An ai-based marketing resource allocation method and system

Abstract

The application discloses a marketing resource allocation method and system based on AI. First, multi-dimensional data of users is acquired, and a user value dynamic evaluation model is input to calculate a comprehensive value score. A performance prediction result is generated by using a revenue prediction and resource consumption rate prediction model. An initial sorting list is generated based on the comprehensive value score and the performance prediction result. Natural conversion users are eliminated by a cause-effect prediction model, and backflow resources are returned to a global resource pool. A multi-objective resource allocation model is constructed, a decay penalty mechanism based on historical violations is introduced to generate an allocation strategy. Online data is collected in real time, KL divergence is calculated, and when the KL divergence exceeds a preset threshold, time window data is traced back and the model is updated. The application can accurately identify high-potential users, realize resource closed-loop optimization and model adaptive updating.

Description

Technical Field

[0001] This invention relates to the field of marketing resource allocation technology, and in particular to an AI-based marketing resource allocation method and system. Background Technology

[0002] Traditional marketing resource allocation methods primarily rely on manual experience rules or automated tools based on simple ranking. They typically use conversion probability or historical spending as the sole basis for allocation, neglecting multi-dimensional factors such as long-term user value and social influence. They also struggle to distinguish between "naturally converting users" and "marketing-sensitive users," resulting in a significant waste of marketing resources on users who would convert even without intervention. Furthermore, most existing methods use static models for decision-making, failing to recognize data distribution shifts caused by changes in the online environment (such as user behavior migration and channel effectiveness decline), leading to severe performance degradation over time. Simultaneously, traditional methods often prioritize short-term profit maximization in resource allocation, ignoring risks such as over-concentration of resources and long-term value loss, and lacking embedded compliance constraints. Therefore, achieving accurate user value assessment, scientific screening of marketing-sensitive users, global optimization of resource allocation, and adaptive model updates within a limited budget are pressing technical challenges in the current marketing technology field. Summary of the Invention

[0003] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide an AI-based marketing resource allocation method and system to solve the problems of the prior art, such as the single dimension of user value assessment, inability to distinguish between natural conversion and marketing-sensitive users, and lack of global optimization in resource allocation.

[0004] To achieve the above and other related objectives, the present invention provides the following technical solution:

[0005] A marketing resource allocation method based on AI includes the following steps: S1, acquiring multi-dimensional behavioral data and social relationship data of target users, and inputting them into a dynamic user value evaluation model; the dynamic user value evaluation model integrates the user's lifecycle stage, conversion probability, customer lifetime value, real-time behavioral activity, and social network dissemination index to calculate the comprehensive value score of each user; S2, using a revenue prediction model and a resource consumption rate prediction model, predicting the expected revenue of each user and the unit resource consumption rate of the corresponding channel, respectively, and generating an efficiency prediction result; S3, generating an initial ranking list based on the comprehensive value score and efficiency prediction result, using a causal prediction model to perform counterfactual incremental estimation of users in the list, removing users whose natural conversion probability is higher than the conversion threshold, filtering out a high-potential user subset, and removing those who were previously excluded. The resources occupied by the user are returned to the global resource pool; S4, a multi-objective resource allocation model is constructed, with the basic allocation time slice as the decision unit, the optimization objective is to maximize the comprehensive value score and minimize the resource consumption cost, and a decay penalty mechanism based on historical violation records is introduced as a constraint to generate a marketing resource allocation strategy; S5, resource delivery is executed according to the allocation strategy, and online execution data is collected in real time; the KL divergence between the distribution of online execution data and the distribution of historical sample data used in training the causal prediction model and the user value dynamic evaluation model is calculated. When the KL divergence exceeds a preset threshold, several preset time slices are backtracked based on the current time slice, and the multi-objective resource allocation model, the user value dynamic evaluation model, and the causal prediction model are updated by gradient descent using the real feedback data in the window.

[0006] Furthermore, the social network propagation index is generated in the following way: a user social relationship graph is constructed, mapping users as nodes and the attention, interaction or transaction relationships between users as edges; the topological feature indicators of each node in the graph are calculated, including degree centrality (measuring the number of direct connections), proximity centrality (measuring the efficiency of information transmission), and betweenness centrality (measuring the bridging effect); based on the topological feature indicators, the social network propagation index, which is the potential for viral spread when a user initiates a marketing campaign, is calculated using a nonlinear mapping function.

[0007] Further, step S3 includes: using a pre-trained causal prediction model, predicting the expected conversion result for each candidate user in the initial ranking list under two conditions: receiving marketing intervention and not receiving marketing intervention, and taking the difference between the two as the individual treatment effect; setting an incremental threshold, retaining only users whose individual treatment effect is greater than the incremental threshold into the high-potential user subset, while judging users whose individual treatment effect is less than or equal to the incremental threshold as natural conversion users and removing them; calculating the total amount of resources originally planned to be allocated to the removed users, and returning this part of the resources to the global resource pool to supplement the resource quota of the high-potential user subset.

[0008] Furthermore, the decay penalty mechanism specifically includes: pre-establishing a historical violation record database to store each event of excessive resource concentration or long-term value destruction and its time slice; when calculating the penalty value of the current allocation strategy, using the current time slice as a benchmark, traversing all past time slices, and calculating an exponential decay weight for each time slice with a violation at a time interval; multiplying the decay weight of each time slice by the violation quantification value of that time slice and summing the results to obtain the total penalty value; and adding this total penalty value as a negative reward item to the objective function of the resource allocation model.

[0009] Furthermore, the objective function of the multi-objective resource allocation model is constructed as follows:

[0010]

[0011] in, The objective function is... Assign tensors to 3D resources. Represents a set of users The total number of users in the region Represents a set of channels The total number of channels in the middle, Represents the set of time slices The total number of time slices in the game; Indicates time slice Through channels Assigned to user The amount of resources and meet the requirements , , This represents the current total budget for the global resource pool. For users The overall value score; For channels The unit resource cost; For dynamic trade-off coefficients; For the first The quantification of the severity of each violation; The time slice in which it occurs; This is the current time slice; The attenuation coefficient; To iterate through all events occurring in the current time slice Previous violations.

[0012] Furthermore, the KL divergence detection and model update mechanism in step S5 specifically includes defining the online real-time data distribution. With training data distribution The KL divergence between them is,

[0013]

[0014] in, Indicates the KL divergence value; This indicates the data distribution generated in the current online real-time environment; This indicates the distribution of historical sample data used during the model training phase; This indicates that the integration operation is performed over the entire feature space; This indicates the distribution of real-time online data. At feature points The probability density function value at that location; Represents the distribution of training data At feature points The probability density function value at that location; This represents the input feature vector; The log-likelihood ratio measures the likelihood of similar eigenvalues. Below, the probability difference between the real-time distribution and the training distribution; when When the divergence exceeds the divergence threshold, concept drift is determined to have occurred, triggering the model update mechanism.

[0015] Furthermore, this invention provides an AI-based marketing resource allocation system comprising: a data acquisition module for acquiring user behavior data, social relationship data, and resource allocation feedback; a user value assessment module for integrating user lifecycle stages, conversion probability, customer lifetime value, behavioral activity, and social influence index to calculate a comprehensive user value score; an effectiveness prediction module for predicting expected user returns and channel resource consumption rates; a causal prediction model for estimating the individual treatment effect of a user receiving marketing intervention and outputting the individual treatment effect to a ranking and initial screening module; a ranking and initial screening module for selecting a subset of high-potential users based on the comprehensive value score and individual treatment effect, and executing resource return; a multi-objective allocation optimization module for generating an allocation strategy by introducing a decay penalty mechanism based on historical violations, with the objectives of maximizing the comprehensive value score and minimizing resource costs; a distribution offset detection module for calculating the KL divergence between the online and training data distributions and determining whether it exceeds a dynamic threshold; an online update module for backtracking time window data and updating the model through gradient descent when the KL divergence exceeds the divergence threshold; and a strategy execution module for allocating resources according to the allocation strategy and providing feedback on the execution results.

[0016] Furthermore, the user value assessment module includes a lifecycle analysis unit, a conversion probability prediction unit, a customer lifetime value calculation unit, a behavior activity analysis unit, a social influence analysis unit, and a weighted fusion unit; the social influence analysis unit is used to construct a user social relationship graph and generate a social network propagation index based on the degree centrality, proximity centrality, and betweenness centrality of the graph.

[0017] Furthermore, the initial ranking and screening module includes an initial ranking generation unit, a causal effect comparison unit, a resource return unit, and a resource transfer unit; the initial ranking generation unit generates an initial ranking list based on the comprehensive value score and the effectiveness prediction result; the causal effect comparison unit compares the individual processing effect of each user with a preset incremental threshold and removes users whose effect is not greater than the incremental threshold; the resource return unit counts the planned resource amount of the removed users and releases it to the global resource pool.

[0018] The resource transfer unit redistributes the returned resources to high-potential users in descending order of individual processing effect.

[0019] Furthermore, the present invention provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method or system of the present invention.

[0020] As described above, the AI-based marketing resource allocation method and system of this invention has the following beneficial effects: This method achieves accurate quantitative evaluation of the comprehensive value of users by integrating multi-dimensional indicators such as user lifecycle stages, conversion probability, customer lifetime value, real-time behavioral activity, and social network dissemination index, breaking through the limitations of traditional methods that rely on only a single indicator. Introducing a causal prediction model for individual treatment effect estimation can accurately distinguish between "marketing-sensitive users" and "naturally converting users," effectively avoiding the waste of marketing resources. Furthermore, a resource return mechanism enables the redistribution of resources from inefficient users to efficient users, improving the overall return on investment without increasing the total budget. A multi-objective resource allocation model is constructed with the goal of maximizing comprehensive value and minimizing resource costs. An exponential decay penalty mechanism based on historical violation records is introduced, embedding compliance constraints into the allocation strategy while pursuing returns, achieving a dynamic balance between value, cost, and risk. In addition, by monitoring the KL divergence between online data distribution and training data distribution in real time, concept drift can be detected in a timely manner and online model updates can be triggered, enabling the system to automatically adapt to environmental changes and significantly improving the long-term stability and effectiveness of marketing strategies. Overall, this technical solution provides an end-to-end intelligent solution for marketing resource allocation, which is accurate, adaptive and maintainable. Attached Figure Description

[0021] Figure 1 The diagram shown is a flowchart of the method of the present invention. Detailed Implementation

[0022] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be noted that, unless otherwise specified, the following embodiments and features described herein can be combined with each other.

[0023] Please see Figure 1 The first embodiment of the present invention provides an AI-based marketing resource allocation method, comprising the following steps: S1, acquiring multi-dimensional behavioral data and social relationship data of target users, and inputting them into a dynamic user value evaluation model; the dynamic user value evaluation model integrates the user's lifecycle stage, conversion probability, customer lifetime value, real-time behavioral activity, and social network dissemination index to calculate the comprehensive value score of each user; S2, using a revenue prediction model and a resource consumption rate prediction model, predicting the expected revenue of each user and the unit resource consumption rate of the corresponding channel, respectively, and generating an efficiency prediction result; S3, generating an initial ranking list based on the comprehensive value score and efficiency prediction result, using a causal prediction model to perform counterfactual incremental estimation of users in the list, eliminating users whose natural conversion probability is higher than the conversion threshold, and selecting high-potential user sub-categories. S4. Collect and return the resources occupied by the removed users to the global resource pool; S5. Construct a multi-objective resource allocation model, using the basic allocation time slice as the decision unit, with the optimization objective being to maximize the comprehensive value score and minimize the resource consumption cost, and introduce a decay penalty mechanism based on historical violation records as a constraint condition to generate a marketing resource allocation strategy; S6. Execute resource placement according to the allocation strategy and collect online execution data in real time; calculate the KL divergence between the distribution of online execution data and the distribution of historical sample data used when training the causal prediction model and the user value dynamic evaluation model. When the KL divergence exceeds a preset threshold, backtrack several preset time slices based on the current time slice, and use the real feedback data within the window to update the multi-objective resource allocation model, the user value dynamic evaluation model, and the causal prediction model through gradient descent.

[0024] This method first acquires multi-dimensional behavioral and social relationship data of target users, inputs it into a dynamic user value evaluation model, and integrates user lifecycle stage, conversion probability, customer lifetime value, real-time behavioral activity, and social network dissemination index to calculate the comprehensive value score for each user. Simultaneously, it uses a revenue prediction model and a resource consumption rate prediction model to predict the expected revenue for each user and the unit resource consumption rate of the corresponding channel, generating performance prediction results. Based on this, an initial ranking list is generated based on the comprehensive value score and performance prediction results. A causal prediction model is used to perform counterfactual incremental estimation on the users in the list, removing users whose natural conversion probability is higher than the conversion threshold, and returning the returned resources to the global resource pool. Subsequently, a multi-objective resource allocation model is constructed with basic allocation time slices as the decision unit, aiming to maximize the comprehensive value score and minimize resource consumption costs. A decay penalty mechanism based on historical violation records is introduced to generate a marketing resource allocation strategy and execute the campaign. During execution, online data is collected in real time, and the KL divergence between the online data distribution and the training data distribution is calculated. When the divergence exceeds a preset threshold, a preset number of time slices are backtracked, and the multi-objective resource allocation model, user value dynamic evaluation model, and causal prediction model are updated through gradient descent using real feedback data within the window, so as to achieve online adaptive updating of the model.

[0025] This method integrates multi-dimensional user value metrics with causal inference techniques to accurately identify high-potential users and eliminate naturally converted users, effectively avoiding wasted marketing resources and improving resource utilization efficiency. It employs a multi-objective resource allocation model to achieve a dynamic balance between maximizing overall user value and minimizing resource costs, and introduces a decay penalty mechanism based on historical violation records to enhance the compliance and long-term stability of the allocation strategy. Furthermore, by monitoring the KL divergence between online data distribution and training data distribution in real time, it can promptly detect concept drift and trigger online model updates, enabling the resource allocation strategy to continuously adapt to environmental changes and significantly improving the marketing system's adaptability and long-term campaign effectiveness.

[0026] The social network propagation index is generated in the following way: a user social relationship graph is constructed, mapping users as nodes and the attention, interaction or transaction relationships between users as edges; the topological feature indicators of each node in the graph are calculated, including degree centrality (measuring the number of direct connections), proximity centrality (measuring the efficiency of information transmission), and betweenness centrality (measuring the bridging effect); based on the topological feature indicators, the social network propagation index, which is the potential for viral spread when a user initiates a marketing campaign, is calculated using a nonlinear mapping function.

[0027] The social network propagation power index is generated as follows: First, a user social relationship graph is constructed, mapping each user to a node in the graph and the attention, interaction, or transaction relationships between users to edges between nodes, forming a graph model reflecting the structure of the user's social network. Then, multiple topological characteristic indicators for each node in the graph are calculated, including: degree centrality (measuring the number of direct connections between users, reflecting the user's local social activity); proximity centrality (measuring the average path length from a user to other nodes, reflecting the user's information transmission efficiency throughout the network); and betweenness centrality (measuring the frequency of a user acting as a bridge on the shortest path, reflecting the user's bridging role in information dissemination). Finally, based on these multiple topological characteristic indicators, a nonlinear mapping function (e.g., weighted summation followed by activation function transformation such as Sigmoid or ReLU) is used to calculate the potential for each user to trigger viral spread when initiating a marketing campaign, ultimately yielding the social network propagation power index, which serves as an input dimension for a dynamic user value evaluation model.

[0028] By introducing a social network dissemination power index, this method quantifies and integrates a user's topological influence within their social network into a value assessment system, overcoming the limitations of traditional marketing resource allocation that relies solely on individual user historical behavior. Specifically, degree centrality identifies socially active "opinion leaders," proximity centrality identifies "efficient disseminators" with the shortest information dissemination paths, and betweenness centrality identifies "bridging nodes" connecting different communities. The combination of these three factors comprehensively captures a user's multiple dissemination roles within social networks. Based on these topological characteristics and the dissemination power index generated through nonlinear mapping, this method can more accurately predict the scale of viral spread that user-initiated marketing campaigns can trigger. This allows users with high social influence to receive more reasonable resource allocation, thereby amplifying the viral diffusion effect of marketing campaigns and improving the marginal return on overall marketing investment.

[0029] Step S3 includes: using a pre-trained causal prediction model, predicting the expected conversion result for each candidate user in the initial ranking list under two conditions: receiving marketing intervention and not receiving marketing intervention, and taking the difference between the two as the individual treatment effect; setting an incremental threshold, retaining only users whose individual treatment effect is greater than the incremental threshold into the high-potential user subset, while judging users whose individual treatment effect is less than or equal to the incremental threshold as natural conversion users and removing them; calculating the total amount of resources originally planned to be allocated to the removed users, and returning this part of the resources to the global resource pool to supplement the resource quota of the high-potential user subset.

[0030] This method utilizes a pre-trained causal prediction model to predict the expected conversion outcome for each candidate user in the initial ranking list under two scenarios: receiving marketing intervention and not receiving marketing intervention. The difference between the two scenarios is taken as the user's individual treatment effect, which quantifies the net causal impact of marketing intervention on user conversion behavior. Subsequently, an incremental threshold is set, retaining only users whose individual treatment effect is greater than the threshold into the high-potential user subset. These users are sensitive to marketing intervention, and the intervention can bring significant incremental conversions. Users whose individual treatment effect is less than or equal to the incremental threshold are classified as naturally converting users, i.e., users who would convert regardless of whether they receive marketing intervention, and are thus removed. Finally, the total amount of resources originally planned to be allocated to the removed users is calculated, and these resources are returned to the global resource pool to supplement the resource quota of the high-potential user subset, thereby realizing the redistribution of resources from inefficient users to efficient users.

[0031] By introducing a causal prediction model to estimate individual treatment effects, this method can accurately distinguish between high-potential users who "truly require marketing intervention to convert" and naturally converting users who "would convert regardless of intervention." This effectively avoids the common problem of wasting marketing resources on naturally converting users, improving the accuracy and cost-effectiveness of resource allocation. Specifically, the individual treatment effect directly quantifies the incremental value brought by the intervention, making it a more causal explanatory and business-guiding criterion than traditional methods based on the absolute value of conversion probability. Simultaneously, the planned resources for excluded users are returned to the global resource pool and redistributed to high-potential users, forming a closed-loop resource optimization mechanism. This amplifies the resource supply to high-potential users without increasing the total budget, thereby improving the overall ROI of marketing campaigns.

[0032] The decay penalty mechanism specifically includes: pre-establishing a historical violation record database to store each event of excessive resource concentration or long-term value destruction and its time slice; when calculating the penalty value of the current allocation strategy, using the current time slice as a benchmark, traversing all past time slices, and calculating an exponential decay weight for each time slice with a violation at a time interval; multiplying the decay weight of each time slice by the violation quantification value of that time slice and summing them to obtain the total penalty value; and adding this total penalty value as a negative reward item to the objective function of the resource allocation model.

[0033] This method introduces a decay penalty mechanism based on historical violation records. First, a historical violation record database is pre-established, storing each event of excessive resource concentration (e.g., allocating a large amount of resources to a few users or channels) or long-term value degradation (e.g., continuous investment leading to a deterioration in ROI) and its time slice. When calculating the allocation strategy for the current time slice, the violation records in each past time slice are traversed, using the current time slice as a benchmark. For each time slice with a violation, an exponential decay weight is calculated based on the interval between the current time and the time of the violation event; the larger the time interval, the smaller the decay weight. Then, the decay weight of each violation time slice is multiplied by the violation quantification value of that time slice, and the results of all violation events are summed to obtain the total penalty value. Finally, this total penalty value is added as a negative reward term (i.e., subtracted from the objective function) to the objective function of the resource allocation model, ensuring that allocation strategies that tend to produce violations receive lower objective function values, thereby guiding the model to avoid violations.

[0034] By introducing a decaying penalty mechanism based on historical violation records, this method effectively suppresses high-risk behaviors in marketing resource allocation and improves the compliance and long-term stability of allocation strategies. Specifically, the design of exponential decay weights makes recent violations more severe in penalizing current strategies, while the impact of long-term violations gradually weakens. This ensures that the model "remembers" historical lessons while avoiding excessively persistent negative impacts of past mistakes on current decisions, reflecting the reasonable penalty principle of "heavier penalties for recent causes and lighter penalties for distant causes." Simultaneously, incorporating the total penalty value as a negative reward term into the objective function allows the multi-objective resource allocation model to automatically balance the relationship between maximizing value, minimizing costs, and compliance during optimization. This embeds risk control capabilities at the algorithmic level, eliminating the need for post-implementation human intervention, thereby improving resource utilization efficiency while reducing operational risks.

[0035] The objective function of the multi-objective resource allocation model is constructed as follows:

[0036]

[0037] in, The objective function is... Assign tensors to 3D resources. Represents a set of users The total number of users in the region Represents a set of channels The total number of channels in the middle, Represents the set of time slices The total number of time slices in the game; Indicates time slice Through channels Assigned to user The amount of resources and meet the requirements , , This represents the current total budget for the global resource pool. For users The overall value score; For channels The unit resource cost; For dynamic trade-off coefficients; For the first The quantification of the severity of each violation; The time slice in which it occurs; This is the current time slice; The attenuation coefficient; To iterate through all events occurring in the current time slice Previous violations.

[0038] This method constructs a multi-objective resource allocation model that unifies "maximizing overall value" and "minimizing resource cost" into a quantifiable objective function. It achieves an explicit trade-off between the two objectives through weighted differences, avoiding the complexity of handling multiple objectives separately in traditional multi-objective optimization. Simultaneously, the objective function embeds an exponentially decaying penalty term, ensuring that historical violations have a continuous but time-decreasing negative impact on current allocation decisions. This guarantees the model's memory of past risks while avoiding policy rigidity caused by excessive penalties. By adjusting the dynamic trade-off coefficients… The system can flexibly adapt to the priority requirements of different business scenarios (such as high-value orientation, low-cost orientation, or strong compliance orientation). Furthermore, the introduction of total resource constraints ensures the feasibility of the allocation strategy within a limited budget. Overall, this objective function design provides a calculable, optimizable, and adjustable mathematical framework for marketing resource allocation, capable of generating optimal allocation strategies that balance value, cost, and compliance.

[0039] The KL divergence detection and model update mechanism in step S5 specifically includes defining the online real-time data distribution. With training data distribution The KL divergence between them is,

[0040]

[0041] in, Indicates the KL divergence value; This indicates the data distribution generated in the current online real-time environment; This indicates the distribution of historical sample data used during the model training phase; This indicates that the integration operation is performed over the entire feature space; This indicates the distribution of real-time online data. At feature points The probability density function value at that location; Represents the distribution of training data At feature points The probability density function value at that location; This represents the input feature vector; The log-likelihood ratio measures the likelihood of similar eigenvalues. Below, the probability difference between the real-time distribution and the training distribution; when When the divergence exceeds the divergence threshold, concept drift is determined to have occurred, triggering the model update mechanism.

[0042] By introducing a distribution shift detection mechanism based on KL divergence, this method can automatically detect changes in the online environment (such as shifts in user behavior patterns, fluctuations in the market environment, and declines in channel effectiveness), and promptly trigger model updates when concept drift occurs. This ensures that the marketing resource allocation strategy always adapts to the current data distribution, avoiding performance degradation caused by inconsistencies between the training data and the online data distribution. Compared with periodic forced updates or manually triggered updates, this method has the following advantages: First, KL divergence provides an objective indicator of distribution differences, making update decisions more accurate and avoiding unnecessary computational overhead; second, updates are triggered only when significant drift occurs, balancing model stability and adaptability; finally, gradient descent updates are performed using real feedback data within the backtracking time window, ensuring that the adjustment direction of model parameters is consistent with actual business feedback, thereby improving the long-term prediction accuracy and resource allocation effectiveness of the model in dynamic environments.

[0043] The second embodiment of the present invention provides an AI-based marketing resource allocation method system, comprising: a data acquisition module for acquiring user behavior data, social relationship data, and resource allocation feedback; a user value assessment module for integrating user lifecycle stages, conversion probability, customer lifetime value, behavioral activity, and social dissemination index to calculate a comprehensive user value score; an efficiency prediction module for predicting user expected returns and channel resource consumption rates; a causal prediction model for estimating the individual treatment effect of users receiving marketing interventions and outputting the individual treatment effect to a ranking and initial screening module; a ranking and initial screening module for screening a subset of high-potential users based on the comprehensive value score and individual treatment effect, and executing resource return; a multi-objective allocation optimization module for generating an allocation strategy by introducing a decay penalty mechanism based on historical violations, with the objectives of maximizing the comprehensive value score and minimizing resource costs; a distribution offset detection module for calculating the KL divergence between the online and training data distributions and determining whether it exceeds a dynamic threshold; an online update module for backtracking time window data and updating the model through gradient descent when the KL divergence exceeds the divergence threshold; and a strategy execution module for allocating resources according to the allocation strategy and providing feedback on the execution results.

[0044] This system translates its methodology into a modular technical architecture, facilitating deployment, maintenance, and expansion. Specifically, the data acquisition module and strategy execution module form a complete "perception-decision-execution-feedback" closed loop, enabling the system to continuously learn from the online environment and optimize allocation strategies. The parallel design of the user value assessment module and the performance prediction module allows for independent modeling and then fusion of the comprehensive value and cost-benefit dimensions, enhancing the system's flexibility and interpretability. The collaborative work of the causal prediction model and the initial ranking screening module enables precise user screening based on incremental value, effectively avoiding resource waste. The introduction of the distribution offset detection module and the online update module gives the system the ability to automatically perceive and adapt to environmental changes, ensuring long-term stable operation without frequent manual intervention. Overall, this system provides an end-to-end intelligent solution for marketing resource allocation, combining accuracy, adaptability, and maintainability.

[0045] The user value assessment module includes a lifecycle analysis unit, a conversion probability prediction unit, a customer lifetime value calculation unit, a behavior activity analysis unit, a social influence analysis unit, and a weighted fusion unit. The social influence analysis unit is used to construct a user social relationship graph and generate a social network dissemination index based on the degree centrality, proximity centrality, and betweenness centrality of the graph.

[0046] By finely breaking down the user value assessment module, this system achieves a comprehensive and multi-dimensional quantitative assessment of user value, overcoming the limitations of traditional methods that rely solely on a single indicator (such as historical spending or conversion probability). Specifically, the lifecycle analysis unit ensures differentiated value assessment strategies for users at different stages; the conversion probability prediction unit, combined with the customer lifetime value calculation unit, considers both short-term conversion opportunities and long-term value contribution; the behavioral activity analysis unit captures real-time changes in user intent, making the assessment results more timely; and the social influence analysis unit quantifies a user's dissemination potential in social networks through graph topology features, expanding individual value into structural value within the social network, enabling users with high social influence to be rationally identified and allocated more resources. The output components of each unit are flexibly combined through a weighted fusion unit, allowing the system to dynamically adjust weight coefficients based on business objectives (such as new user acquisition, activation, conversion, or brand promotion), aligning the assessment strategy with business goals. Overall, this modular design provides a more accurate, comprehensive, and configurable user value measurement foundation for marketing resource allocation.

[0047] The initial ranking and screening module includes: an initial ranking generation unit, a causal effect comparison unit, a resource return unit, and a resource transfer unit; the initial ranking generation unit generates an initial ranking list based on the comprehensive value score and the effectiveness prediction results; the causal effect comparison unit compares the individual processing effect of each user with a preset incremental threshold and removes users whose individual processing effect is not greater than the incremental threshold; the resource return unit counts the planned resource amount of the removed users and releases it to the global resource pool; the resource transfer unit redistributes the returned resources to high-potential users in descending order of individual processing effect.

[0048] By finely breaking down the initial ranking and screening module, this system achieves fully automated processing across the entire chain, from candidate user generation to high-potential user screening and resource reallocation. Specifically, the initial ranking generation unit combines comprehensive value scores with performance prediction results, ensuring that the initial ranking reflects both the static value and dynamic performance of users, resulting in a more scientific and reasonable ranking. The causal effect comparison unit uses individual treatment effects as the screening criterion, rather than the traditional absolute value of conversion probability, accurately identifying users whose incremental conversions are truly driven by marketing intervention, effectively avoiding the waste of resources on naturally converting users. The resource return unit releases the planned resources of removed users back into the global resource pool, achieving closed-loop resource recycling and improving resource utilization efficiency without increasing the total budget.

[0049] A third embodiment of the present invention provides a storage medium storing a computer program thereon, characterized in that the program, when executed by a processor, is the method or system of the present invention.

[0050] This storage medium stores a computer program that, when executed by a processor, can implement all steps of an AI-based marketing resource allocation method or all module functions of the system. This storage medium can exist independently of specific hardware devices and can be installed in servers, cloud platforms, edge computing nodes, or terminal devices, greatly expanding the breadth of industrial applications for the technology solution.

[0051] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. All equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this invention should still be covered by the claims of this invention.

Claims

1. A marketing resource allocation method based on AI, characterized in that, Includes the following steps: S1. Obtain multi-dimensional behavioral data and social relationship data of the target user and input them into the dynamic user value evaluation model. The dynamic user value evaluation model integrates the user's life cycle stage, conversion probability, customer lifetime value, real-time behavioral activity and social network dissemination index to calculate the comprehensive value score of each user. S2, using the revenue prediction model and the resource consumption rate prediction model, predicts the expected revenue of each user and the unit resource consumption rate of the corresponding channel, and generates performance prediction results. S3. Based on the comprehensive value score and performance prediction results, an initial ranking list is generated. The causal prediction model is used to make counterfactual incremental predictions on the users in the list, users whose natural conversion probability is higher than the conversion threshold are removed, a subset of high-potential users is selected, and the resources occupied by the removed users are returned to the global resource pool. S4. Construct a multi-objective resource allocation model, using the basic allocation time slice as the decision unit. The optimization objectives are to maximize the comprehensive value score and minimize the resource consumption cost. Introduce a decay penalty mechanism based on historical violation records as a constraint to generate a marketing resource allocation strategy. S5, execute resource allocation according to the allocation strategy and collect online execution data in real time; calculate the KL divergence between the distribution of online execution data and the distribution of historical sample data used in training the causal prediction model and the dynamic evaluation model of user value; when the KL divergence exceeds a preset threshold, backtrack several preset time slices based on the current time slice, and update the multi-objective resource allocation model, the dynamic evaluation model of user value and the causal prediction model through gradient descent using the real feedback data in the window.

2. The AI-based marketing resource allocation method according to claim 1, characterized in that, The social network dissemination power index is generated in the following way: Construct a user social relationship graph, mapping users as nodes and the attention, interaction or transaction relationships between users as edges; The topological characteristics of each node in the computational graph are evaluated, including degree centrality (measuring the number of direct connections), proximity centrality (measuring the efficiency of information transfer), and betweenness centrality (measuring the bridging effect). Based on the aforementioned topological feature indicators, a social network propagation index is calculated using a nonlinear mapping function to assess the potential for viral spread when a user initiates a marketing campaign.

3. The AI-based marketing resource allocation method according to claim 1, characterized in that, Step S3 includes using a pre-trained causal prediction model to predict the expected conversion result for each candidate user in the initial ranking list under two conditions: receiving marketing intervention and not receiving marketing intervention, and using the difference between the two as the individual treatment effect. Set an incremental threshold, and only retain users whose individual processing effect is greater than the incremental threshold into the high-potential user subset, while users whose individual processing effect is less than or equal to the incremental threshold are judged as naturally converted users and are removed. The total amount of resources originally allocated to the removed users is calculated, and these resources are returned to the global resource pool to supplement the resource quotas of the high-potential user subset.

4. The AI-based marketing resource allocation method according to claim 1, characterized in that, The attenuation penalty mechanism specifically includes: pre-establishing a historical violation record database to store each event of excessive resource concentration or long-term value destruction and its time slice; When calculating the penalty value of the current allocation strategy, the current time slice is used as the reference, and all past time slices are traversed. For each time slice with a violation, an exponentially decaying weight is calculated according to the time interval. The total penalty value is obtained by multiplying the decay weight of each time slice by the violation quantification value of that time slice and summing the results. This total penalty value is then added as a negative reward term to the objective function of the resource allocation model.

5. The AI-based marketing resource allocation method according to claim 1, characterized in that, The objective function of the multi-objective resource allocation model is constructed as follows: in, The objective function is... Assign tensors to 3D resources. Represents a set of users The total number of users in the region Represents a set of channels The total number of channels in the middle, Represents the set of time slices The total number of time slices in the game; Indicates time slice Through channels Assigned to user The amount of resources and meet the requirements , , This represents the current total budget for the global resource pool. For users The overall value score; For channels The unit resource cost; For dynamic trade-off coefficients; For the first The quantification of the severity of each violation; The time slice in which it occurs; This is the current time slice; The attenuation coefficient; To iterate through all events occurring in the current time slice Previous violations.

6. The AI-based marketing resource allocation method according to claim 1, characterized in that, The KL divergence detection and model update mechanism in step S5 specifically includes defining the online real-time data distribution. With training data distribution The KL divergence between them is, in, Indicates the KL divergence value; This indicates the data distribution generated in the current online real-time environment; This indicates the distribution of historical sample data used during the model training phase; This indicates that the integration operation is performed over the entire feature space; This indicates the online real-time data distribution. At feature points The probability density function value at that location; Represents the distribution of training data At feature points The probability density function value at that location; This represents the input feature vector; The log-likelihood ratio measures the likelihood of similar eigenvalues. Below, the probability difference between the real-time distribution and the training distribution; when When the divergence exceeds the divergence threshold, concept drift is determined to have occurred, triggering the model update mechanism.

7. A system for implementing the method as described in claims 1-6, characterized in that, include: The data acquisition module is used to obtain user behavior data, social relationship data, and resource allocation feedback. The user value assessment module is used to integrate user lifecycle stages, conversion probability, customer lifetime value, behavioral activity and social communication power index to calculate the user's comprehensive value score. The performance prediction module is used to predict users' expected revenue and the rate of channel resource consumption. A causal prediction model is used to estimate the individual treatment effect of users accepting marketing interventions, and the individual treatment effect is output to the ranking screening module. The initial ranking and screening module is used to select a subset of high-potential users based on their comprehensive value scores and individual treatment effects, and to perform resource recycling. The multi-objective allocation optimization module is used to generate allocation strategies by introducing a decay penalty mechanism based on historical violations, with the objectives of maximizing the comprehensive value score and minimizing resource costs. The distribution offset detection module is used to calculate the KL divergence between the online data distribution and the training data distribution and determine whether it exceeds the dynamic threshold. The online update module is used to backtrack the time window data when the KL divergence exceeds the divergence threshold and update the model through gradient descent. The strategy execution module is used to allocate resources according to the allocation strategy and provide feedback on the execution results.

8. The AI-based marketing resource allocation system according to claim 7, characterized in that, The user value assessment module includes: The system includes a lifecycle analysis unit, a conversion probability prediction unit, a customer lifetime value calculation unit, a behavioral activity analysis unit, a social influence analysis unit, and a weighted fusion unit. The social influence analysis unit is used to construct a user social relationship graph and generate a social network propagation index based on the degree centrality, proximity centrality, and betweenness centrality of the graph.

9. The system according to claim 8, characterized in that, The initial sorting and screening module includes: an initial sorting generation unit, a causal effect comparison unit, a resource return unit, and a resource transfer unit; The initial ranking generation unit generates an initial ranking list based on the comprehensive value score and the effectiveness prediction result. The causal effect comparison unit compares the individual treatment effect of each user with a preset incremental threshold and removes users whose effect is not greater than the incremental threshold. The resource return unit counts the planned resource amount of the removed users and releases it to the global resource pool; The resource transfer unit redistributes the returned resources to high-potential users in descending order of individual processing effect.

10. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method or system described in claims 1-9.

Images