A precise marketing resource allocation method and system based on user marketing value life cycle

Abstract

This invention discloses a method and system for precise marketing resource allocation based on the user's marketing value lifecycle, specifically relating to the field of marketing resource allocation technology. The method includes acquiring multi-source heterogeneous data of the target user group and constructing a user state representation that integrates static user attributes, dynamic behavioral sequences, and market dynamic characteristics. Based on the user state representation, a dynamic CLV prediction model is used to simultaneously predict the user's future lifecycle value and user state transition probability. This invention solves the technical problems of static prediction, disconnect between optimization and execution, and rigid budgeting in traditional marketing resource allocation by constructing a closed-loop collaborative architecture of dynamic CLV prediction, heterogeneous MDP planning, and contextual multi-armed learning. This enables the marketing resource allocation process to combine the foresight of long-term planning with the agility of short-term execution, improving the accuracy of CLV and state transition probability predictions. Through asynchronous updates and knowledge transfer mechanisms, it ensures the system's stable adaptation in dynamic environments.

Description

Technical Field

[0001] This invention relates to the field of marketing resource allocation technology, and more specifically, to a method and system for precise marketing resource allocation based on the lifecycle of user marketing value. Background Technology

[0002] In the field of precision marketing, resource allocation strategies based on Customer Lifecycle Value (CLV) have evolved from auxiliary reference indicators to core decision-making tools for enterprises to improve marketing efficiency and ROI. CLV theory, by quantifying the total expected value created by users for enterprises throughout the complete relationship lifecycle, drives marketing strategies from short-term transaction-oriented to long-term relationship maintenance, achieving precise matching of resource investment with potential customer value. However, with the increasing dynamism of the market environment, the complexity of user behavior, and the diversification of channels, traditional CLV models based on static calculations of historical transaction data are struggling to adapt to real-time decision-making needs.

[0003] Existing technical solutions have limitations in terms of dynamism, collaboration, and adaptability, which restricts their practical effectiveness. First, CLV prediction models are static and isolated. Mainstream methods rely on historical transaction data for static evaluation and clustering, which cannot respond to changes in behavior in real time and completely ignore external effects such as market competition and user social influence, resulting in value assessment lagging behind the market environment. Second, optimization and execution are severely disconnected. Long-term value planning and short-term marketing execution are often deployed separately. The former has fixed strategies and is difficult to integrate with real-time feedback, while the latter pursues short-term indicators such as click-through rates, which can easily damage long-term customer relationships. Moreover, the system is rigid and lacks intelligent response. Budget allocation is usually fixed in advance and cannot be dynamically adjusted according to the results. Model updates rely on manual or fixed cycles and cannot be quickly adjusted when competitors take action or the market changes, resulting in resource misallocation and decreased efficiency. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the present invention provides a method and system for precise marketing resource allocation based on the user marketing value lifecycle, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for precise marketing resource allocation based on the lifetime value of user marketing value, comprising the following steps: S1: Acquire multi-source heterogeneous data of the target user group and construct a user state representation that integrates user static attributes, dynamic behavior sequences and market dynamic characteristics; S2: Based on the user state representation, a dynamic CLV prediction model is used to simultaneously predict the lifecycle of the user's future marketing value and the probability of user state transition; S3: Construct a Markov decision process model that considers user heterogeneity, use the prediction results of S2 as the basis for constructing the state space, and solve for the set of candidate marketing strategies with the long-term marketing return on investment as the objective. S4: Parameterize the candidate marketing strategy set into the arm space of a multi-armed slot machine, and based on real-time feedback data, use a contextual multi-armed slot machine algorithm to learn online and select the optimal strategy for execution; S5: Collect user feedback data after the strategy is executed, update the state transition probabilities of the dynamic CLV prediction model and the MDP model, and form a closed-loop optimization system.

[0006] Preferably, the dynamic CLV prediction model in S2 adopts a multi-task learning framework, which simultaneously completes the prediction of the user's future lifecycle value and the prediction of the user's state transition probability by sharing the underlying feature representation; wherein, in the process of building the model, a competition effect modeling unit is integrated, which is used to learn and quantify the externality influence between users based on the user relationship graph through graph neural network, and uses this influence factor as one of the input features for prediction.

[0007] Preferably, in S3: The state space of the Markov decision process model contains a latent categorical variable obtained from the user's historical behavior sequence through unsupervised learning, which is used to characterize the heterogeneity of the user. The reward function of the Markov decision process model consists of a first term reflecting the change in lifecycle value brought about by the action, a second term reflecting the cost of the action, and an information gain term used to explicitly incentivize the exploration of unknown states.

[0008] Preferably, the solution of the Markov decision process model in S3 is carried out under the preset global marketing budget constraint and coupled with a budget dynamic reallocation mechanism; this mechanism allows for temporary budget adjustments among users, and for each adjustment, the decision is based on an assessment of the change in the user's total expected lifetime value before and after the adjustment. The cost of the adjustment is positively correlated with the degree of difference in the marketing strategies of the users on both sides of the adjustment.

[0009] Preferably, in S4: During execution, the contextual multi-armed slot machine algorithm uses the collected real-time state transition data to perform online incremental updates of the state transition probabilities in the Markov decision process model. The algorithm also includes a process for processing delayed conversion rewards, which decomposes and attributes the observed composite reward signal with time delay to the corresponding marketing actions that historically led to the conversion.

[0010] Preferably, it further includes S6: establishing an asynchronous collaborative update mechanism between the prediction layer, the optimization layer, and the execution layer, wherein: The dynamic CLV prediction model, the Markov decision process model, and the contextual multi-armed slot machine algorithm are updated at different time scales. The update of the Markov decision process model is triggered by a fixed-time-period event or an event in which the model uncertainty exceeds a predetermined threshold. The update processes of the dynamic CLV prediction model, the Markov decision process model, and the contextual multi-armed slot machine algorithm respectively adopt update step size or learning rate parameters that decrease from large to small in order to maintain the overall stability of the system.

[0011] A precision marketing resource allocation system for implementing the above method includes: The user state feature construction module is used to acquire multi-source heterogeneous data and generate user state representations. The user marketing value and switching probability prediction module is used to receive the user state representation and output the user marketing value lifecycle prediction and state switching probability prediction. The marketing strategy planning module is used to solve the Markov decision process model based on the prediction results to generate a set of candidate marketing strategies. The online strategy learning and execution module is used to parameterize the candidate marketing strategy set into a multi-armed slot machine, and to perform online learning and strategy selection based on real-time feedback to execute resource allocation; The model update and control module is used to update the model parameters of the prediction module and the planning module based on the policy execution feedback data, and to coordinate the update rhythm of each module.

[0012] Preferably, the user marketing value and conversion probability prediction module includes a multi-task learning network, a competition effect modeling unit, and an incremental update unit; the incremental update unit uses stochastic variational inference to realize the streaming update of model parameters, and its update triggering condition is decoupled from the marketing activity execution cycle.

[0013] Preferably, the state space constructor in the marketing strategy planning module uses a Dirichlet process mixture model to perform unsupervised clustering of user historical behavior sequences to generate latent categorical variables representing user heterogeneity. ; The number of clusters in the Dirichlet process hybrid model can be dynamically adjusted according to the data distribution without needing to be preset.

[0014] Preferably, the model update and control module includes: The trigger logic unit is configured to dynamically trigger updates of different modules based on at least one of the following: a fixed period, deviation of system performance indicators from expectations, and changes in the external market environment. The knowledge transfer unit is configured to transform the strategy experience gained by the online strategy learning and execution module in real-time interaction into initial information for the marketing strategy planning module to solve strategies, so as to accelerate the strategy optimization process. The consistency arbitration unit is configured to adaptively adjust the system's dependence on both when there is a significant difference between the prediction results of the user marketing value and conversion probability prediction module and the real-time feedback statistics of the online strategy learning and execution module, and to trigger the adaptive adjustment of the prediction module.

[0015] The technical effects and advantages of this invention are as follows: By constructing a multi-layered collaborative architecture integrating dynamic prediction, offline planning, online learning, and closed-loop control, this approach effectively overcomes the technical limitations of existing marketing resource allocation schemes, such as static prediction, delayed decision-making, and rigid execution. It uses the value and transition probability output by the dynamic CLV prediction model as the basis for constructing the MDP state space, enabling long-term decisions to be based on real-time user states. Furthermore, it parameterizes the candidate strategy set generated by the MDP into the arm space of a multi-armed slot machine, achieving strategy-level online learning and rapid optimization. Finally, through model updates and control modules, it achieves cross-layer feedback and adaptive optimization, forming a closed-loop system that connects data, models, and strategies. This allows the marketing resource allocation process to combine the foresight of long-term planning with the agility of short-term execution, improving marketing ROI and increasing high-value user retention under the same budget constraints. By introducing a multi-task learning framework that integrates competitive effect modeling units in the prediction phase, and quantifying the externality impact among users through graph neural networks, the accuracy of CLV and state transition probability prediction is improved, overcoming the shortcomings of traditional models that ignore market interaction effects. An information gain term is innovatively introduced into the MDP reward function, and the explicit incentive system explores uncertain states, effectively solving the decision bias problem in cold start and data sparsity scenarios. By designing an elastic budget pool mechanism based on policy KL divergence, the budget is dynamically and safely adjusted among users, avoiding resource misallocation caused by policy conflicts. This enables the system to accurately characterize user heterogeneity, market dynamics, and policy interaction, achieving truly personalized optimal resource allocation. Through asynchronous collaborative update mechanisms and intelligent control strategies, the stability and reliability of complex learning systems in real-world environments are ensured. Differentiated update cycles and progressive learning rates are set for the prediction, planning, and execution layers to guarantee the robustness of high-level strategies and the sensitivity of low-level controls. An event-triggered update mechanism and a prediction-feedback conflict arbitration mechanism are designed to enable the system to proactively perceive, respond quickly, and self-correct. The knowledge transfer unit enables the hot-start of online strategy experience into the offline planning model, significantly accelerating model convergence, effectively reducing system operation and maintenance costs, and improving the feasibility of large-scale commercial use. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow structure of the present invention.

[0017] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figures 1-2 The method for precise marketing resource allocation based on the user's marketing value lifecycle, as shown, includes the following steps: S1: Constructing a user state representation Data Acquisition and Preprocessing: User static attribute data: extracted from enterprise customer relationship management (CRM) systems or user databases, including but not limited to: age, gender, geographic region, registration duration, historical cumulative spending amount, and membership level. Continuous variables are standardized, such as Z-score normalization, and categorical variables are one-hot encoded.

[0020] User dynamic behavior sequence data: Time series data is extracted from application backend logs or data warehouses. For each user, the sequence of behavioral events within the most recent T time units is extracted, for example, T=90 days. Each event includes the behavior type such as "browsing product details page", "adding to cart", "completing payment", "customer service inquiry", the timestamp of the event, and related entity information such as product ID and order amount. The sequence is sorted in ascending order of timestamp.

[0021] Market dynamic characteristic data: This data is obtained periodically from external data sources via application programming interface (API), including: the overall industry sales index, real-time promotional information of major competitors, data obtained by web crawling monitoring their official websites or applications, and sentiment index of related brands or categories on social media. This data is uniformly processed into time-series characteristics with the same time granularity as user behavior data.

[0022] Feature fusion and representation generation: Static attributes, after being encoded, form a dense vector of fixed dimensions. .

[0023] Dynamic behavioral sequences are modeled using a Transformer encoder. Specifically, each behavioral event is mapped to an embedding vector, which is then fed into the Transformer encoding layer after being encoded with position. The Transformer layer utilizes its self-attention mechanism to capture long-term dependencies between behaviors. The output of the last layer (CLS) is used, or the outputs of all positions in the sequence are pooled (e.g., average pooling) to obtain a vector representing the user's recent behavioral patterns. .

[0024] Market dynamic characteristics, after smoothing and standardization, form an environmental context vector. .

[0025] Finally, the three vectors are concatenated to form a complete user state representation. This vector comprehensively represents the user's inherent attributes, recent dynamic intentions, and external market environment, providing a comprehensive information foundation for subsequent predictions.

[0026] S2: Dynamic CLV and State Transition Probability Prediction Model architecture: Shared Feature Extraction Layer: The user state representation s is first passed through a shared multilayer perceptron (MLP), which consists of several fully connected layers and activation functions such as ReLU, to learn higher-level abstract features. .

[0027] Competition Effect Modeling Unit: To quantify the mutual influence among users, such as peer effect and competition effect, this unit constructs a user relationship graph. Among them, nodes For users, if two users have exhibited similar behavioral patterns over a period of time—for example, purchased the same type of goods, clicked on the same ads, or belonged to the same social network community—an edge is established between them. A two-layer graph attention network (GAT) is used to aggregate neighbor information for the target user node. Its characteristics after the aggregation of competitive effects The calculation is as follows: ; in, User The neighborhood group, It is a learnable weight matrix. It is a non-linear activation function. It is a normalized attention coefficient calculated through an attention mechanism, representing the user's attention level. For users Ultimately, the influence weight is determined by combining the characteristics of the entity itself with those of the competitive effect: .

[0028] Multi-task prediction head: CLV Prediction Head: Will Input a regression subnetwork, typically consisting of several fully connected layers, and output a scalar. As a point estimate of the discounted CLV for a future period of time, such as the next 12 months, this subnetwork can also output the predicted confidence interval, such as through quantile regression or by directly predicting the variance parameter.

[0029] State transition probability prediction head: The input is a classification subnetwork, typically consisting of several fully connected layers with a Softmax activation function. The output dimension of this network corresponds to the number of categories for the next possible user state. For a given marketing action 'a', such as "no intervention", "send a 5 yuan coupon", or "push personalized ads", the network outputs a probability distribution. , indicating the state Next action Then, transition to state The probability of.

[0030] Model training and incremental updates: Initial training: The model is trained offline using historical data. The loss function is designed as a weighted sum of the losses from the two tasks: ;in, Mean squared error (MSE) or quantile loss can be used. Cross-entropy loss is used.

[0031] Streaming Incremental Updates: After the model goes live, its "incremental update unit" is responsible for continuous learning. This unit uses stochastic variational inference (SVI) technology; whenever a certain amount of new feedback data is accumulated, for example, every time 1000 new state transition samples are collected, SVI treats this small batch of data as an observation and updates the parameters of the model's variational posterior distribution online. The update trigger is decoupled from the specific marketing campaign execution cycle and can be automatically triggered based on fixed time intervals, such as every 6 hours or a data volume threshold, ensuring that the model can autonomously and continuously adapt to changes in data distribution.

[0032] S3: Constructing and Solving Heterogeneous MDP Models State space construction: Status in marketing decisions It is a discrete or discretized vector that represents the continuous states of S2. It is derived from the Chinese; generally, .

[0033] : The CLV value predicted by S2 It is divided into several intervals, such as "high value", "medium value" and "low value", corresponding to discrete levels.

[0034] Based on a user's recent behavioral sequence, such as purchase frequency and activity level, determine their lifecycle stage, for example, "new user", "active user", "dormant user" and "user at risk of churn".

[0035] This is key to reflecting user heterogeneity. The state space builder uses the Dirichlet Process Mixture Model (DPMM) to perform unsupervised clustering of the historical behavior sequences of all users. DPMM is a non-parametric Bayesian model, whose advantage is that it does not require pre-specifying the number of clusters K; the model can determine the appropriate number of categories based on the data. Each cluster represents a potential user behavior pattern or segment, such as "price-driven," "quality-seeking," or "impulsive consumer." The assigned cluster ID is This allows MDP to explicitly distinguish and utilize the differentiated response patterns of different potential users.

[0036] Action space: Defined as a combination of executable marketing interventions. For example, an action 'a' can be a triple: {marketing channel: email, incentive type: percentage discount coupon, incentive strength: 20% off}. The system pre-sets a set A of actions containing a finite number of such combinations.

[0037] Reward function design: reward function The design directly guides the optimization direction, and the present invention adopts the following structure: in: It is a long-term value term, based on the state transition probability predicted by S2. and the state after the transition The corresponding expected CLV is used to calculate the execution action. This will result in an expected increase in long-term CLV.

[0038] It is a cost item, that is, the action performed. Direct financial costs, such as the face value of coupons.

[0039] It is an information gain term, an exploration incentive term; it calculates the execution action. Then, the expected reduction in uncertainty of the state transition model; for example, it can be achieved using actions. Transition probability distribution before and after execution The reduction in entropy is used as a metric, and this explicitly encourages the system to explore state-action pairs where the model is uncertain, thereby accelerating learning.

[0040] and It is a hyperparameter used to balance short-term costs, long-term benefits, and exploration value.

[0041] Budget constraints and flexible budget pools: The solution to the entire MDP is based on the total marketing budget. It is carried out under the constraints of [the relevant regulations / regulations].

[0042] Dynamic budget reallocation mechanism: This is the key to solving budget rigidity in this invention. This mechanism allows for the temporary allocation of unused or inefficient budgets from some users to other users within the planning cycle; when considering the allocation from users... Adjusting the budget For users At that time, the decision-making criterion is to maximize the global expected value; specifically, the following optimization problem is solved: ; in, Indicates the user under a given budget The expected CLV is estimated based on the current MDP model.

[0043] Adjustment costs: To prevent unreasonable budget shifts, adjustment costs are set. This cost and the current optimal strategy for both users are considered. and The degree of difference between the two policy distributions is positively correlated, and one way to quantify this is to calculate the KL divergence between the two policy distributions: in, Representation Strategy and The KL divergence between users makes the system tend to allocate budgets among users with similar strategies, thus avoiding efficiency losses due to strategy conflicts.

[0044] Model Solution: The Proximal Policy Optimization (PPO) algorithm with budget constraints is used to solve the above MDP. The PPO algorithm iteratively updates the policy network by interactively sampling with the environment, i.e., the prediction model of S2. parameters To maximize the expected cumulative reward; after solving, output a set of high-performance candidate policy parameters. Each Corresponding to a complete marketing strategy These together constitute a set of candidate marketing strategies.

[0045] S4: Online Strategy Learning and Execution Arm space parameterization: The M candidate policies obtained from S3 are... Each arm is parameterized as an "arm" in a contextual multi-armed slot machine, with each arm representing a complete MDP strategy. This is done when a state is needed. When a user makes a decision (based on the original continuous state representation), the arm will act according to their strategy. Provide action suggestions.

[0046] Online selection and execution, Thompson sampling: Each arm Maintain a posterior distribution for its expected reward. For example, if the reward follows a Bernoulli distribution, the posterior is a Beta distribution; for continuous rewards, a Gaussian distribution can be used.

[0047] At every decision moment Regarding the current user state (In context), the algorithm starts from each arm Sample an expected reward value from the posterior distribution. .

[0048] Select the arm with the largest sample value. As the winning arm of the current round.

[0049] Execute the strategy corresponding to this arm. Recommended actions For example, sending a text message with specific content to a user while recording the context. and the chosen arm .

[0050] Key Mechanism 1: Online Incremental Update of MDP Transfer Probability: During the execution of a slot machine, real interaction data, known as triples, is continuously collected. ,in It is the new user state observed after the action is performed.

[0051] This real-time data is used to update the state transition probability prediction model in S2 online. Specifically, a meta-learning approach based on policy gradients is employed: a model parameter for the transition probability is defined. The meta-loss for the objective is the negative value of the cumulative reward gained from online interactions. ; By calculating this loss gradient Update using gradient descent. This allows the transition probability model to better reflect the real, dynamically changing environment, thus keeping the MDP model relevant.

[0052] The update formula can be approximated as: in, Indicating in strategy The expectation of the downsampled state, action, and next state triple. It is a state-action value function estimate. The parameter is State transition probability model, It is the model's log-likelihood relative to the parameters. The gradient.

[0053] Key Mechanism Two: Delayed Conversion Reward Processing Marketing results, such as purchases, are often delayed. The "Delayed Feedback Processing Unit" in this module maintains a reward buffer that needs to be attributed.

[0054] When a delayed conversion event occurs in time Observed and rewarded At that time, the unit adopted an exponentially decaying kernel attribution model, assuming that the transformation might be due to past... If marketing activities within a specific timeframe work together, then... According to the attenuation coefficient Assign to historical actions: ,in ; in, It is attributed to time steps The partial rewards for the actions performed are immediately used to update the posterior distribution of the corresponding arm, i.e., the policy selected at that time, after the attribution is completed.

[0055] Step S5 / S6: Closed-loop feedback and asynchronous collaborative update Trigger logic unit; Periodic triggering: Set the basic update cycle for different modules, following... The principles are as follows: for example, the CLV prediction model is updated fully or incrementally every 24 hours (T1); the MDP strategy planning is re-solved based on the latest prediction model every hour (T2); and the reward posterior distribution of the slot machine arm is updated every minute (T3) based on new feedback.

[0056] Performance Deviation Trigger: Continuously monitor key performance indicators. If the actual average return on investment (ROI) of the system is lower than the theoretical expected value of the current MDP model for N=10 consecutive evaluation periods. If %=15%, the MDP module will be immediately re-solved to adjust the strategy.

[0057] External event triggering: Real-time monitoring of market dynamics. If a competitor suddenly launches a large-scale promotion or the industry sentiment index fluctuates beyond a preset threshold (e.g., 20%) within a short period, the CLV prediction model is immediately updated to quickly adapt to changes in the external environment.

[0058] Knowledge transfer unit: Regularly check, for example every 30 minutes, the online strategy learning and execution module to see which strategy arm, i.e., which... It has performed best recently, as mentioned above, Thompson sampling has the highest average sampling reward.

[0059] The strategy parameters corresponding to the winning arm The extracted parameters are used as "hot start" parameters for initializing the strategy network during the next iteration of the PPO algorithm in the marketing strategy planning module. This enables positive knowledge transfer from online real-time learning to offline global planning, which can significantly accelerate the convergence of MDP strategies and improve planning quality.

[0060] Consistency Arbitration Unit: Continuously compare the output of the user marketing value and conversion probability prediction module, for example, the predicted conversion rate of a certain type of user with the statistical results actually observed by the online strategy learning and execution module, for example, the actual conversion rate of that type of user.

[0061] When there is a statistically significant difference between the two, for example, when the p-value is less than 0.01 through hypothesis testing, the unit determines that a "prediction-reality" conflict has occurred.

[0062] The unit adopts the idea of ​​Bayesian model averaging to dynamically reduce the dependence weight of the prediction results of the problem module in the system decision-making. At the same time, it generates a small batch training set containing recent conflict data and triggers the competitive effect modeling subnetwork in the prediction module to perform a fast targeted fine-tuning, updating only the parameters of the subnetwork, so as to quickly correct the erroneous estimates of market interaction relationships.

[0063] Stability guaranteed, progressive learning rate: When training models at different levels, a learning rate that decreases from large to small is used: For example, setting .

[0064] This design ensures that the parameters of the high-level model's CLV prediction are updated robustly and slowly, avoiding drastic fluctuations due to short-term noise; the mid-level model's MDP planning can smoothly adjust the strategy; and the low-level controller's slot machine execution can quickly adapt to real-time feedback. This hierarchical stability control is the key to the reliable operation of complex learning systems.

[0065] A precision marketing resource allocation system for implementing the above method includes: The user state feature construction module is used to acquire multi-source heterogeneous data and generate user state representations. The user marketing value and transition probability prediction module is used to receive user state representations and output lifecycle predictions of user marketing value and state transition probability predictions. The marketing strategy planning module is used to solve the Markov decision process model based on the prediction results to generate a set of candidate marketing strategies. The online strategy learning and execution module is used to parameterize the candidate marketing strategy set into a multi-armed slot machine, and to perform online learning and strategy selection based on real-time feedback to execute resource allocation; The model update and control module is used to update the model parameters of the prediction and planning modules based on the policy execution feedback data, and to coordinate the update rhythm of each module.

[0066] The user marketing value and conversion probability prediction module includes a multi-task learning network, a competition effect modeling unit, and an incremental update unit. The incremental update unit uses stochastic variational inference to achieve streaming updates of model parameters, and its update triggering conditions are decoupled from the marketing campaign execution cycle.

[0067] The state space builder in the marketing strategy planning module uses a Dirichlet process mixture model to perform unsupervised clustering of user historical behavior sequences to generate latent categorical variables representing user heterogeneity. ; In particular, the number of clusters in the Dirichlet process hybrid model can be dynamically adjusted according to the data distribution without needing to be preset.

[0068] The model update and control module includes: The trigger logic unit is configured to dynamically trigger updates of different modules based on at least one of the following: a fixed period, deviation of system performance indicators from expectations, and changes in the external market environment. The knowledge transfer unit is configured to transform the strategic experience gained by the online strategy learning and execution module in real-time interaction into initial information for the marketing strategy planning module to solve strategies, thereby accelerating the strategy optimization process. The consistency arbitration unit is configured to adaptively adjust the system's dependence on both the prediction results of the user marketing value and conversion probability prediction module and trigger adaptive adjustments to the prediction module when there is a significant difference between the prediction results of the online strategy learning and execution module and the real-time feedback statistics of the online strategy learning and execution module.

[0069] Finally, it should be noted that the accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for precise marketing resource allocation based on the user's marketing value lifecycle, characterized in that: Includes the following steps: S1: Acquire multi-source heterogeneous data of the target user group and construct a user state representation that integrates user static attributes, dynamic behavior sequences and market dynamic characteristics; S2: Based on the user state representation, a dynamic CLV prediction model is used to simultaneously predict the lifecycle of the user's future marketing value and the probability of user state transition; S3: Construct a Markov decision process model that considers user heterogeneity, use the prediction results of S2 as the basis for constructing the state space, and solve for the set of candidate marketing strategies with the long-term marketing return on investment as the objective. S4: Parameterize the candidate marketing strategy set into the arm space of a multi-armed slot machine, and based on real-time feedback data, use a contextual multi-armed slot machine algorithm to learn online and select the optimal strategy for execution; S5: Collect user feedback data after the strategy is executed, update the state transition probabilities of the dynamic CLV prediction model and the MDP model, and form a closed-loop optimization system.

2. The method for precise marketing resource allocation based on the user marketing value lifecycle as described in claim 1, characterized in that: The dynamic CLV prediction model in S2 adopts a multi-task learning framework. By sharing the underlying feature representation, it simultaneously completes the prediction of the user's future lifetime value and the prediction of the user's state transition probability. In the process of building the model, a competition effect modeling unit is integrated to learn and quantify the externality influence between users based on the user relationship graph through graph neural network, and to use this influence factor as one of the input features for prediction.

3. The method for precise marketing resource allocation based on the user marketing value lifecycle as described in claim 1, characterized in that: In S3: The state space of the Markov decision process model contains a latent categorical variable obtained from the user's historical behavior sequence through unsupervised learning, which is used to characterize the heterogeneity of the user. The reward function of the Markov decision process model consists of a first term reflecting the change in lifecycle value brought about by the action, a second term reflecting the cost of the action, and an information gain term used to explicitly incentivize the exploration of unknown states.

4. The method for precise marketing resource allocation based on the user marketing value lifecycle as described in claim 3, characterized in that: The solution of the Markov decision process model in S3 is carried out under the preset global marketing budget constraint and coupled with a budget dynamic reallocation mechanism. This mechanism allows for temporary budget adjustments among users, and for each adjustment, the decision is based on the assessment of the change in the user's total expected lifetime value before and after the adjustment. The cost of the adjustment is positively correlated with the degree of difference in the marketing strategies of the users on both sides of the adjustment.

5. The method for precise marketing resource allocation based on the user marketing value lifecycle as described in claim 1, characterized in that: In S4: During execution, the contextual multi-armed slot machine algorithm uses the collected real-time state transition data to perform online incremental updates of the state transition probabilities in the Markov decision process model. The algorithm also includes a process for processing delayed conversion rewards, which decomposes and attributes the observed composite reward signal with time delay to the corresponding marketing actions that historically led to the conversion.

6. The method for precise marketing resource allocation based on the user marketing value lifecycle as described in claim 1, characterized in that: It also includes S6: establishing an asynchronous collaborative update mechanism between the prediction layer, optimization layer, and execution layer, wherein: The dynamic CLV prediction model, the Markov decision process model, and the contextual multi-armed slot machine algorithm are updated at different time scales. The update of the Markov decision process model is triggered by a fixed-time-period event or an event in which the model uncertainty exceeds a predetermined threshold. The update processes of the dynamic CLV prediction model, the Markov decision process model, and the contextual multi-armed slot machine algorithm respectively adopt update step size or learning rate parameters that decrease from large to small in order to maintain the overall stability of the system.

7. A precision marketing resource allocation system for implementing the method of any one of claims 1-6, characterized in that, include: The user state feature construction module is used to acquire multi-source heterogeneous data and generate user state representations. The user marketing value and switching probability prediction module is used to receive the user state representation and output the user marketing value lifecycle prediction and state switching probability prediction. The marketing strategy planning module is used to solve the Markov decision process model based on the prediction results to generate a set of candidate marketing strategies. The online strategy learning and execution module is used to parameterize the candidate marketing strategy set into a multi-armed slot machine, and to perform online learning and strategy selection based on real-time feedback to execute resource allocation; The model update and control module is used to update the model parameters of the prediction module and the planning module based on the policy execution feedback data, and to coordinate the update rhythm of each module.

8. A precision marketing resource allocation system based on the user marketing value lifecycle as described in claim 7, characterized in that: The user marketing value and conversion probability prediction module includes a multi-task learning network, a competition effect modeling unit, and an incremental update unit. The incremental update unit uses stochastic variational inference to achieve streaming updates of model parameters, and its update triggering conditions are decoupled from the marketing activity execution cycle.

9. A precision marketing resource allocation system based on the user marketing value lifecycle as described in claim 7, characterized in that: The state space constructor in the marketing strategy planning module uses a Dirichlet process mixture model to perform unsupervised clustering of user historical behavior sequences to generate latent categorical variables representing user heterogeneity. ; The number of clusters in the Dirichlet process hybrid model can be dynamically adjusted according to the data distribution without needing to be preset.

10. A precision marketing resource allocation system based on the user marketing value lifecycle as described in claim 7, characterized in that: The model update and control module includes: The trigger logic unit is configured to dynamically trigger updates of different modules based on at least one of the following: a fixed period, deviation of system performance indicators from expectations, and changes in the external market environment. The knowledge transfer unit is configured to transform the strategy experience gained by the online strategy learning and execution module in real-time interaction into initial information for the marketing strategy planning module to solve strategies, so as to accelerate the strategy optimization process. The consistency arbitration unit is configured to adaptively adjust the system's dependence on both when there is a significant difference between the prediction results of the user marketing value and conversion probability prediction module and the real-time feedback statistics of the online strategy learning and execution module, and to trigger the adaptive adjustment of the prediction module.

Images