Consumption coupon issuing control method and system based on big data

Abstract

The invention discloses a consumption coupon issuing control method and system based on big data, and aims to solve the technical problems that an existing coupon issuing strategy is low in budgeting efficiency due to dependency rather than causality, and long-term value of a user is damaged due to short view of an optimization target. The method comprises the following steps: firstly, under a federated learning architecture, fusing multi-source heterogeneous data through a topology and distribution perception aggregation algorithm on the premise of protecting data privacy, constructing a knowledge graph and learning entity representation vectors; secondly, on the basis of the representation vector, an individual causal gain value of the consumption coupon to the user is accurately calculated by adopting an R-Learner causal inference model; then, constructing a multi-target composite reward function of the predicted user long-term value variation; and finally, strategy learning is carried out by using a deep reinforcement learning agent to maximize the composite reward as a target, and dynamic optimization of coupon issuing decision is realized. According to the application, the marketing ROI and the long-term loyalty of the user can be remarkably improved.

Description

Technical Field

[0001] This application relates to the fields of big data processing and artificial intelligence technology, and in particular to a method and system for controlling the issuance of consumer vouchers based on big data. Background Technology

[0002] With the widespread adoption of e-commerce, consumer vouchers have become an important tool for online and offline merchants to conduct marketing activities. To improve marketing efficiency, the industry has widely adopted big data and artificial intelligence technologies to optimize consumer voucher distribution strategies. However, in practice, existing technologies mainly face the following two intractable technical bottlenecks: First, at the data fusion level, building accurate user profiles requires integrating fragmented data from different platforms, merchants, and even user terminals. Due to compliance requirements for data security and privacy protection, federated learning is considered a viable technical approach. However, in real-world business scenarios, data from various data sources (clients) exhibit significant differences in user behavior patterns, data distribution, and graph structure, demonstrating a high degree of non-independent and identically distributed (Non-IID) characteristics. This leads to a severe performance degradation of the global model when standard federated learning algorithms aggregate models from various sources, resulting in unstable training processes or even failure to converge, making it difficult to build a high-quality foundational model that can effectively integrate multi-source heterogeneous data.

[0003] Secondly, at the decision optimization level, even when using cutting-edge algorithms such as deep reinforcement learning to automatically generate coupon distribution strategies, their optimization goals generally suffer from "strategic short-sightedness." Specifically, the system is usually designed to maximize short-term, easily measurable metrics such as click-through rate and instant conversion rate. This optimization approach ignores the long-term impact of a single coupon distribution action on users' future behavior. A recommendation that seems successful in the short term (e.g., a user claims and uses a coupon) may, in the long run, reduce future platform activity and loyalty because it does not meet the user's deeper needs. Therefore, how to enable intelligent decision-making systems to simultaneously pursue short-term business returns while considering and optimizing the long-term lifetime value of users is a pressing technological challenge that needs to be addressed. Summary of the Invention

[0004] This application provides a consumer voucher issuance control method and system based on big data, aiming to solve the technical problems in the existing technology, such as voucher issuance strategies relying on correlation rather than causation, short-sighted optimization goals, and inability to integrate multi-source data while protecting data privacy.

[0005] A method for controlling the issuance of consumer vouchers based on big data, characterized by the following steps: S1. In a central server-multi-client federated learning architecture, distributed user data with different structures and distributions deployed on multiple data terminals are aggregated. By executing a topology and distribution-aware federated aggregation algorithm, an entity representation vector of a global heterogeneous knowledge graph containing users, vouchers, goods and their interaction relationships is constructed and learned. When aggregating the model parameters of each client, the aggregation algorithm dynamically weights the data by comprehensively considering the similarity of the topological structure and the difference in vector distribution of each client's data, so as to overcome the model training difficulties caused by the non-independent and identically distributed data. S2. Based on the entity representation vector obtained in step S1, establish a causal inference model to identify the true causal effect of the consumer voucher intervention on user purchasing behavior, and then calculate the individual causal gain value that can quantify the net increase in purchase probability for the target user and the consumer voucher to be selected. S3. To guide the optimization of the coupon issuance strategy, a multi-objective composite reward function is constructed. This reward function integrates: the immediate causal gain directly related to the marketing ROI of a single coupon issuance behavior, the long-term value increment used to measure future user retention and activity, and the ecological health index used to ensure the diversity of coupon issuance strategies and budget compliance. S4. The sequential distribution process of consumption vouchers is modeled as a Markov decision process, and a deep reinforcement learning agent is used to perform end-to-end policy learning with the goal of maximizing the long-term cumulative compound reward defined in step S3. This enables the agent to make dynamic decisions and output the optimal target consumption voucher when the real-time status of the target user is received.

[0006] Preferably, in step S2, calculating the individual causal gain value of the consumption voucher for the target user specifically includes: S21. The issue of issuing consumption vouchers is constructed as a causal inference task. The individual causal gain (Conditional Average Treatment Effect, CATE) to be calculated aims to measure the net improvement effect on user purchasing behavior (outcome variable Y) when a specific consumption voucher is issued (treatment variable T=1) compared to when it is not issued (control variable T=0). S22. Using the R-Learner meta-learner model, the individual causal gain value is estimated through a two-stage process. In the first stage, using historical interaction data and the representation vector generated in step S1 as input, an outcome model and a propensity score model are trained. The outcome model is used to predict the user's natural purchase probability in the absence of voucher intervention, and the propensity score model is used to predict the probability that the user has been allocated vouchers historically, in order to quantify and control the selection bias introduced by historical distribution strategy preferences. S23. In the second stage, based on the outputs of the outcome model and the propensity score model, a residual-on-residual regression method is used to solve the problem. Specifically, a regression model is constructed with the outcome residual (i.e., the difference between the actual purchase result and the predicted value of the outcome model) as the dependent variable and the treatment residual (i.e., the difference between the actual treatment state and the predicted value of the propensity score model) as the independent variable. The individual causal gain value is then obtained. This value accurately reflects the net increase in the purchase probability of a specific user due to the consumption voucher.

[0007] Preferably, in step S3, the multi-objective composite reward function is constructed, specifically defined as a weighted sum of immediate causal gain reward, long-term value impact reward, and ecosystem health reward: in, These are adjustable weighting coefficients; S31. The instantaneous causal gain reward Its value is equal to the individual causal gain value calculated in step S2, which is used to guide the reinforcement learning model to maximize the direct business increment of a single coupon issuance behavior; S32. The long-term value impact on rewards The value represents the change in the expected Customer Lifetime Value (CLV) within a preset future time window after a user interacts with the recommended coupon. Preferably, the Customer Lifetime Value is predicted using a deep Pareto / NBD model. This model constructs a neural network that receives the user representation vector generated in step S1 as input and predicts personalized Pareto / NBD model parameters for each user, thereby achieving dynamic and accurate prediction of the user's future purchase frequency and activity probability. S33. Ecosystem health reward The penalty term includes a diversity penalty and a budget control penalty. The diversity penalty applies a negative reward when the model continuously recommends homogeneous vouchers to encourage strategy exploration and improve user experience. The budget control penalty applies a negative reward when the cost of issuing vouchers exceeds a preset budget to ensure the commercial viability of the strategy.

[0008] Preferably, in step S4, training the coupon issuance strategy based on a deep reinforcement learning model specifically includes: S41. The serialized coupon issuance process is modeled as a Markov Decision Process (MDP), where the State is formed by concatenating the target user representation vector generated in step S1 with the user's recent interaction history sequence; the Action Space is the set of all available coupons; the concatenation refers to concatenating two or more vectors in a dimensional manner to form a longer vector. The 'recent interaction history sequence' needs to be encoded into a fixed-dimensional historical vector before it can be input into the DRL model. This encoding process can be handled using sequence models such as Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), or Transformer encoders to capture the temporal dependencies of user behavior. S42. To address the challenge of the massive action space resulting from the large number of consumption vouchers, an action embedding method is employed. The policy network of the deep reinforcement learning model outputs an action vector with the same dimension as the consumption voucher representation vector. The final voucher issuance decision is determined by performing an efficient Approximate Nearest Neighbor Search in this consumption voucher representation vector space, i.e., selecting the consumption voucher with the highest cosine similarity to the action vector output by the policy network as the optimal action. S43. A deep reinforcement learning algorithm based on the actor-critic framework (such as SoftActor-Critic, SAC, or Twin Delayed DDPG, TD3) is used for policy learning. The critic network evaluates the expected value of the long-term cumulative compound reward obtained after performing a specific action (issuing a voucher) in a specific state; the actor network, as the policy network, aims to learn to select actions that maximize the critic network's evaluation value. The voucher issuance strategy is continuously optimized through offline training on historical interaction logs and continuous feedback from online services.

[0009] This application also provides a consumer voucher issuance control system based on big data, characterized in that it includes: The federated graph management module is used to receive model update parameters uploaded by each client in a federated learning architecture of central server-multiple clients, execute topology and distribution-aware federated aggregation algorithms, and generate and manage entity representation vectors of the global knowledge graph. The causal gain evaluation module is used to connect to the federated graph management module, receive entity representation vectors, and use the R-Learner meta-learner model to calculate individual causal gain values ​​for target users and candidate vouchers. The long-term value prediction module is used to connect to the federated graph management module. It uses a deep Pareto / NBD model to predict changes in the customer lifetime value based on the user's representation vector. A multi-objective reward synthesis module is used to connect the causal gain evaluation module and the long-term value prediction module, and calculates and generates a composite reward signal for reinforcement learning training based on its output and preset ecosystem health indicators. The strategy learning and execution module is used to train the coupon issuance strategy based on the composite reward signal using a deep reinforcement learning algorithm based on the actor-critic framework. During the execution phase, the module receives the user state, generates action embedding vectors by the policy network, and determines the specific consumer coupons to be issued by the end through approximate nearest neighbor search. The recommendation interface module is used to distribute the consumption vouchers determined by the strategy learning and execution module to target users through application programming interfaces or push notifications.

[0010] Compared with the prior art, this application has at least the following beneficial effects: This application achieves effective fusion and modeling of multi-source, heterogeneous, and non-independent identically distributed (Non-IID) big data within a privacy-preserving federated learning framework. By introducing a topology- and distribution-aware federated aggregation algorithm, it intelligently assigns aggregation weights to different data sources, significantly improving the training stability and final performance of the global model in real-world, complex data environments, thus laying a high-quality data foundation for accurate personalized coupon issuance.

[0011] Overcoming the short-sightedness of traditional reinforcement learning strategies, this application expands the optimization objective from pursuing immediate rewards to maximizing the user's long-term value throughout their entire lifecycle. By creatively quantifying the dynamic changes in long-term user value (CLV) and integrating them into the reward function, this application guides the coupon system to make decisions that balance short-term ROI and long-term user loyalty. This helps build healthy and sustainable user relationships and platform ecosystems, achieving a unity between business goals and user experience. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall process of a consumer voucher issuance control method based on big data, provided in an embodiment of this application.

[0013] Figure 2 This is a schematic diagram of the federated learning architecture and topology distribution-aware aggregation in the embodiments of this application.

[0014] Figure 3 This is a schematic diagram of the process for calculating individual causal gain (CATE) based on R-Learner in an embodiment of this application.

[0015] Figure 4 This is a schematic diagram of the multi-objective composite reward function in the embodiments of this application.

[0016] Figure 5 This is a schematic diagram of the deep reinforcement learning decision-making process based on the actor-critic framework in the embodiments of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] In the description of this application, unless otherwise stated, "multiple" means two or more. The term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0019] See also Figure 1 This document illustrates the overall flow of a big data-based consumer voucher issuance control method provided in an embodiment of this application. This method can be executed by a computer program deployed on a server or server cluster. In a specific application scenario, such as a large e-commerce platform, the platform aims to intelligently issue massive amounts of consumer vouchers from tens of thousands of merchants within its system to improve marketing budget efficiency and long-term user value. Specific steps of the method may include: Step S100: Data Fusion and Representation Learning. First, distributed user data deployed on multiple data terminals is collected to construct a heterogeneous knowledge graph containing users, vouchers, products, and their interaction relationships. Based on the knowledge graph, entity representation vectors that can represent user preferences and market dynamics are learned.

[0020] Step S200: Causal Gain Calculation. Next, based on the entity representation vector obtained in step S100, a causal inference model is established to identify the true causal effect of the consumer voucher intervention on user purchasing behavior, thereby calculating the individual causal gain value that can quantify the net increase in purchase probability for the target user and the consumer vouchers to be selected.

[0021] Step S300: Composite Reward Construction. Next, to guide the optimization of the coupon issuance strategy, a multi-objective composite reward function is constructed, which aligns with multiple core business objectives of the platform.

[0022] Step S400: Policy Learning and Decision Execution. Finally, the sequential distribution process of consumption vouchers is modeled as a Markov decision process, and a deep reinforcement learning agent is used to perform end-to-end policy learning and online decision-making with the goal of maximizing the long-term cumulative compound reward defined in step S300.

[0023] The above steps will be explained in detail below through two specific embodiments.

[0024] Example 1 This embodiment focuses on the implementation details of step S100, namely, how to effectively integrate multi-source heterogeneous big data and generate high-quality entity representation vectors while protecting privacy.

[0025] Please refer to the following: Figure 2 This demonstrates the central server-multiple client federated learning architecture used in this embodiment. In this architecture, the central server 201 is maintained by the platform (such as a large e-commerce platform). The multiple clients can be users' personal devices 202 (such as mobile phones with e-commerce apps installed), the business systems 203 of merchants (such as backend servers that manage their products and orders), or even third-party content platforms 204 that cooperate with the platform (such as social media or review websites).

[0026] First, a global heterogeneous knowledge graph needs to be defined. Node types in the graph can include: users, vouchers, products, categories, brands, merchants, livestreamers, etc.; edges represent the rich relationships between them, such as a user "clicking" a product, a user "claiming" a voucher, a user "following" a livestreamer, a voucher "applicable" to a certain category, a product "belonging" to a certain brand, etc. Importantly, this raw data is physically distributed. For example, a user's detailed browsing and click history is stored on the user's personal device 202, a merchant's product inventory and sales records are stored in their business system 203, and a third-party platform 204 holds the user's content consumption data on its platform. This architecture naturally ensures that the raw data of each party does not leave their local device, complying with increasingly stringent data privacy regulations.

[0027] The core of this embodiment lies in executing a topology- and distribution-aware federated aggregation algorithm to train a global graph neural network model (such as GraphSAGE, GAT, GIN, etc.), thereby learning a low-dimensional, dense representation vector (Embedding) for each entity node in the graph. In each round of communication during federated learning, the central server 201 distributes the current global model parameters to the clients participating in this round of training. Each client performs multiple rounds of training on its local subgraph and uploads the updated model parameters (or gradients).

[0028] When aggregating these parameters from different clients, the central server 201 does not simply calculate an average (as in the FedAvg algorithm), but instead performs a process such as... Figure 2 The intelligent weighting process is shown in the "aggregate weight calculation unit" 205. The core of this process is assigning a dynamic weight to the contribution of each client k. This weighting calculation takes into account two dimensions: Topological similarity: This dimension aims to assess the generalizability of a client's local graph structure. For example, the user behavior patterns (such as the browsing-compare-add to cart-purchase path) of client A (e.g., multiple user groups in a typical city) are similar to most other clients, while the graph structure of client B (e.g., a newly joined merchant with a small amount of data) is very sparse and specific. By calculating and comparing the spectral information of the Laplacian matrix of each client's subgraph (a mathematical tool reflecting the macroscopic characteristics of the graph structure), it can be considered that the structural knowledge of client A is more valuable for building a global model, and therefore it is given a higher weight.

[0029] Vector Distribution Dissimilarity: This dimension aims to assess the degree to which a client's node feature distribution deviates from the global average distribution. During each round of aggregation, the server can estimate the average distribution of a global node representation vector. If the node representation vector distribution uploaded by client C differs significantly from the global average distribution (measured by statistical distance such as Maximum Mean Difference (MMD)), it may indicate statistical anomalies in that client's data or a highly specific user group. To prevent it from excessively "slanting" the global model and affecting its generalization ability, the system will appropriately reduce its weight.

[0030] In this way, this embodiment can effectively overcome the non-independent and identically distributed (Non-IID) problem of data from different clients in terms of graph structure, node features, and label distribution, ensuring the stability and efficiency of model training. After multiple rounds of federated learning, the final entity representation vector integrates global information from all parties while preserving data privacy, providing high-quality, high-dimensional feature input for subsequent accurate causal inference and reinforcement learning decisions.

[0031] Example 2 This embodiment focuses on the implementation details of steps S200 to S400, namely, how to realize an end-to-end intelligent coupon issuance decision closed loop based on the representation vector generated in the previous step.

[0032] First, in step S200, the system calculates individual causal gains. The aim is to accurately identify "persuadable users"—those whose purchasing behavior will be positively changed by receiving the voucher—from all users interested in a particular voucher, thus ensuring that the budget is spent wisely.

[0033] See also Figure 3 This embodiment employs a meta-learner model called R-Learner to estimate individual causal gain (CATE). The model's input is the representation vector 301 of the target user and the candidate vouchers (generated in step S100). R-Learner separates the "causal" effect through a two-stage process: In the first phase, two auxiliary models are trained in parallel using historical coupon issuance and purchase data. The Outcome Model (302) predicts the probability of a user making a purchase based solely on their inherent characteristics without coupon intervention. For example, a highly active, brand-loyal user has a high probability of making a purchase. The PropensityScore Model (303) predicts the probability that a user has historically received a certain type of coupon from the system. This corrects for biases in historical distribution strategies, such as the system's past tendency to issue coupons to all high-spending users.

[0034] In the second stage, a residual-to-residual regression model 304 is used, leveraging the outputs of the two models from the first stage, to regress the user's "outcome residual" (the difference between actual purchase behavior and the probability of natural purchase) and "processing residual" (the difference between whether the coupon was actually received and the probability of being pushed to the user). This ingenious mathematical transformation effectively eliminates the influence of confounding variables, ultimately calculating a precise CATE value 305 for each (user, coupon) pair. A high CATE value indicates that the coupon is highly "persuasive" to the user, representing a potential high-ROI campaign.

[0035] Next, in step S300, the system constructs a multi-objective composite reward function. Traditional reinforcement learning coupon distribution models may only reward "redemption" behavior, but this can lead the system to tend to issue coupons without any threshold, or harm the user experience.

[0036] See also Figure 4 The composite reward R 401 designed in this embodiment is a weighted sum, designed to balance multiple core business objectives: Instant causal gain reward 402: This reward directly uses the CATE value calculated in step S200. If a user uses the recommended voucher and completes a purchase, the agent receives a positive reward equal to the CATE value. This directly guides the model to find "persuasive" users.

[0037] Long-term value impacts rewards 403: This is a key innovation of this application. Its goal is to quantify the long-term impact of a single coupon issuance action on the future value of users. To this end, this embodiment proposes a deep Pareto / NBD model to realize the impact on the long-term value of users ( The traditional Pareto / NBD model predicts a user's future activity probability based on historical transaction statistics (such as transaction frequency and the time of the most recent transaction), and assumes that its core parameters (such as r, α, s, β) follow a fixed prior probability distribution across the entire population. The key improvement of this invention is that it abandons this static assumption and instead utilizes the information-rich entity representation vector learned in step S1. Specifically, a small neural network is constructed, whose input is the target user's entity representation vector X. user Its output is directly a set of personalized Pareto / NBD model parameters (r', α', s', β') predicted for the user. In this way, CLV prediction no longer relies solely on sparse transaction history, but can utilize comprehensive and dynamically changing user profile information to make more accurate and personalized predictions of the user's long-term value. This refers to the change in the number of transactions a user is expected to make within a predetermined future time window (e.g., 90 days) after interacting with the recommended coupon, as predicted by the deep Pareto / NBD model. Thus, the reward... This can be calculated as the change in a user's expected number of transactions over the next 90 days after the interaction. A good coupon distribution experience (such as helping users discover new brands) should improve their long-term activity, thus generating a positive return. award.

[0038] Ecosystem health rewards 404: This is a negative penalty used to maintain the long-term health of the platform. For example, if the system continuously recommends the same brand of coupons to the same user, it will trigger a "diversity penalty" to encourage exploration; if the total cost of issuing coupons within the current time window exceeds the budget, it will trigger a "budget control penalty" to ensure business viability.

[0039] Finally, in step S400, the system performs strategy learning and decision execution.

[0040] See also Figure 5This embodiment employs an advanced actor-critic deep reinforcement learning framework. The entire decision-making process can be viewed as an interaction between two roles: actor and critic.

[0041] The user's current state 501 (consisting of its representation vector and recent behavior sequence) is fed into the Actor network 502 and the Critic network 503.

[0042] The actor's role is that of a "decision-maker," responsible for directly outputting an action based on the current state. However, in real-world e-commerce scenarios, the number of vouchers can reach millions, resulting in a massive action space. Therefore, the actor network outputs not a specific voucher ID, but an action embedding vector with the same dimension as the voucher's representation vector. The system then utilizes an efficient approximate nearest neighbor search technique (such as FAISS) to quickly find the one most similar to this action embedding vector in the vector library of all vouchers, which is then selected as the final voucher to be issued (voucher number 504).

[0043] After the voucher is recommended to a user, the user interacts with environment 505 (i.e., the e-commerce app) and generates feedback (purchase, click, or ignore). Based on this feedback, the system calculates the multi-objective composite reward R 401 defined in step S300.

[0044] The critic acts as an "evaluator," receiving the state and the actor's chosen action, and learning to predict the expected value (Q-score) of the long-term cumulative compound reward that the (state, action) can bring. This Q-score is used to guide the updates of the actor's network: if the critic considers an actor's action to be good (high Q-score), the actor will adjust its parameters to be more inclined to perform that action in similar states in the future; and vice versa.

[0045] Through offline training on massive amounts of historical user interaction logs and continuous learning in online services, the actor and critic network iterates and optimizes, eventually learning an intelligent strategy that can make optimal serialized coupon issuance decisions in complex environments to maximize long-term composite value.

[0046] This application also provides a consumer voucher issuance control system based on big data, which can implement the above-mentioned method. Embodiments of this system may include: The Federated Graph Management module is responsible for executing the logic of step S100, especially managing and updating the representation vectors of the global knowledge graph under the federated learning architecture.

[0047] The causal gain evaluation module is responsible for executing the logic of step S200. It has a built-in R-Learner model and calculates the CATE value.

[0048] The long-term value prediction module is responsible for executing the calculation logic for long-term value rewards in step S300, and has a built-in deep Pareto / NBD model.

[0049] The multi-objective reward synthesis module is responsible for integrating the outputs from the causal gain assessment module, the long-term value prediction module, and other business rules to calculate the final composite reward signal.

[0050] The strategy learning and execution module is the core of the system's decision-making process. It has a built-in Actor-Critic network and is responsible for executing the logic of step S400, outputting the final consumer voucher decision.

[0051] The recommendation interface module is responsible for reliably delivering the vouchers selected by the strategy learning and execution module to the target users through pop-ups, push notifications, or in-app messages in the platform app.

[0052] These modules work together to form a complete, closed-loop, and intelligent consumer voucher distribution control system.

[0053] It is understood that those skilled in the art will not limit this application to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered exemplary and non-limiting from any point of view.

[0054] To implement the above methods and systems, corresponding hardware system support and software architecture integration are required.

[0055] At the hardware system level, this application involves large-scale graph computation, causal inference, and the training and inference of deep reinforcement learning models, thus placing high demands on computing resources. A typical hardware deployment scheme may include: (1) Training server cluster: Used for offline training of graph neural network models, R-Learner causal inference models, and Actor-Critic reinforcement learning policy networks for the global knowledge graph. This cluster should be equipped with high-performance computing cards (GPUs), for example, a server node or node cluster consisting of 8 to 16 NVIDIA A100 or H100 Tensor Core GPUs. Such a configuration can provide petaFLOP-level computing power (for example, the FP32 computing power of a single A100 is about 19.5 TFLOPS, and an 8-card node can provide more than 150 TFLOPS of computing power) to meet the training needs of knowledge graphs with hundreds of millions of nodes and billions of edges and massive historical interaction logs.

[0056] (2) Online Inference Server: Used to process real-time coupon issuance requests. This server cluster has high requirements for low latency response. GPUs such as NVIDIA T4 or L4, which are optimized for inference, can be used. The number of servers scales dynamically based on the QPS (queries per second) of online requests. For example, on a large e-commerce platform with tens of millions of daily active users, dozens or even hundreds of inference servers may need to be deployed during peak periods to ensure that the response time for user requests is within 50 milliseconds.

[0057] (3) Federated Learning Center Server: As the coordinator of federated learning, this server is mainly responsible for the aggregation and distribution of model parameters and the calculation of dynamic weights. It has high requirements for CPU computing power, memory capacity and network I / O bandwidth. Servers equipped with multi-core (such as 64-core or 128-core) CPUs, TB-level memory and 10 Gigabit (10GbE) or higher speed network interfaces can be selected.

[0058] (4) Data storage system: used to store raw log data, processed intermediate data, knowledge graph data, and trained model files. Distributed storage systems, such as Hadoop HDFS, or object storage services provided by cloud service providers (such as Amazon S3 and Google Cloud Storage) can be used in conjunction with distributed databases (such as HBase) and data warehouses (such as Hive and ClickHouse) to support the storage and retrieval of massive amounts of data.

[0059] At the software architecture level, this application's system needs to be embedded as a service into the existing e-commerce platform's technology stack. The integration method is as follows: (1) Software Dependencies and Frameworks: This system can be developed based on mainstream deep learning frameworks (such as TensorFlow or PyTorch). Specifically, the graph neural network part can utilize the PyTorch Geometric (PyG) or DeepGraph Library (DGL) libraries; the federated learning implementation can be based on the TensorFlow Federated (TFF) or Flower framework; the causal inference model can utilize dedicated libraries such as EconML or CausalML; the reinforcement learning part can be based on RLlib or implemented by ourselves; and the efficient approximate nearest neighbor search uses the FAISS (Facebook AI Similarity Search) library. All services can be containerized using technologies such as Docker and orchestrated and managed uniformly by Kubernetes.

[0060] (2) Connection with e-commerce platforms and databases: The interaction between this system and existing platforms is mainly completed through API and service calls. On the one hand, it needs to periodically pull historical interaction logs that have been anonymized and desensitized from the platform's data warehouse or data lake (e.g., a user behavior database based on the Hadoop or Spark ecosystem) in batch processing for offline training and updating of the model. On the other hand, it needs to communicate in real time with the front-end application or back-end business system of the e-commerce platform. When a user triggers a coupon issuance scenario (such as opening the App or entering the activity page), the platform backend will send a request to the "recommendation interface module" of this system through internal RPC (Remote Procedure Call) or message queue (such as Apache Kafka). The request contains the user's real-time status information. After making a decision, this system returns the optimal coupon ID to the platform backend through the API, which is responsible for the final page rendering and presentation.

[0061] To verify the superiority of the technical solution proposed in this application, a comparative experiment was conducted with two representative traditional methods in a simulated e-commerce platform environment. The experiment lasted for 90 days and was carried out on a dataset containing 10 million users, 500,000 products, and 100,000 types of consumer vouchers.

[0062] Comparison Method 1: Collaborative Filtering (CF). This method employs the classic item-based collaborative filtering (Item-based CF) algorithm to recommend coupons associated with other items similar to those found in the user's historical behavior (such as purchases and add-to-cart). This represents a traditional, non-deep learning-based recommendation method.

[0063] Comparison Method Two: Short-Term Objective DRL. A standard Deep Reinforcement Learning (DRL) model is employed, with a structure similar to this application (using the Actor-Critic framework), but its reward function is designed only to optimize short-term business metrics, namely the user's coupon redemption rate (CVR). This represents a state-of-the-art but short-sighted AI approach.

[0064] Comparison metrics: (1) Marketing Return on Investment (ROI): Defined as (incremental sales revenue generated by vouchers - cost of vouchers) / cost of vouchers. This metric directly measures the efficiency of budget utilization.

[0065] (2) Customer Long-Term Value Lift (CLV Lift): This measures the percentage increase in average Customer Lifetime Value (CLV) of the experimental group compared to the control group at the end of the experimental period. CLV is estimated using the expected total purchase amount over the next 90 days. This metric measures the impact on long-term customer loyalty.

[0066] (3) Recommendation Diversity: This is measured by calculating the average number of product categories covered by the vouchers recommended to a single user over a period of time. This metric reflects user experience and the ability to avoid information cocoons.

[0067] The results of the comparative experiment are shown in the table below: The data in the table above shows that: Collaborative filtering has the lowest ROI because it cannot distinguish "inevitable buyers," resulting in a significant waste of budget. Furthermore, its recommendation logic may annoy users and have a slight negative impact on long-term value.

[0068] While short-term goal-oriented DRL methods improve ROI by optimizing redemption rates, they come at the cost of severely damaging long-term user value. In pursuit of short-term CVR, the model continuously recommends homogeneous, low-threshold popular product vouchers to users, resulting in extremely low recommendation diversity and ultimately user churn.

[0069] This application achieved optimal results across all key metrics. By introducing causal inference to precisely target "persuasive" users, its ROI improved by approximately 66.7% compared to the short-term goal of DRL. More importantly, by making CLV one of the core optimization objectives, this application not only avoided harming long-term user value but also achieved a significant improvement of over 15%. Simultaneously, the ecosystem health reward ensured the diversity of recommendations, providing users with a superior discovery experience. The experimental results fully demonstrate the significant beneficial effects of this application in improving marketing efficiency and promoting the long-term healthy development of the platform.

[0070] The foregoing has provided a detailed description of one embodiment of this application, but the description is merely a preferred embodiment and should not be construed as limiting the scope of this application. All equivalent variations and improvements made within the scope of this application should still fall within the patent coverage of this application.

Claims

1. A method for controlling the issuance of consumer vouchers based on big data, characterized in that, Includes the following steps: S1. In a central server-multi-client federated learning architecture, distributed user data with varying structures and distributions deployed on multiple data terminals are aggregated. A topology-aware federated aggregation algorithm is executed to construct and learn entity representation vectors for a global heterogeneous knowledge graph containing users, vouchers, goods, and their interaction relationships. When aggregating the model parameters of each client, the aggregation algorithm dynamically weights the data based on both the similarity of the topological structure and the differences in vector distribution. This dynamic weighting specifically includes: Based on the spectral information of the local subgraphs of each client, the topological similarity between the clients is calculated and measured; and based on the entity representation vectors uploaded by each client, the difference between them and the global average distribution is calculated using the maximum mean difference (MMD) metric. The weight values ​​for the dynamic weighting are determined based on both the topological similarity and the distribution differences. S2. Based on the entity representation vector obtained in step S1, establish a causal inference model to identify the true causal effect of the consumer voucher intervention on user purchasing behavior, and then calculate the individual causal gain value that can quantify the net increase in purchase probability for the target user and the consumer voucher to be selected. S3. To guide the optimization of the coupon issuance strategy, a multi-objective composite reward function is constructed. This reward function integrates the immediate causal gain directly related to the marketing investment return rate of a single coupon issuance behavior, the long-term value increment used to measure the future user retention and activity, and the ecological health index used to ensure the diversity of coupon issuance strategies and budget compliance. S4. The sequential distribution process of consumption vouchers is modeled as a Markov decision process, and a deep reinforcement learning agent is used to perform end-to-end policy learning with the goal of maximizing the long-term cumulative compound reward defined in step S3. This enables the agent to make dynamic decisions and output the optimal target consumption voucher when the real-time status of the target user is received.

2. The method according to claim 1, characterized in that, In step S2, calculating the individual causal gain value specifically includes: S21. The problem of issuing consumption vouchers is constructed as a causal inference task, wherein the individual causal gain is intended to measure the net boost effect on user purchasing behavior when a specific consumption voucher is issued compared to when it is not issued; S22. Using the R-Learner meta-learner model, and taking historical interaction data and the representation vector generated in step S1 as input, train an outcome model and a propensity score model, wherein the outcome model is used to predict the user's natural purchase probability in the absence of voucher intervention, and the propensity score model is used to predict the probability that the user has been allocated vouchers in the past. S23. Based on the outputs of the outcome model and the propensity score model, a regression model is constructed using the residual-to-residual regression method, with the outcome residual as the dependent variable and the treatment residual as the independent variable, and the individual causal gain value is obtained by solving the model.

3. The method according to claim 1, characterized in that, In step S3, the multi-objective composite reward function is constructed, specifically defined as the weighted sum of immediate causal gain reward, long-term value impact reward, and ecosystem health reward, i.e. ,in, These are adjustable weighting coefficients.

4. The method according to claim 3, characterized in that, The long-term value impact reward is the change in the expected customer lifetime value within a preset time window after a user interacts with the recommended voucher. The customer lifetime value is predicted using a deep Pareto / NBD model, characterized in that the deep Pareto / NBD model includes a neural network. The neural network receives the user representation vector generated in step S1 as input, predicts personalized Pareto / NBD model parameters for each user, and calculates the expected customer lifetime value based on the personalized parameters and the user's historical transaction data.

5. The method according to claim 3, characterized in that, The instant causal gain reward Its value is equal to the individual causal gain value calculated in step S2 and / or the ecosystem health reward. , is a penalty term that includes a diversity penalty and a budget control penalty, wherein the diversity penalty is used to impose a negative reward when the model continuously recommends homogeneous consumer vouchers, and the budget control penalty is used to impose a negative reward when the cost of issuing vouchers exceeds a preset budget.

6. The method according to claim 1, characterized in that, In step S4, the serialized distribution process of the consumption vouchers is modeled as a Markov decision process, specifically including: The state of the Markov decision process is defined as a state vector formed by concatenating the target user representation vector generated in step S1 with the user's recent interaction history sequence. Furthermore, the action space of the Markov decision process is defined as the set of all available vouchers.

7. The method according to claim 6, characterized in that, To address the decision-making challenges posed by the vast action space, step S4 involves dynamically deciding and outputting the optimal target consumption voucher, specifically including: Using an action embedding method, the policy network of the deep reinforcement learning agent outputs an action vector with the same dimension as the consumption voucher representation vector; In the representation vector space of all candidate vouchers, an approximate nearest neighbor search is performed to select the voucher with the highest similarity to the action vector output by the policy network as the optimal target voucher.

8. The method according to claim 1, characterized in that, In step S4, the deep reinforcement learning agent learns policies using an actor-critic framework. The critic network is used to learn to evaluate the expected value of the long-term cumulative compound reward that can be obtained after performing a specific action in a specific state. The actor network, as a policy network, aims to learn to select actions that maximize the expected value evaluated by the critic network.

9. A consumer voucher issuance control system based on big data, characterized in that, include: The federated graph management module is used to receive model update parameters uploaded by each client in a federated learning architecture of central server-multiple clients, execute topology and distribution-aware federated aggregation algorithms, and generate and manage entity representation vectors of the global knowledge graph. The causal gain evaluation module is used to connect to the federated graph management module, receive entity representation vectors, and use the R-Learner meta-learner model to calculate individual causal gain values ​​for target users and candidate vouchers. The multi-objective reward synthesis module is used to calculate and generate a composite reward signal for reinforcement learning training based on the individual causal gain value, the change in customer lifetime value predicted based on the deep Pareto / NBD model, and the preset ecosystem health indicators. The strategy learning and execution module is used to train the coupon issuance strategy based on the composite reward signal using a deep reinforcement learning algorithm based on the actor-critic framework, and to determine the specific consumer coupons to be issued in the final execution phase.

10. The system according to claim 9, characterized in that, The specific configuration of the strategy learning and execution module is as follows: it receives the user's state, generates an action embedding vector from its internal actor network, and determines the specific consumer voucher to be issued by performing an approximate nearest neighbor search.

Images