simulation trajectory data.

[0053] Collect complete simulation trajectory data generated by all parallel simulation instances to form multiple simulation trajectory data containing environmental state evolution and entity action sequences.

[0054] In some embodiments, multiple simulated trajectory data are aggregated and analyzed to identify and quantify the competitive vulnerability patterns and word-of-mouth tipping point patterns triggered by the marketing campaign data to be evaluated in dynamic game simulation, generating a pre-simulation insight report containing a competitive vulnerability map and a word-of-mouth tipping point probability cloud map, including:

[0055] Filter through multiple simulated trajectory data and extract negative outcome trajectory data from all simulated trajectory data, such as the final decline in market share or damage to brand health of our entity;

[0056] By performing time-series pattern mining on negative outcome trajectory data, we can pinpoint frequently co-occurring action pairs between marketing actions initiated by our entity and adversarial actions taken by competitors within the time window before a significant decline in market share.

[0057] Calculate the support and confidence of each action pair combination in all negative result trajectory data, and identify action pair combinations whose support and confidence both exceed a preset threshold as competitive vulnerability patterns.

[0058] Based on the competitive vulnerability pattern, a competitive vulnerability map is generated by visually annotating the competitive ecosystem map in the multi-agent interaction environment model. The competitive vulnerability map shows the typical attack paths from one's own actions to the competitor's counterattack and their frequency of occurrence.

[0059] Filter through multiple simulated trajectory data and extract the positive result trajectory data from all simulated trajectory data where the total sound volume increases beyond the preset burst threshold;

[0060] Feature backtracking analysis is performed on the positive result trajectory data to extract common initial environmental state features and early key event features that existed before the sound volume explosion, forming a potential tipping point feature set.

[0061] Calculate the conditional probability of a sound burst event occurring given that each feature combination in the potential detonation point feature set appears in the entire simulated trajectory dataset.

[0062] Feature combinations whose conditional probability exceeds a preset probability threshold are identified as word-of-mouth marketing tipping point patterns, and leverage conditional parameters corresponding to word-of-mouth marketing tipping point patterns are calculated.

[0063] Based on the word-of-mouth propagation tipping point model and its conditional probability and leverage conditional parameters, a word-of-mouth tipping point probability cloud map is generated by visual rendering on the user social network graph and cross-platform influence network graph in the multi-agent interaction environment model.

[0064] Integrate competitive vulnerability maps and word-of-mouth tipping point probability cloud maps, and add statistical summaries and key findings descriptions to generate a pre-launch insight report.

[0065] In some embodiments, based on the pre-visualization insight report, key parameters in the marketing campaign data to be evaluated are iteratively optimized using counterfactual reasoning to generate optimized marketing campaign data, including:

[0066] Analyze the pre-launch insight report to identify high-frequency and high-risk competitive vulnerability patterns from the competitive vulnerability map, and identify high-probability and high-leverage potential word-of-mouth dissemination tipping point patterns from the word-of-mouth tipping point probability cloud map.

[0067] For the identified competitive vulnerability patterns, the adjustable marketing action parameters associated with them in the marketing program data to be evaluated are identified as the first type of key parameters;

[0068] For the identified word-of-mouth marketing tipping point patterns, the adjustable marketing action parameters or resource allocation parameters associated with them in the marketing plan data to be evaluated are identified as the second type of key parameters.

[0069] Construct a counterfactual reasoning experimental framework that keeps the random seed of the hybrid inference engine and all other input conditions except for the first and second type of key parameters unchanged.

[0070] Within the framework of counterfactual reasoning experiments, the values ​​of the first and second types of key parameters are systematically adjusted to generate multiple candidate marketing plan data variants.

[0071] Multiple candidate marketing plan data variants are converted into corresponding initial action sequences and then sequentially input into the hybrid inference engine for dynamic game simulation, generating multiple sets of candidate simulation trajectory data;

[0072] The simulation trajectory data of multiple candidate projects are evaluated, and the strategy robustness index corresponding to each candidate marketing plan data variant is calculated. The strategy robustness index comprehensively evaluates the expected market share, the probability of risk exposure, and the potential for word-of-mouth explosion.

[0073] Select the candidate marketing campaign data variant with the best strategy robustness index and identify it as the optimized marketing campaign data.

[0074] In a second aspect, the present invention also provides a marketing risk pre-simulation system based on dynamic simulation, applicable to the method described in the first aspect. The system includes a data receiving and processing module, an environment and decision model construction module, a scheme transformation and hybrid simulation module, a trajectory analysis and insight generation module, and an iterative optimization and output module. The data receiving and processing module receives marketing scheme data to be evaluated, historical market environment data, and multi-source social network data. The environment and decision model construction module constructs a multi-agent interaction environment model containing competitor entities, user circle entities, and media opinion entities based on historical market environment data and multi-source social network data, and configures parameterized decision models for competitor entities, user circle entities, and media opinion entities respectively. The scheme transformation and hybrid simulation module transforms the marketing scheme data to be evaluated into the initial action sequence of the user entity in the multi-agent interaction environment model, and performs multi-agent interaction... The environmental model and parameterized decision model, through a hybrid inference engine combining Monte Carlo simulation and multi-agent reinforcement learning, execute multiple dynamic game simulations in parallel, generating multiple simulated trajectory data containing environmental state evolution and entity action sequences. The trajectory analysis and insight generation module aggregates and analyzes multiple simulated trajectory data, identifies and quantifies the competitive vulnerability patterns and word-of-mouth tipping point patterns triggered by the marketing plan data to be evaluated in the dynamic game simulation, and generates a pre-simulation insight report containing a competitive vulnerability map and a word-of-mouth tipping point probability cloud map. The iterative optimization and output module, based on the pre-simulation insight report, iteratively optimizes the key parameters in the marketing plan data to be evaluated through counterfactual reasoning, generates optimized marketing plan data, and re-inputs the optimized marketing plan data into the hybrid inference engine for dynamic game simulation and evaluation until the predetermined strategy robustness conditions are met, outputting the final risk assessment and strategy optimization suggestions.

[0075] Unlike existing technologies, the above-mentioned technical solution provides a marketing risk prediction method and system based on dynamic simulation. It constructs a multi-agent interaction environment model based on historical market environment data and multi-source social network data, and configures parameterized decision models for competitor entities, user circle entities, and media opinion entities. Through a hybrid inference engine combining Monte Carlo simulation and multi-agent reinforcement learning, it performs multiple dynamic game simulations on the initial action sequences of its own entities in the multi-agent interaction environment model obtained from marketing plan data transformation, generating simulated trajectory data. It aggregates and analyzes multiple simulated trajectory data to identify and quantify competitive vulnerability patterns and word-of-mouth dissemination tipping point patterns, generating a prediction insight report. Based on this, it iteratively optimizes the marketing plan through counterfactual reasoning until the strategy robustness conditions are met. This invention achieves a forward-looking quantitative assessment and proactive optimization of the potential chain risks of marketing plans in a dynamic game environment.

[0076] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0077] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0078] In the accompanying drawings of the instruction manual:

[0079] Figure 1 This is a schematic diagram illustrating steps S101 to S105 of the method described in a specific embodiment.

[0080] Figure 2 This is a schematic diagram illustrating steps S201 to S206 of the method described in a specific implementation.

[0081] Figure 3 This is a schematic diagram of the marketing risk simulation system described in a specific implementation.

[0082] The reference numerals used in the above figures are explained as follows:

[0083] 1. Marketing risk simulation system;

[0084] 11. Data receiving and processing module;

[0085] 12. Environment and Decision Model Construction Module;

[0086] 13. Solution transformation and hybrid simulation module;

[0087] 14. Trajectory Analysis and Insight Generation Module;

[0088] 15. Iterative optimization and output module. Detailed Implementation

[0089] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended only as examples, not as limiting the scope of protection of this application.

[0090] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0091] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0092] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0093] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order between these entities or operations.

[0094] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0095] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0096] The processor described in the embodiments of this application can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0097] The computer program involved in the embodiments can be stored in a computer device readable storage medium, which includes, but is not limited to, disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical, or chemical structures that can achieve the same or equivalent functions as the storage media listed above, such as DNA, RNA, proteins, and other units with information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-mentioned media types, or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiments can be centrally stored in a single medium, or distributed and stored in multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into the device, or can be connected to the device involved in the embodiments as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0098] Please see Figure 1 In a first aspect, this embodiment provides a marketing risk prediction method based on dynamic simulation, including:

[0099] S101. Receive marketing plan data, historical market environment data, and multi-source social network data to be evaluated;

[0100] S102. Based on historical market environment data and multi-source social network data, construct a multi-agent interaction environment model that includes competitor entities, user circle entities, and media opinion entities, and configure parameterized decision models for competitor entities, user circle entities, and media opinion entities respectively.

[0101] S103. The marketing plan data to be evaluated is transformed into the initial action sequence of the entity in the multi-agent interaction environment model. Based on the multi-agent interaction environment model and the parameterized decision model, multiple dynamic game simulations are executed in parallel through a hybrid inference engine of Monte Carlo simulation and multi-agent reinforcement learning to generate multiple simulation trajectory data containing environmental state evolution and entity action sequence.

[0102] S104. Aggregate and analyze multiple simulated trajectory data to identify and quantify the competitive vulnerability patterns and word-of-mouth initiation point patterns triggered by the marketing plan data to be evaluated in the dynamic game simulation, and generate a pre-simulation insight report containing competitive vulnerability maps and word-of-mouth initiation point probability cloud maps.

[0103] S105. Based on the pre-simulation insight report, iteratively optimize the key parameters in the marketing plan data to be evaluated through counterfactual reasoning to generate optimized marketing plan data. Then, re-input the optimized marketing plan data into the hybrid inference engine for dynamic game simulation and evaluation until the predetermined strategy robustness conditions are met, and output the final risk assessment and strategy optimization suggestions.

[0104] In step S101, the marketing plan data to be evaluated specifically includes planned promotional strategies, advertising creatives, channel scheduling, budget allocation, and execution timelines; historical market environment data covers the market share timeline of various brands, marketing activity records, price changes, and macroeconomic indicators; multi-source social network data is collected from public platforms such as social media and forums, including user social relationships, interactive behaviors, content dissemination links, and account information of key opinion leaders. These three types of data define the projection objectives and construct the historical and relational context of the projection.

[0105] In step S102, the multi-agent interaction environment model abstracts key market participants into three types of entities: competitor entities simulate competing brands in reality, user circle entities represent consumer groups with similar behavioral characteristics, and media opinion entities correspond to nodes with information dissemination influence. Parameterized decision models are configured for these entities to endow them with the ability to simulate real-world decision-making logic. The competitor entity's decision model simulates its process of formulating competitive strategies based on market conditions; the user circle entity's decision model integrates social influence and individual preferences to simulate the evolution of group behavior; and the media opinion entity's decision model simulates its logic of information filtering and distribution based on content value and audience feedback. These models collectively constitute the behavioral rule base for each entity in the dynamic game environment.

[0106] In step S103, the marketing plan data is transformed into the initial action sequence of the entity, which encodes the marketing plan into a standard set of executable instructions for the model. Monte Carlo simulation covers the uncertainty of the market environment through random sampling, while multi-agent reinforcement learning drives each entity to make adaptive strategy adjustments based on its own decision-making model and environmental feedback. The hybrid inference engine of Monte Carlo simulation and multi-agent reinforcement learning executes multiple simulations in parallel. Each simulation runs the entity's plan in a possible market scenario and simultaneously infers the reactions and interactions of competitors, user groups, and media nodes. The generated simulation trajectory data fully records the evolution sequence of environmental states (such as volume and market share) and the actions taken by each entity at each step in each inference.

[0107] In step S104, multiple simulated trajectory data are aggregated and analyzed to extract statistically significant patterns from a large number of simulation results. Competitive vulnerability patterns refer to interaction sequences where specific marketing actions of the company are associated with high-frequency counter-attacks from competitors, resulting in damage to the company's objectives. Word-of-mouth virality tipping point patterns refer to common pre-existing conditions or event combinations observed in the simulation before an explosive increase in buzz. Identifying and quantifying these patterns transforms simulation data into quantitative insights into the weaknesses and opportunities of the strategy. The generated pre-simulation insight report visualizes high-risk confrontation paths through a competitive vulnerability map and displays potential trigger points and their probability distribution for buzz explosions through a word-of-mouth tipping point probability cloud map.

[0108] In step S105, based on the pre-simulation insight report, iterative optimization is performed using counterfactual reasoning. Counterfactual reasoning constructs an analytical framework: keeping other conditions constant, only key parameters in the marketing plan (such as resource allocation and action timing) are adjusted to generate multiple plan variants, which are then re-entered into the hybrid simulation engine for evaluation. By comparing the performance of each variant in the simulation, the effectiveness of parameter adjustments in mitigating vulnerabilities and leveraging tipping points is evaluated. This process is repeated until the resulting plan's performance in the simulation, including indicators such as goal achievement stability, meets predetermined strategy robustness conditions. The final output includes a quantitative risk assessment and an optimized plan validated by simulation, along with the basis for adjustments.

[0109] This embodiment abstracts the market into a dynamic game system by constructing an interactive environment model containing multiple entities and configuring a parameterized decision model for it. It then utilizes a hybrid inference engine to execute numerous parallel simulations covering various future market scenarios. Through aggregated analysis of the simulation trajectories, it extracts two key risk patterns: competitive vulnerability and word-of-mouth tipping points, generating visual insights. Finally, based on these insights, iterative optimization of the original solution is achieved using counterfactual reasoning until a strategy that meets robustness requirements is obtained. This method enables proactive quantitative assessment and optimization of the potential cascading risks of marketing plans in a dynamic game environment.

[0110] Please see Figure 2 In some embodiments, based on historical market environment data and multi-source social network data, a multi-agent interaction environment model is constructed, including competitor entities, user circle entities, and media opinion entities, including:

[0111] S201. Extract and quantify entity relationships from historical market environment data to generate a competitive relationship intensity matrix and market resource constraint parameters;

[0112] S202. Perform graph structure modeling and community discovery on multi-source social network data to generate user social network graphs and cross-platform influence network graphs. In the user social network graphs, nodes are labeled with multiple user circle labels divided by clustering algorithms.

[0113] S203. Based on the competitive relationship intensity matrix, construct a competitive ecosystem map with competitor entities as nodes, and configure a resource constraint model for competitor entities that includes budget, production capacity and channel scheduling.

[0114] S204. Based on the user social network graph and multiple user circle tags, construct a group behavior model with user circle entities as agents. The group behavior model integrates social network propagation dynamics and behavioral economics theory to simulate the process of individual decision-making being influenced by social factors and cognitive biases.

[0115] S205. Based on the cross-platform influence network graph, construct an influence propagation model for media opinion entities, wherein the influence propagation model encodes the characteristic differences of different information dissemination platforms and the position preference parameters of media opinion entities.

[0116] S206. The competitive ecosystem map, the group behavior model and the influence propagation model are coupled to form a multi-agent interaction environment model. The multi-agent interaction environment model defines the state perception and interaction rules between competitor entities, user circle entities and media opinion entities.

[0117] In step S201, entity relationship extraction and quantification are achieved through named entity recognition technology. Key entities such as brands and products are identified from historical market environment data, and historical interaction records between entities are analyzed. The quantification process calculates the strength of the association between entities, such as statistically analyzing the frequency of mutual responses between brand marketing activities within a time window, analyzing the correlation of waxing and waning in market share time-series data, or assessing the degree of emotional antagonism accompanying simultaneous mentions in public text data. Multiple association measurement results are integrated to generate a numerical competitive relationship strength matrix. Market resource constraint parameters are obtained through regression analysis of the input costs and resulting market effect indicators of historical marketing activities. This parameter describes the functional relationship or constraint boundary between resource input and effect output.

[0118] In step S202, graph structure modeling abstracts user accounts and content publishers in multi-source social network data into nodes, and abstracts interactive relationships such as following, forwarding, and replying into edges, constructing a graph representing the connection relationships. Community discovery applies a clustering algorithm to divide the nodes in the graph, classifying them into different clusters based on the tightness of connections and attribute similarity between nodes, with each cluster assigned a user circle label. The user social network graph is the graph structure with these circle labels attached. When constructing the cross-platform influence network graph, the correspondence between the same opinion leader's accounts on different social media platforms is first established. Then, with platforms and opinion leaders as two types of nodes, and the relationship of cross-platform content synchronization or referencing as edges, this bipartite graph is constructed and analyzed, ultimately forming a weighted network graph reflecting the path and intensity of cross-platform information flow.

[0119] In step S203, the competitive ecosystem graph is constructed using the competitive relationship strength matrix as the adjacency matrix. Nodes in the graph represent competitor entities, and edge weights represent competitive intensity. The resource constraint model configured for competitor entities includes multiple dimensions: budget constraints limit the total amount of marketing funds available during the simulation period; capacity constraints reflect the upper limit of their product production or service delivery capabilities; and channel scheduling constraints define their time-occupancy rules and exclusivity agreements on various promotional channels. The values ​​of these constraint parameters can be set by analyzing the competitor's historical operating data or industry benchmarks.

[0120] In step S204, a group behavior model with user circle entities as agents is constructed, defining an algorithmic framework for simulating individual decision-making. This framework integrates social network propagation dynamics, for example, by employing a threshold model to set the probability of an individual adopting a certain behavior (such as purchasing or sharing) to increase non-linearly with the proportion of their social neighbors who have already adopted that behavior. Simultaneously, the framework incorporates behavioral economics theory, introducing cognitive bias parameters, such as reference point dependence and loss aversion coefficients, into the individual's utility calculation function, making the individual's perception of the negative utility of loss stronger than their perception of the positive utility of an equivalent gain. The decision output of the user circle entities is an action probability distribution calculated using the aforementioned algorithmic framework, based on individual attributes, currently received information stimuli, and their local social network state.

[0121] In step S205, the differences in characteristics of information dissemination platforms may include the platform's content decay coefficient, average user activity, and coverage of users in different circles. These parameters can be obtained by fitting publicly available data or historical dissemination data from the platform. The position preference parameters of media opinion entities can be quantified by performing thematic clustering and sentiment analysis on their historical published content to determine their inclination on specific issues. The decision logic of the influence dissemination model is expressed as a content selection function: for a set of candidate content, the model calculates the matching degree between each piece of content and its own position parameters, the estimated potential audience size, and possible commercial incentives, and selects the content with the highest comprehensive score after weighted summation to execute the dissemination action.

[0122] In step S206, the state-awareness rules clarify the observable environmental information for each entity: competitor entities can perceive market share and competitor actions; user circle entities can perceive information and behaviors from media and social neighbors; media opinion entities can perceive public opinion hotspots and audience feedback. The interaction rules define how entity actions affect the environmental state: competitor entities' marketing actions (such as price reductions or new product launches) directly change the market resource landscape, thus affecting the utility calculations of other competitors and user entities; user circle entities' purchasing and sharing behaviors are transformed into market sales and buzz data, which are then fed back to competitor entities and media opinion entities; media opinion entities' dissemination actions directly become information input for user circle entities. These perception-action loops are tightly connected through the state transition function defined by the model, enabling any entity's action to trigger a chain reaction of environmental state evolution.

[0123] This embodiment extracts quantifiable relationships from historical and social data, segments user communities, and constructs a competitive graph. It then configures parameterized decision-making models incorporating domain knowledge (such as behavioral economics and communication dynamics) for three types of entities: competitors, user communities, and media opinions. Finally, it couples these entities into an organic whole through clearly defined state perception and interaction rules. This model is no longer a simple statistical or regression model, but a dynamic computational environment capable of characterizing multi-party games, social impact, and information dissemination. This lays a solid foundation for subsequent high-fidelity dynamic game simulations, enabling risk prediction to be conducted in a highly realistic complex system environment.

[0124] In some embodiments, entity relationships are extracted and quantified from historical market environment data to generate a competitive relationship strength matrix and market resource constraint parameters, including:

[0125] Extract historical marketing campaign sequences, historical market share time-series data, and historical brand co-occurrence data from historical market environment data;

[0126] Based on historical marketing campaign sequences, the mutual response strength of marketing actions between different brand entities is calculated using a causal inference algorithm, serving as the primary quantitative indicator of competitive relationship strength.

[0127] Based on historical market share time series data, the negative correlation between changes in market share of different brands is calculated through time series correlation analysis, which serves as a second quantitative indicator of the intensity of competitive relationship.

[0128] Based on historical brand co-occurrence data, the intensity of opposing sentiment associations between different brand entities in the public opinion field is calculated through sentiment analysis in natural language processing, serving as a third quantitative indicator of the intensity of competitive relationships.

[0129] The first, second, and third quantitative indicators of competitive relationship strength are combined and weighted to generate a competitive relationship strength matrix.

[0130] By analyzing historical resource input data of each brand entity from historical market environment data, market resource constraint parameters are obtained through statistical regression fitting. These parameters characterize the constraint relationship between resource input and marketing effectiveness output.

[0131] In this embodiment, the historical marketing campaign sequence records the marketing events initiated by each brand in the past time period, such as promotions, advertisements, and new product launches, and their timestamps; the historical market share time series data is the market share change data of each brand, which is statistically analyzed at fixed time intervals (such as monthly or quarterly); the historical brand co-occurrence data comes from texts such as news, social media, and forums, and records the context in which the brand name is mentioned at the same time.

[0132] The first quantitative indicator for calculating the strength of competitive relationships based on historical marketing campaign sequences is analyzed using causal inference algorithms. For example, the Granger causality test can be applied to examine whether a brand's marketing campaign time series statistically significantly leads another brand's market share or brand awareness change series. If the null hypothesis of "no causal relationship" can be rejected, the calculated test statistic or p-value can be converted into a response strength indicator. Preferably, methods based on transfer entropy or convergent cross-mapping can be used to quantify the mutual predictive ability of marketing actions between brands over time, thus serving as a measure of response strength.

[0133] The second quantitative indicator is calculated based on historical market share time-series data, achieved through time-series correlation analysis. The correlation coefficient between the market share time series of different brands is calculated. Since competitive relationships are typically inversely related, the degree of negative correlation is emphasized. Specifically, the Pearson correlation coefficient can be calculated and negative values ​​taken; or the Spearman rank correlation coefficient can be calculated to capture non-linear negative correlations. The larger the absolute value of the resulting negative correlation coefficient, the higher the value of the second quantitative indicator characterizing the intensity of the competitive relationship.

[0134] The third quantitative indicator is calculated based on historical brand co-occurrence data, utilizing sentiment analysis technology from natural language processing. Sentiment polarity analysis is performed on text fragments containing brand co-occurrences to determine whether the overall sentiment of the text or for each brand is positive, negative, or neutral. The proportion of co-occurrences with opposing sentiments when two brands are mentioned simultaneously is calculated, or the cosine similarity of the sentiment vectors is calculated (negative values ​​indicate opposition), which serves as the strength of the opposing sentiment association.

[0135] A competitive relationship strength matrix is ​​generated by integrating the three quantitative indicators mentioned above, using a weighted summation method. Weight coefficients are assigned to the first, second, and third quantitative indicators, which can be determined based on business experience or verified through historical data. For each brand pair, the values ​​of its three corresponding indicators are multiplied by their respective weights and then summed. The result is the element value of that brand pair in the competitive relationship strength matrix. This weighted fusion comprehensively reflects the competitive relationship across three different dimensions: behavioral confrontation, market outcomes, and public opinion / sentiment.

[0136] Historical resource input data for each brand entity is analyzed from historical market environment data, including advertising expenses, promotion costs, and channel construction investment; corresponding marketing effect output data can include sales revenue, market share, and brand search index. Market resource constraint parameters are obtained through statistical regression fitting. For example, a linear or nonlinear regression model (such as a log-linear model or a saturation curve model) is established between resource input (independent variable) and effect output (dependent variable). The model coefficients obtained from the fitting (such as marginal effect and saturation point) are the market resource constraint parameters. These parameters describe the change in output increment brought about by an increase in input, providing quantitative constraints for resource allocation decisions in the simulation.

[0137] This embodiment extracts three dimensions of quantitative competitive intensity indicators from multi-source heterogeneous data, including marketing campaign sequences, market share time series, and brand mention text, using algorithms such as causal inference, time series correlation, and sentiment analysis. These indicators are then weighted and fused to generate a more comprehensive and robust competitive relationship strength matrix that accurately depicts the adversarial relationships between brands. Simultaneously, regression analysis is used to fit historical input-output data to obtain quantified market resource constraint parameters. This embodiment provides a reliable and computable foundation for constructing a high-fidelity multi-agent interaction environment model, enhancing the scientific rigor and credibility of the entire simulation and inference system.

[0138] In some embodiments, parameterized decision models are configured for competitor entities, user community entities, and media opinion entities, including:

[0139] Configure a first parameterized decision model for the competitor entity. The first parameterized decision model includes a policy network built based on a deep reinforcement learning algorithm. The input of the policy network is an environmental state vector containing the entity's own state, the competitor's state, and the overall market state. The output of the policy network is the probability distribution of the competitor entity's actions or deterministic action values ​​in the action space. The objective function of the first parameterized decision model is to maximize the long-term cumulative reward. The long-term cumulative reward is calculated by the changes in market share, changes in brand equity, and action costs.

[0140] A second parameterized decision-making model is configured for user circle entities. The second parameterized decision-making model includes a decision function constructed based on the intelligent agent modeling framework and behavioral economics theory. The input of the decision function is an individual attribute vector, a local social network state vector, and a received information stimulus vector. The decision function simulates individual decision-making by calculating the comprehensive utility value and introducing random noise. The comprehensive utility value integrates personal preference utility, social conformity utility, and risk perception utility under the prospect theory framework.

[0141] A third parametric decision-making model is configured for media opinion entities. The third parametric decision-making model includes a content selection function built based on the utility maximization theory. The input of the content selection function is a vector of candidate content attributes, a vector of audience matching degree, and a vector of potential commercial incentives. The content selection function calculates the expected influence return and commercial return by weighting, and selects the content with the highest comprehensive return to perform the dissemination action.

[0142] In this embodiment, the first parameterized decision model configured for the competitor entity typically employs a deep neural network architecture, such as a multilayer perceptron, in its strategy network built based on deep reinforcement learning algorithms. Its input layer receives an encoded environmental state vector. This environmental state vector integrates the competitor entity's own resource status (e.g., current budget, capacity utilization), recent market actions of major competitors (e.g., last month's promotional efforts, advertising spending), and overall market indicators (e.g., industry total demand growth rate, raw material price index). The output layer of the strategy network corresponds to the competitor entity's action space, which includes discrete or continuous actions such as price reductions, new product launches, increased advertising spending, and channel expansion. The output can be the probability distribution of each action, generated through a softmax function; or it can be the deterministic Q-value or advantage value of each action. The objective function of the first parameterized decision model is set to maximize long-term cumulative rewards. This reward function is calculated at each simulation time step: the reward value equals the increase in market share and the positive change in brand equity (quantifiable by improvements in brand search index or public sentiment) gained at that step, minus the cost of taking action (e.g., marketing expenses, capacity adjustment costs). The long-term cumulative reward is the sum of the discounted present value of rewards from all future steps. When training the first parameterized decision model, a large number of simulated rounds are run in the constructed multi-agent interaction environment model to collect state-action-reward sequences. The gradient update mechanism of deep reinforcement learning algorithms (such as Proximal Policy Optimization (PPO) and Deep Deterministic Policy Gradient (DDPG)) is used to iteratively optimize the parameters of the policy network, so that the competitor entities learn to adopt strategies that maximize long-term interests in complex game environments.

[0143] The decision function, constructed based on an intelligent agent modeling framework and behavioral economics theory, takes as input an individual attribute vector representing individual characteristics (e.g., income level, brand loyalty, innovation adoption tendency), a local social network state vector reflecting the current state and behavior of the individual's social network neighbors, and a vector of information stimuli received from media or other users (e.g., advertising content, word-of-mouth information). The decision function simulates the individual's decision-making process by calculating a comprehensive utility value. This comprehensive utility value is a weighted sum of three parts: personal preference utility, which calculates the matching degree between product attributes and the individual's preference vector; social conformity utility, calculated based on the proportion of the individual's social neighbors who have already adopted the target behavior (the higher the proportion, the greater the conformity utility); and risk perception utility under the prospect theory framework. This part of the utility calculation is non-linear, assigning a higher negative weight to perceived losses than to equivalent gains, and introducing a probability weighting function that causes the individual to overestimate low-probability events and underestimate medium- to high-probability events. After calculating the combined utility value of each possible action, the decision function does not directly select the action with the highest utility. Instead, it introduces a random noise term (for example, by converting the utility value into a selection probability through a logistic function, or by adding Gumbel distribution noise). Finally, it randomly samples the probability distribution after adding noise to determine the individual's action, thereby simulating the uncertainty and individual differences in real-world decision-making.

[0144] The content selection function, built upon utility maximization theory, takes as input a vector of candidate content attributes (encoding topic category, sentiment, novelty, source credibility, etc.), an audience fit vector (representing the degree of matching between the content and the interests of various user groups, calculated using embedding vector similarity trained from historical interaction data), and a vector of potential commercial incentives (such as advertising fees and revenue sharing). The function evaluates each candidate content by weighting the expected impact return and commercial return. The expected impact return estimates potential metrics such as reposts, views, and positive comments after content publication based on the candidate content attribute vector and audience fit vector, converting this into a scalar return value. Commercial return is directly linked to the potential commercial incentive vector. The function assigns pre-defined weights to these two types of returns and sums them to obtain the overall return for each candidate content. Ultimately, the media entity selects the candidate content with the highest overall return for its dissemination action. The media entity's stance preference parameters are implicit in the audience fit calculation or return weights; for example, content aligned with its stance may receive higher weight or valuation in the impact return estimation.

[0145] This embodiment configures an adaptive policy network based on deep reinforcement learning for competitor entities, a probabilistic decision function incorporating behavioral economics theory for user circle entities, and a content selection function based on utility maximization for media opinion entities. It endows the three core entities in the multi-agent interaction environment model with highly realistic and computable behavioral logic, enabling each entity to act as an intelligent agent in the simulation to make complex decisions based on environmental state, its own goals, and intrinsic preferences. This ensures the depth, realism, and uncertainty of subsequent dynamic game simulations, laying the foundation for extracting valuable risk patterns from the simulation.

[0146] In some embodiments, graph structure modeling and community discovery are performed on multi-source social network data to generate user social network graphs and cross-platform influence network graphs, including:

[0147] Clean and align user interaction behavior data and content association data in multi-source social network data to identify unique user nodes and content nodes;

[0148] Based on user interaction behavior data, an initial social network graph is constructed with user nodes as vertices and follow relationships, forwarding relationships, or co-occurrence relationships as edges;

[0149] Applying a graph embedding algorithm to the initial social network graph yields a low-dimensional vector representation of user nodes;

[0150] Based on the low-dimensional vector representation of user nodes, a community detection algorithm is used to cluster user nodes, the clustering results are labeled as multiple user circle labels, and the multiple user circle labels are mapped back to the initial social network graph to generate a user social network graph with circle labels.

[0151] By integrating cross-platform account association data and content dissemination path data from multi-source social network data, media opinion leader nodes and their influence range on different information dissemination platforms can be identified.

[0152] Based on content dissemination path data, a bipartite graph is constructed with media opinion leader nodes and platforms as vertices and cross-platform content migration relationships as edges;

[0153] Projecting and weighting the bipartite graph generates a cross-platform influence network graph. The edge weights in the cross-platform influence network graph represent the probability or intensity of information migration between platforms.

[0154] In this embodiment, cleaning and aligning user interaction behavior data and content-related data involves removing duplicate records, filling in missing values ​​in key fields, and normalizing the same user or content identifier from different data sources. For example, user IDs from different platforms are mapped to a unified internal ID through account binding relationships or device fingerprints, thereby ensuring that each user node and content node is unique in the subsequent graph structure.

[0155] An initial social network graph is constructed based on user interaction behavior data. This is achieved by parsing the relationship records in the data: follow relationships originate from a user's follow list, forwarding relationships originate from content forwarding records, and co-occurrence relationships are obtained by statistically analyzing the frequency of user interactions on the same topic or within the same time period. With users as vertices and follow, forwarding, or co-occurrence relationships as edges, a social network graph representing the connections between users is formed.

[0156] Graph embedding algorithms are applied to transform each user node in the initial social network graph into a low-dimensional dense vector. By learning the local neighborhood structure of nodes in the network, graph embedding algorithms ensure that nodes that are closely connected or structurally similar in the original graph are also closer in the low-dimensional vector space. Through training, each user node obtains a fixed-length low-dimensional vector that contains its structural position information within the social network.

[0157] Based on the low-dimensional vector representation of user nodes, a community detection algorithm is used for clustering. The algorithm divides all user nodes into several clusters according to the similarity between vectors (such as cosine similarity, Euclidean distance) or distribution density. The node vectors in each cluster are clustered in the feature space, representing that these users have high homogeneity in social behavior or interest preferences. The cluster numbers obtained from clustering are multiple user circle labels. These labels are attached as attributes to the corresponding nodes in the initial social network graph to generate a user social network graph with circle labels.

[0158] By integrating cross-platform account association data, and matching the similarity of account names, homepage links, published content, or official certification association information provided by the platform, different platform accounts belonging to the same opinion leader entity can be identified; content dissemination path data records the complete link from the publication of content to its citation or synchronization across platforms; the scope of influence can be quantitatively assessed by analyzing indicators such as the number of followers of the opinion leader on each platform and the average interaction volume of historical content (such as reposts, comments, and likes).

[0159] A bipartite graph is constructed based on content dissemination path data: one type of vertex is media opinion leader nodes, and the other type of vertex is information dissemination platforms. When it is detected that a piece of content is disseminated from an opinion leader's account on platform A to their account on platform B, an edge is established in the bipartite graph from the corresponding "opinion leader-platform A" vertex to the "opinion leader-platform B" vertex.

[0160] Projecting and weighting a bipartite graph typically involves projecting the graph onto media opinion leader nodes. After projection, the edge weights connecting two opinion leader nodes can be calculated by counting the number of times content migrates between them through shared platforms, or by accumulating the total interaction volume generated by migrated content. This weight characterizes the probability or strength of information migration between the platforms influenced by the two opinion leaders, ultimately generating a cross-platform influence network graph.

[0161] This embodiment transforms social network interaction data into a user social network graph with a clear structure and semantic labels through data cleaning, graph construction, embedding representation, and community segmentation processes. By constructing and analyzing a bipartite graph of cross-platform content migration, it generates a relationship network that reflects the cross-platform influence flow patterns among opinion leaders, providing a structured quantitative basis for simulated social interaction and information dissemination.

[0162] In some embodiments, the marketing campaign data to be evaluated is transformed into an initial sequence of actions for the local entity in a multi-agent interaction environment model, including:

[0163] Analyze the marketing plan data to be evaluated, and extract the core marketing action elements, resource allocation plans, and execution timelines contained in the marketing plan data;

[0164] The core marketing action elements are mapped to standard action codes in the action space that the entity can execute in the multi-agent interaction environment model. The standard action codes include promotion action codes, advertising action codes, channel strategy action codes, and public opinion guidance action codes.

[0165] Based on the resource allocation plan, allocate corresponding resource quantification parameters to the mapped standard action codes;

[0166] Based on the execution timeline, the standard action codes for allocating resource quantification parameters are arranged and combined in chronological order to generate an initial action sequence.

[0167] In this embodiment, parsing the marketing campaign data to be evaluated involves extracting structured information from the input data. For unstructured text formats, named entity recognition and relation extraction techniques from natural language processing can be used to automatically identify entities describing marketing activities, resource allocation, and scheduling, as well as the relationships between them. For structured or semi-structured data, the corresponding data fields are directly read. Through parsing, the core marketing action elements describing specific marketing behaviors, resource allocation plans indicating the amount and type of resource input, and execution timelines specifying the start and end dates or sequence of each action are extracted.

[0168] The core marketing action elements are mapped to standard action codes based on a predefined coding mapping table that fully corresponds to the action space of the entity in the multi-agent interaction environment model. This mapping table establishes a correspondence between natural language descriptions or business terms and internally unique action identifiers. The mapping process queries the mapping table to match the parsed action element descriptions to the closest standard action codes. For example, a description about price adjustments is mapped to a promotion action code, a description about media use is mapped to an advertising placement action code, a description about sales channel selection is mapped to a channel strategy action code, and a description about information dissemination and guidance is mapped to a public opinion guidance action code.

[0169] Based on the resource allocation plan, corresponding resource quantification parameters are assigned to the mapped standard action codes. This involves transforming the abstract resource descriptions such as budget, manpower, and materials in the plan into numerical parameters that the model can calculate and binding them to specific actions. The type of resource quantification parameters is related to the type of action code. For example, promotional actions may be associated with parameters such as discount rates and coupon values, while advertising actions may be associated with parameters such as budget amount, target audience feature vectors, and display frequency. This allocation process fills in the required parameter values ​​for each action code based on the parsed resource allocation plan.

[0170] The initial action sequence is generated based on the execution timeline. First, the continuous timeline is discretized into a sequence of time units consistent with the simulation time step. Then, based on the planned execution time point of each standard action code and its attached resource quantification parameters in the original execution timeline, it is assigned to the corresponding time unit. If multiple actions are planned to be executed within the same time unit, these action codes and their parameters are combined into an action set for that time unit. The actions (sets) of all time units are arranged in chronological order and linked together to form a complete initial action sequence.

[0171] This embodiment achieves the conversion from human-readable marketing plans to machine-executable precise instruction sequences through standardized information extraction, encoding mapping, parameter binding, and time alignment processes, ensuring that the input of subsequent dynamic game simulations remains consistent with the original strategic intent in terms of logic, resources, and time.

[0172] In some embodiments, based on a multi-agent interaction environment model and a parameterized decision model, a hybrid inference engine combining Monte Carlo simulation and multi-agent reinforcement learning is used to execute multiple dynamic game simulations in parallel, generating multiple simulated trajectory data containing environmental state evolution and entity action sequences, including:

[0173] Initialize the hybrid simulation engine, and set the total number of parallel simulation instances and the maximum simulation time step for each simulation instance;

[0174] Generate an independent random seed for each parallel simulation instance, and initialize the initial state of other entities in the multi-agent interaction environment model, excluding the agent's own entity, based on the independent random seed;

[0175] Within each time step of each simulation instance, perform the following steps:

[0176] Invoke the first parameterized decision model of the competitor entity and generate competitive actions of the competitor entity based on the current environmental state;

[0177] The second parameterized decision model of the user circle entity is invoked to generate the group behavior actions of the user circle entity based on the current social network state and the intensity of information stimuli.

[0178] The third parameterized decision-making model of the media opinion entity is invoked to generate the dissemination actions of the media opinion entity based on the current content ecosystem and commercial incentives;

[0179] The actions of our own entity, competitive actions, group behavior actions, and propagation actions are input into the multi-agent interaction environment model. Based on the state transition rules defined in the multi-agent interaction environment model, the environment state of the next time step is calculated and updated.

[0180] During the environment state update process, check whether the preset key event triggering conditions are met. If they are met, activate the corresponding key event rule and modify the subsequent state transition rules.

[0181] Record the environmental state, the actions of all entities, and the key events triggered at the current time step to form a trajectory segment at the current time step;

[0182] When a simulation instance reaches the maximum time step or meets the simulation termination condition, the simulation instance ends, and all trajectory segments connected in chronological order in the simulation instance are summarized into complete simulation trajectory data.

[0183] Collect complete simulation trajectory data generated by all parallel simulation instances to form multiple simulation trajectory data containing environmental state evolution and entity action sequences.

[0184] In this embodiment, when initializing the hybrid simulation engine, the total number of parallel simulation instances is set according to the computing resources and the requirements for statistical significance of the results, in order to cover the uncertainty of the market environment; the maximum simulation time step of each simulation instance is determined according to the planning cycle of the marketing plan and the time scale of the market dynamics that are of interest, for example, the entire plan cycle is discretized into tens to hundreds of simulation steps.

[0185] An independent random seed is generated for each parallel simulation instance. This seed is used to initialize the pseudo-random number generator, ensuring the independence of all random processes in each instance (such as user decision noise, media content selection randomness, and initial environmental perturbations). Based on the independent random seed, the initial states of other entities in the multi-agent interaction environment model, excluding the agent's own entity, are initialized. For example, random initial resource levels are assigned to each competitor entity, random initial opinion distributions are generated for user circle entities, and random initial agenda preferences are set for media opinion entities, thereby constructing diverse initial market scenarios.

[0186] Within each simulation instance and each time step, the competitor entity's first parameterized decision model is invoked to receive the current environment state vector, which encodes information such as current market share, competitor action history, and total market demand. This vector is then calculated through its internal policy network to output the competitive actions that the competitor entity plans to execute in this time step, such as adjusting prices or launching promotions.

[0187] The second parameterized decision model of the user circle entity is invoked. It receives individual attributes, the current state and behavior of its social neighbors, and information stimulus vectors received from the media and its own entity. By integrating personal preferences, social influence and risk perception, the comprehensive utility calculation is performed to output the group behavior action that the user circle entity may take at this time step, such as purchasing, sharing or posting negative comments.

[0188] The third parameterized decision-making model of media opinion entities is invoked. Based on the distribution of hot topics in the current content ecosystem, the emotional feedback of various user groups, and potential commercial cooperation invitations, the content with the highest expected comprehensive benefits is selected from the candidate content through utility maximization calculation, and its dissemination action is generated at the current time step.

[0189] The actions of our own entities (from the initial action sequence), competitive actions, group behavior actions, and propagation actions are all input into the multi-agent interaction environment model. Based on the state transition rules defined in the multi-agent interaction environment model (usually a set of predefined mathematical functions or logical conditions, whose inputs are the current environment state and the actions of all entities, and whose output is the updated environment state; the specific function form is based on market dynamics principles, such as supply and demand relationships and information diffusion models), the impact of all the above actions on indicators such as market resources, user opinions, and network volume is comprehensively calculated, thereby calculating and updating the environment state for the next time step.

[0190] During the environmental state update process, it is checked whether the updated state meets the preset key event trigger conditions. Key event trigger conditions can be set to certain state variables exceeding specific thresholds, such as a competitor's market share suddenly dropping to a dangerous level, or a surge in negative sentiment about the company in a short period of time. If the conditions are met, the corresponding key event rule is activated. This rule may temporarily modify certain state transition parameters, introduce new constraints, or change the behavior patterns of specific entities to simulate structural changes brought about by major market events.

[0191] The simulation instance records the environmental state (values ​​of all state variables), the actions of all entities (action codes and parameters), and triggered key events (if any) at the current time step, forming a trajectory segment for the current time step. The simulation instance ends when it reaches its maximum extrapolation time step or meets the simulation termination condition (e.g., one's own entity's resources are exhausted, or a preset market objective is achieved). All trajectory segments connected chronologically within the simulation instance are aggregated to form a complete simulation trajectory dataset, which fully records the complete evolution sequence of the environmental state and the actions of each party over time. Collecting all parallel simulation instances forms multiple simulation trajectory datasets containing environmental state evolution and entity action sequences, constituting a high-dimensional dataset covering multiple possible market futures and recording the detailed game process, providing a foundation for subsequent aggregation analysis.

[0192] This embodiment fully realizes a parallel, event-driven multi-agent dynamic game simulation loop by initializing parallel instances, sequentially driving three types of intelligent entities to make actions based on their parameterized decision models in each step, updating the global state according to the environment model, and dynamically responding to key events. This hybrid inference engine organically integrates the random sampling idea of ​​Monte Carlo simulation with the mechanism of agents making decisions based on environmental feedback in multi-agent reinforcement learning. It can efficiently generate a large number of high-fidelity market evolution trajectories, thus providing a rich and reliable data foundation for identifying potential risk patterns.

[0193] In some embodiments, multiple simulated trajectory data are aggregated and analyzed to identify and quantify the competitive vulnerability patterns and word-of-mouth tipping point patterns triggered by the marketing campaign data to be evaluated in dynamic game simulation, generating a pre-simulation insight report containing a competitive vulnerability map and a word-of-mouth tipping point probability cloud map, including:

[0194] Filter through multiple simulated trajectory data and extract negative outcome trajectory data from all simulated trajectory data, such as the final decline in market share or damage to brand health of our entity;

[0195] By performing time-series pattern mining on negative outcome trajectory data, we can pinpoint frequently co-occurring action pairs between marketing actions initiated by our entity and adversarial actions taken by competitors within the time window before a significant decline in market share.

[0196] Calculate the support and confidence of each action pair combination in all negative result trajectory data, and identify action pair combinations whose support and confidence both exceed a preset threshold as competitive vulnerability patterns.

[0197] Based on the competitive vulnerability pattern, a competitive vulnerability map is generated by visually annotating the competitive ecosystem map in the multi-agent interaction environment model. The competitive vulnerability map shows the typical attack paths from one's own actions to the competitor's counterattack and their frequency of occurrence.

[0198] Filter through multiple simulated trajectory data and extract the positive result trajectory data from all simulated trajectory data where the total sound volume increases beyond the preset burst threshold;

[0199] Feature backtracking analysis is performed on the positive result trajectory data to extract common initial environmental state features and early key event features that existed before the sound volume explosion, forming a potential tipping point feature set.

[0200] Calculate the conditional probability of a sound burst event occurring given that each feature combination in the potential detonation point feature set appears in the entire simulated trajectory dataset.

[0201] Feature combinations whose conditional probability exceeds a preset probability threshold are identified as word-of-mouth marketing tipping point patterns, and leverage conditional parameters corresponding to word-of-mouth marketing tipping point patterns are calculated.

[0202] Based on the word-of-mouth propagation tipping point model and its conditional probability and leverage conditional parameters, a word-of-mouth tipping point probability cloud map is generated by visual rendering on the user social network graph and cross-platform influence network graph in the multi-agent interaction environment model.

[0203] Integrate competitive vulnerability maps and word-of-mouth tipping point probability cloud maps, and add statistical summaries and key findings descriptions to generate a pre-launch insight report.

[0204] In this embodiment, damage to brand health is determined based on whether multiple indicators, including the comprehensive public opinion sentiment index and brand search index, fall below a safety threshold set according to historical normal fluctuation ranges at the end of the simulation. The preset outbreak threshold is determined by analyzing the distribution of voice growth data from historical successful marketing cases and taking the value of a specific quantile.

[0205] The time-series pattern mining employs a sequence pattern mining algorithm. Within a pre-defined time window preceding a significant market share decline inflection point, it analyzes the co-occurrence relationship between the company's marketing action encoding sequences and competitors' adversarial action encoding sequences. Support is calculated as the ratio of the number of negative trajectories containing a specific action pair to the total number of negative trajectories. Confidence is calculated as the ratio of the number of trajectories showing a significant market share decline after the occurrence of that action pair to the total number of trajectories containing that action pair. The pre-defined thresholds for support and confidence are set based on business experience or a trade-off between historical false positive and false negative rates.

[0206] The competitive vulnerability graph is generated by adding directed edges between the player's entity nodes and the competitor's entity nodes on the competitive ecosystem graph. The visual attributes of the edges are encoded based on the frequency of occurrence of the corresponding competitive vulnerability patterns or the calculated confidence level.

[0207] Total volume growth is calculated using cumulative interaction data such as mentions and reposts related to the platform within the simulated period. Feature backtracking analysis traces back a fixed time window from the volume surge point, extracting environmental characteristics such as average sentiment values ​​of each user group, the volume share of major competitors, and the concentration of media topics within that window, as well as early key event characteristics. Common features are extracted from the positive trajectory set through cluster analysis or frequent itemset mining.

[0208] Conditional probability is calculated as the ratio of the number of trajectories in the entire simulated trajectory dataset that exhibit a specific feature combination and ultimately result in a volume surge, to the total number of trajectories exhibiting that feature combination. The preset probability threshold is set based on the business requirements for prediction accuracy and recall. The leverage conditional parameter is obtained by calculating the ratio of the probability of a volume surge when the feature combination occurs to the probability of a volume surge when the feature combination does not occur.

[0209] The generation of the word-of-mouth tipping point probability cloud map is achieved by highlighting the relevant user circle nodes, media opinion leader nodes, or dissemination paths on the user social network graph and cross-platform influence network graph. The color or size of the nodes is gradually rendered according to the conditional probability or leverage condition parameters of the corresponding word-of-mouth dissemination tipping point mode.

[0210] The pre-launch insight report integrates a competitive vulnerability map and a probability cloud map of word-of-mouth tipping points, and includes summaries of key statistical indicators such as the total number of identified patterns, the highest risk value, and the most potential leverage points, as well as textual descriptions of key findings and business recommendations.

[0211] This embodiment sets quantitative judgment thresholds and applies data analysis methods such as sequence pattern mining, feature backtracking, conditional probability and lift calculation to automatically extract and quantify competitive vulnerability patterns and word-of-mouth dissemination tipping point patterns from simulated trajectories. It also uses network visualization technology to generate intuitive insight maps, transforming complex simulated data into clear and actionable insights into risks and opportunities.

[0212] In some embodiments, based on the pre-visualization insight report, key parameters in the marketing campaign data to be evaluated are iteratively optimized using counterfactual reasoning to generate optimized marketing campaign data, including:

[0213] Analyze the pre-launch insight report to identify high-frequency and high-risk competitive vulnerability patterns from the competitive vulnerability map, and identify high-probability and high-leverage potential word-of-mouth dissemination tipping point patterns from the word-of-mouth tipping point probability cloud map.

[0214] For the identified competitive vulnerability patterns, the adjustable marketing action parameters associated with them in the marketing program data to be evaluated are identified as the first type of key parameters;

[0215] For the identified word-of-mouth marketing tipping point patterns, the adjustable marketing action parameters or resource allocation parameters associated with them in the marketing plan data to be evaluated are identified as the second type of key parameters.

[0216] Construct a counterfactual reasoning experimental framework that keeps the random seed of the hybrid inference engine and all other input conditions except for the first and second type of key parameters unchanged.

[0217] Within the framework of counterfactual reasoning experiments, the values ​​of the first and second types of key parameters are systematically adjusted to generate multiple candidate marketing plan data variants.

[0218] Multiple candidate marketing plan data variants are converted into corresponding initial action sequences and then sequentially input into the hybrid inference engine for dynamic game simulation, generating multiple sets of candidate simulation trajectory data;

[0219] The simulation trajectory data of multiple candidate projects are evaluated, and the strategy robustness index corresponding to each candidate marketing plan data variant is calculated. The strategy robustness index comprehensively evaluates the expected market share, the probability of risk exposure, and the potential for word-of-mouth explosion.

[0220] Select the candidate marketing campaign data variant with the best strategy robustness index and identify it as the optimized marketing campaign data.

[0221] In this embodiment, high-frequency and high-risk competitive vulnerability patterns are determined by a comprehensive ranking based on their frequency of occurrence in the competitive vulnerability map and the confidence weight of the corresponding edge; high-probability and high-leverage potential word-of-mouth dissemination tipping point patterns are determined by a comprehensive ranking based on their conditional probability and leverage condition parameters in the word-of-mouth tipping point probability cloud map.

[0222] The identification of the first and second categories of key parameters is achieved by establishing a mapping relationship between the model and the elements of the marketing plan. For example, the self-action codes involved in the competitive vulnerability model are directly linked to the specific parameters of the corresponding marketing activities in the marketing plan data to be evaluated, such as the price adjustment range and the intensity of promotional activities; the early key events or environmental characteristics involved in the word-of-mouth tipping point model are linked to marketing action parameters (such as specific advertising content attributes) or resource allocation parameters (such as the budget proportion for a specific user group) that may affect the event or characteristic.

[0223] When constructing the counterfactual reasoning experimental framework, the random seed of the hybrid inference engine is fixed to ensure that all simulated random processes are reproducible except for the key parameters; at the same time, all other contents in the marketing plan data to be evaluated, except for the first and second types of key parameters, are kept unchanged, so that the differences in simulation results are attributed to the adjustment of key parameters.

[0224] The values ​​of the first and second types of key parameters are systematically adjusted using a parameter scanning method. A reasonable value range and step size are defined for each key parameter. By traversing the combinations of parameter value ranges, or by using more efficient experimental design methods (such as Latin hypercube sampling), a set of candidate marketing plan data variants covering the parameter space is generated.

[0225] Each candidate marketing plan data variant is transformed into a corresponding initial action sequence according to the conversion method defined in the aforementioned embodiments, and then sequentially input into the hybrid inference engine to re-execute the dynamic game simulation, generating multiple sets of candidate simulation trajectory data corresponding to each variant.

[0226] Multiple sets of candidate simulated trajectory data are evaluated to calculate a strategy robustness index corresponding to each candidate marketing plan data variant. This index is obtained by weighted aggregation of multiple sub-indicators: expected market share is calculated from the average market share at the end of the simulation; probability of risk exposure is calculated from the proportion of trajectories in the simulation that exhibit competitive vulnerability patterns or damaged brand health; and potential for word-of-mouth explosion is calculated from the proportion of trajectories in the simulation whose volume growth exceeds the explosion threshold and their average growth rate. The weights of each sub-indicator are set according to business objectives.

[0227] This embodiment identifies key risk and opportunity patterns from the insight report and associates them with specific adjustable parameters, constructing a counterfactual experimental environment with controlled variables. By systematically adjusting these parameters and re-simulating and evaluating, the impact of different parameter adjustments on the robustness of the overall strategy can be quantified. This allows for data-driven identification of optimized solutions to avoid major risks and utilize key opportunities, achieving a closed loop from risk insight to strategy improvement.

[0228] Please see Figure 3 In a second aspect, this embodiment also provides a marketing risk pre-simulation system 1 based on dynamic simulation, applicable to the method described in the first aspect. The system includes a data receiving and processing module 11, an environment and decision model construction module 12, a scheme transformation and hybrid inference module 13, a trajectory analysis and insight generation module 14, and an iterative optimization and output module 15. The data receiving and processing module 11 is used to receive marketing scheme data to be evaluated, historical market environment data, and multi-source social network data. The environment and decision model construction module 12 is used to construct a multi-agent interaction environment model containing competitor entities, user circle entities, and media opinion entities based on historical market environment data and multi-source social network data, and to configure parameterized decision models for competitor entities, user circle entities, and media opinion entities respectively. The scheme transformation and hybrid inference module 13 is used to transform the marketing scheme data to be evaluated into the initial action sequence of the local entity in the multi-agent interaction environment model, and... Based on a multi-agent interaction environment model and a parameterized decision-making model, a hybrid inference engine combining Monte Carlo simulation and multi-agent reinforcement learning is used to execute multiple dynamic game simulations in parallel, generating multiple simulated trajectory data containing environmental state evolution and entity action sequences. The trajectory analysis and insight generation module 14 is used to aggregate and analyze multiple simulated trajectory data, identify and quantify the competitive vulnerability patterns and word-of-mouth initiation point patterns triggered by the marketing plan data to be evaluated in the dynamic game simulation, and generate a pre-simulation insight report containing a competitive vulnerability map and a word-of-mouth initiation point probability cloud map. The iterative optimization and output module 15 is used to iteratively optimize the key parameters in the marketing plan data to be evaluated based on the pre-simulation insight report through counterfactual reasoning, generate optimized marketing plan data, and re-input the optimized marketing plan data into the hybrid inference engine for dynamic game simulation and evaluation until the predetermined strategy robustness conditions are met, and output the final risk assessment and strategy optimization suggestions.

[0229] In this embodiment, the data receiving and processing module 11 acquires raw data; the environment and decision model construction module 12 establishes a realistic multi-agent game environment based on this data; the scheme transformation and hybrid inference module 13 injects specific marketing schemes into this environment and infers a large number of possible future market trajectories through parallel simulation; the trajectory analysis and insight generation module 14 automatically extracts key risk and opportunity patterns from these trajectory data; and the iterative optimization and output module 15 uses these insights to automatically adjust the scheme parameters and re-evaluate them in a counterfactual manner, iterating cyclically until a robust strategy scheme is obtained. This system integrates complex market game simulation, risk quantification analysis, and strategy optimization processes, providing an efficient and automated pre-simulation and optimization tool for marketing decisions.

[0230] By adopting the above technical solutions, this invention differs from existing technologies and possesses the following beneficial effects: By constructing a multi-agent interactive environment model and configuring a parameterized decision-making model, the market is abstracted into a dynamic game system, providing a realistic foundation for simulation; a hybrid inference engine is used to execute a large number of parallel dynamic game simulations, generating evolutionary trajectories covering multiple market scenarios, achieving a comprehensive inference of the chain reaction of marketing plans; through the aggregation analysis of simulation trajectories, competitive vulnerability patterns and word-of-mouth dissemination tipping point patterns are automatically identified and quantified, and a visual insight map is generated, transforming complex data into intuitive risk and opportunity insights. Based on this insight, key parameters of the plan are iteratively optimized through counterfactual reasoning, outputting an optimized strategy and risk assessment verified by simulation. This invention achieves a forward-looking, quantitative assessment and proactive optimization of the potential chain risks of marketing plans in a complex dynamic game environment, significantly improving the scientific nature, risk resistance, and strategic robustness of marketing decisions.

[0231] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A marketing risk prediction method based on dynamic simulation, characterized in that, include: Receive marketing campaign data to be evaluated, historical market environment data, and multi-source social network data; Based on the historical market environment data and the multi-source social network data, a multi-agent interaction environment model is constructed, which includes competitor entities, user circle entities, and media opinion entities. Parameterized decision models are then configured for the competitor entities, user circle entities, and media opinion entities, respectively. The marketing plan data to be evaluated is transformed into the initial action sequence of the local entity in the multi-agent interaction environment model. Based on the multi-agent interaction environment model and the parameterized decision model, multiple dynamic game simulations are executed in parallel through a hybrid inference engine of Monte Carlo simulation and multi-agent reinforcement learning to generate multiple simulated trajectory data containing environmental state evolution and entity action sequences. By aggregating and analyzing multiple simulated trajectory data, the competitive vulnerability patterns and word-of-mouth initiation point patterns triggered by the marketing plan data to be evaluated in the dynamic game simulation are identified and quantified, and a pre-simulation insight report containing competitive vulnerability maps and word-of-mouth initiation point probability cloud maps is generated. Based on the pre-simulation insight report, the key parameters in the marketing plan data to be evaluated are iteratively optimized through counterfactual reasoning to generate optimized marketing plan data. The optimized marketing plan data is then re-input into the hybrid inference engine for dynamic game simulation and evaluation until the predetermined strategy robustness conditions are met, and the final risk assessment and strategy optimization suggestions are output.

2. The marketing risk prediction method based on dynamic simulation according to claim 1, characterized in that, Based on the historical market environment data and the multi-source social network data, a multi-agent interaction environment model is constructed, including competitor entities, user circle entities, and media opinion entities, comprising: The historical market environment data is used to extract and quantify entity relationships, generating a competitive relationship intensity matrix and market resource constraint parameters; Graph structure modeling and community discovery are performed on the multi-source social network data to generate user social network graphs and cross-platform influence network graphs. Nodes in the user social network graphs are labeled with multiple user circle tags divided by clustering algorithms. Based on the competitive relationship strength matrix, a competitive ecosystem graph is constructed with the competitor entities as nodes, and a resource constraint model including budget, production capacity and channel scheduling is configured for the competitor entities. Based on the user social network graph and the multiple user circle tags, a group behavior model is constructed with the user circle entities as agents. The group behavior model integrates social network propagation dynamics and behavioral economics theory to simulate the process of individual decision-making being influenced by social factors and cognitive biases. Based on the cross-platform influence network graph, an influence propagation model for the media opinion entity is constructed, wherein the influence propagation model encodes the characteristic differences of different information dissemination platforms and the position preference parameters of the media opinion entity. The competitive ecosystem map, the group behavior model, and the influence propagation model are coupled to form the multi-agent interaction environment model, which defines the state perception and interaction rules between the competitor entity, the user circle entity, and the media opinion entity.

3. The marketing risk prediction method based on dynamic simulation according to claim 2, characterized in that, The historical market environment data is used to extract and quantify entity relationships, generating a competitive relationship strength matrix and market resource constraint parameters, including: Extract historical marketing campaign sequences, historical market share time-series data, and historical brand mention co-occurrence data from the historical market environment data; Based on the aforementioned historical marketing activity sequence, the mutual response intensity of marketing actions between different brand entities is calculated using a causal inference algorithm, serving as the first quantitative indicator of the intensity of competitive relationships. Based on the historical market share time series data, the negative correlation between the changes in market share of different brand entities is calculated through time series correlation analysis, which serves as a second quantitative indicator of the intensity of competitive relationship. Based on the historical brand co-occurrence data, the intensity of the opposing sentiment association between different brand entities in the public opinion field is calculated through sentiment analysis in natural language processing, which serves as the third quantitative indicator of the intensity of competitive relationship. The first quantitative indicator of the intensity of the competitive relationship, the second quantitative indicator of the intensity of the competitive relationship, and the third quantitative indicator of the intensity of the competitive relationship are combined and weighted to generate the competitive relationship intensity matrix; The historical resource input data of each brand entity is analyzed from the historical market environment data, and the market resource constraint parameters are obtained through statistical regression fitting. The market resource constraint parameters represent the constraint relationship between resource input and marketing effect output.

4. The marketing risk prediction method based on dynamic simulation according to claim 2, characterized in that, And parameterized decision models are configured for the competitor entity, the user circle entity, and the media opinion entity, respectively, including: A first parameterized decision model is configured for the competitor entity. The first parameterized decision model includes a policy network constructed based on a deep reinforcement learning algorithm. The input of the policy network is an environmental state vector containing its own state, the competitor's state, and the overall market state. The output of the policy network is the action probability distribution or deterministic action value of the competitor entity in the action space. The objective function of the first parameterized decision model is to maximize the long-term cumulative reward. The long-term cumulative reward is calculated by combining changes in market share, changes in brand equity, and action costs. A second parameterized decision model is configured for the user circle entity. The second parameterized decision model includes a decision function constructed based on the intelligent agent modeling framework and behavioral economics theory. The input of the decision function is an individual attribute vector, a local social network state vector, and a received information stimulus vector. The decision function simulates individual decision-making by calculating a comprehensive utility value and introducing random noise. The comprehensive utility value integrates personal preference utility, social conformity utility, and risk perception utility under the prospect theory framework. A third parameterized decision model is configured for the media opinion entity. The third parameterized decision model includes a content selection function constructed based on the utility maximization theory. The input of the content selection function is a vector of candidate content attributes, a vector of audience matching degree, and a vector of potential commercial incentives. The content selection function calculates the expected influence return and commercial return by weighting, and selects the content with the highest comprehensive return to perform the dissemination action.

5. The marketing risk prediction method based on dynamic simulation according to claim 2, characterized in that, Graph structure modeling and community discovery are performed on the multi-source social network data to generate user social network graphs and cross-platform influence network graphs, including: The user interaction behavior data and content association data in the multi-source social network data are cleaned and aligned to identify unique user nodes and content nodes. Based on the user interaction behavior data, an initial social network graph is constructed with user nodes as vertices and follow relationships, forwarding relationships, or co-occurrence relationships as edges; A graph embedding algorithm is applied to the initial social network graph to obtain a low-dimensional vector representation of the user nodes; Based on the low-dimensional vector representation of the user nodes, a community detection algorithm is used to cluster the user nodes, the clustering results are labeled as multiple user circle labels, and the multiple user circle labels are mapped back to the initial social network graph to generate the user social network graph with circle labels. The cross-platform account association data and content dissemination path data in the multi-source social network data are integrated to identify media opinion leader nodes and their influence range on different information dissemination platforms; Based on the content dissemination path data, a bipartite graph is constructed with media opinion leader nodes and platforms as vertices and cross-platform content migration relationships as edges; The bipartite graph is projected and weights are calculated to generate the cross-platform influence network graph. The edge weights in the cross-platform influence network graph represent the probability or intensity of information migration between platforms.

6. The marketing risk prediction method based on dynamic simulation according to claim 1, characterized in that, The marketing plan data to be evaluated is transformed into the initial action sequence of the local entity in the multi-agent interaction environment model, including: Analyze the marketing plan data to be evaluated, and extract the core marketing action elements, resource allocation plan and execution timeline contained in the marketing plan data to be evaluated; The core marketing action elements are mapped to standard action codes in the action space that the entity can execute in the multi-agent interaction environment model. The standard action codes include promotion action codes, advertising action codes, channel strategy action codes, and public opinion guidance action codes. Based on the resource allocation plan, allocate corresponding resource quantization parameters to the mapped standard action codes; Based on the execution timeline, the standard action codes that allocate the resource quantification parameters are arranged and combined in chronological order to generate the initial action sequence.

7. The marketing risk prediction method based on dynamic simulation according to claim 1, characterized in that, Based on the multi-agent interaction environment model and the parameterized decision model, a hybrid inference engine combining Monte Carlo simulation and multi-agent reinforcement learning is used to execute multiple dynamic game simulations in parallel, generating multiple simulated trajectory data containing environmental state evolution and entity action sequences, including: Initialize the hybrid simulation engine, and set the total number of parallel simulation instances and the maximum simulation time step for each simulation instance; An independent random seed is generated for each parallel simulation instance, and the initial state of other entities in the multi-agent interaction environment model, excluding the local entity, is initialized based on the independent random seed. Within each time step of each simulation instance, perform the following steps: The first parameterized decision model of the competitor entity is invoked to generate competitive actions of the competitor entity based on the current environmental state; The second parameterized decision model of the user circle entity is invoked to generate the group behavior actions of the user circle entity based on the current social network state and the intensity of information stimuli. The third parameterized decision model of the media opinion entity is invoked to generate the dissemination action of the media opinion entity based on the current content ecosystem and commercial incentives; The actions of the local entity, the competitive actions, the group behavior actions, and the propagation actions are input into the multi-agent interaction environment model. Based on the state transition rules defined in the multi-agent interaction environment model, the environment state at the next time step is calculated and updated. During the environmental state update process, it is checked whether the preset key event triggering conditions are met. If they are met, the corresponding key event rule is activated and the subsequent state transition rules are modified. Record the environmental state, the actions of all entities, and the key events triggered at the current time step to form a trajectory segment at the current time step; When the simulation instance reaches the maximum simulation time step or meets the simulation termination condition, the simulation instance ends, and all trajectory segments connected in chronological order in the simulation instance are summarized into complete simulation trajectory data. Collect complete simulation trajectory data generated by all parallel simulation instances to form the multiple simulation trajectory data containing environmental state evolution and entity action sequences.

8. The marketing risk prediction method based on dynamic simulation according to claim 1, characterized in that, Aggregate and analyze the multiple simulated trajectory data to identify and quantify the competitive vulnerability patterns and word-of-mouth tipping point patterns triggered by the marketing plan data to be evaluated in the dynamic game simulation. Generate a pre-simulation insight report containing a competitive vulnerability map and a word-of-mouth tipping point probability cloud map, including: Filter through multiple simulated trajectory data and extract negative outcome trajectory data from all simulated trajectory data, such as the final decline in market share or damage to brand health of our entity; Time-series pattern mining is performed on the negative outcome trajectory data to locate frequent co-occurring action pairs between marketing actions initiated by our entity and adversarial actions taken by competitors within the time window before a significant decline in market share. Calculate the support and confidence of each action pair combination in all negative result trajectory data, and identify action pair combinations whose support and confidence both exceed a preset threshold as the competitive vulnerability pattern; Based on the aforementioned competitive vulnerability pattern, a visual annotation is performed on the competitive ecosystem map in the multi-agent interaction environment model to generate the competitive vulnerability map. The competitive vulnerability map displays typical attack paths from one's own actions to the competitor's counterattack and their frequency of occurrence. The multiple simulated trajectory data are filtered to extract the positive result trajectory data in which the total sound volume increases beyond the preset burst threshold. Feature backtracking analysis is performed on the positive result trajectory data to extract common initial environmental state features and early key event features that existed before the sound volume explosion, forming a potential trigger point feature set. Calculate the conditional probability of a sound burst event occurring given that each feature combination in the potential detonation point feature set appears in the entire simulated trajectory dataset; The feature combination whose conditional probability exceeds a preset probability threshold is identified as the word-of-mouth dissemination tipping point pattern, and the leverage condition parameter corresponding to the word-of-mouth dissemination tipping point pattern is calculated. Based on the word-of-mouth dissemination tipping point pattern and its conditional probability and leverage condition parameters, the user social network graph and cross-platform influence network graph in the multi-agent interaction environment model are visualized and rendered to generate the word-of-mouth tipping point probability cloud map. The competitive vulnerability map and the word-of-mouth tipping point probability cloud map are integrated, and statistical summaries and key findings descriptions are added to generate the pre-analysis insight report.

9. The marketing risk prediction method based on dynamic simulation according to claim 1, characterized in that, Based on the aforementioned pre-performance insight report, key parameters in the marketing plan data to be evaluated are iteratively optimized using counterfactual reasoning to generate optimized marketing plan data, including: The pre-rehearsal insight report is analyzed to identify high-frequency and high-risk competitive vulnerability patterns from the competitive vulnerability map, and high-probability and high-leverage potential word-of-mouth dissemination tipping point patterns are identified from the word-of-mouth tipping point probability cloud map. For the identified competitive vulnerability patterns, the adjustable marketing action parameters associated with them in the marketing program data to be evaluated are determined as the first type of key parameters; For the identified word-of-mouth dissemination tipping point pattern, the adjustable marketing action parameters or resource allocation parameters associated with it in the marketing plan data to be evaluated are determined as the second type of key parameters; Construct a counterfactual reasoning experimental framework, wherein the counterfactual reasoning experimental framework keeps the random seed of the hybrid inference engine and all other input conditions except for the first type of key parameters and the second type of key parameters unchanged; Within the counterfactual reasoning experimental framework, the values ​​of the first type of key parameters and the second type of key parameters are systematically adjusted to generate multiple candidate marketing scheme data variants; The multiple candidate marketing scheme data variants are converted into corresponding initial action sequences, and then sequentially input into the hybrid inference engine for dynamic game simulation to generate multiple sets of candidate simulation trajectory data. The multiple sets of candidate simulated trajectory data are evaluated, and the strategy robustness index corresponding to each candidate marketing plan data variant is calculated. The strategy robustness index comprehensively evaluates the expected market share, the probability of risk exposure, and the potential for word-of-mouth explosion. The candidate marketing plan data variant with the best robustness index is selected and identified as the optimized marketing plan data.

10. A marketing risk prediction system based on dynamic simulation, characterized in that, The system applicable to the method of any one of claims 1 to 9 comprises: The data receiving and processing module is used to receive marketing plan data to be evaluated, historical market environment data, and multi-source social network data. The environment and decision model construction module is used to construct a multi-agent interaction environment model containing competitor entities, user circle entities, and media opinion entities based on the historical market environment data and the multi-source social network data, and to configure parameterized decision models for the competitor entities, user circle entities, and media opinion entities respectively. The scheme transformation and hybrid inference module is used to transform the marketing scheme data to be evaluated into the initial action sequence of the entity in the multi-agent interaction environment model. Based on the multi-agent interaction environment model and the parameterized decision model, the module executes multiple dynamic game simulations in parallel through a hybrid inference engine of Monte Carlo simulation and multi-agent reinforcement learning to generate multiple simulated trajectory data containing environmental state evolution and entity action sequences. The trajectory analysis and insight generation module is used to aggregate and analyze the multiple simulated trajectory data, identify and quantify the competitive vulnerability patterns and word-of-mouth initiation point patterns triggered by the marketing plan data to be evaluated in the dynamic game simulation, and generate a pre-simulation insight report containing a competitive vulnerability map and a word-of-mouth initiation point probability cloud map. The iterative optimization and output module is used to iteratively optimize the key parameters in the marketing plan data to be evaluated based on the pre-simulation insight report through counterfactual reasoning, generate optimized marketing plan data, and re-input the optimized marketing plan data into the hybrid inference engine for dynamic game simulation and evaluation until the predetermined strategy robustness conditions are met, and output the final risk assessment and strategy optimization suggestions.

Images