Consumption finance marketing strategy reaching model

Abstract

The invention discloses a consumer finance marketing strategy reaching model. According to the model, user behaviors and service data are acquired in real time through a data acquisition system, a dynamic user portrait is constructed by adopting a multi-head self-attention mechanism of a Transform model, a user behavior sequence is coded into a 512-dimensional dynamic interest vector, and the 512-dimensional dynamic interest vector and a 128-dimensional static attribute vector are fused to generate a comprehensive portrait. And training a strategy network based on a PPO reinforcement learning algorithm, modeling a marketing decision into a Markov decision process, and outputting a personalized strategy including a product type, a reach channel, an incentive limit and a reach opportunity. The system collects user feedback through full-link data burying points, maintains experience playback pools with the capacity of 1 million, carries out incremental learning every 4 hours, and adopts an elastic weight consolidation technology to avoid disastrous forgetting. A new strategy effect is verified through an A / B test, and automatic hyper-parameter tuning is carried out through Bayesian optimization. Practical application shows that the marketing conversion rate of the model is improved from 2.3% to 3.1%, the ROI is improved from 1.5 to 2.1, and the complaint rate is reduced by 40%.

Description

Technical Field

[0001] This invention relates to the field of consumer finance technology, and in particular to a consumer finance marketing strategy outreach model. Background Technology

[0002] With the rapid development of the consumer finance market and the increasing diversification of user needs, precision marketing has become a key means for financial institutions to improve business conversion rates and user experience. Traditional consumer finance marketing mainly relies on batch push and rule-driven methods, pushing financial product information to target users through preset user segmentation rules and fixed marketing strategies. However, this static marketing model often adopts a "one-size-fits-all" strategy, failing to fully consider users' personalized needs and dynamically changing interests and preferences, resulting in poor marketing effectiveness, a poor user experience, and the potential for user resentment and complaints due to excessive intrusion.

[0003] While existing marketing technologies have incorporated user profiling and recommendation algorithms to some extent, they still have many limitations. First, traditional user profiling is mostly based on static analysis of historical data, failing to capture real-time changes in user needs and leading to inaccurate marketing timing. Second, marketing strategy formulation relies heavily on human experience and simple rule engines, lacking intelligent decision-making mechanisms and struggling to cope with complex and ever-changing marketing scenarios. Third, channel selection and script design often use fixed templates, unable to be personalized based on user characteristics. Finally, most systems lack effective feedback and learning mechanisms, failing to continuously optimize strategies from historical marketing results and easily falling into the trap of "model aging."

[0004] A more critical technical challenge lies in the fact that consumer finance marketing is a typical sequential decision-making problem, requiring different marketing strategies at different stages of the user lifecycle. This involves complex temporal dependencies and long-term return optimization. Traditional supervised learning methods struggle to handle such dynamic decision-making problems, while simple rule systems cannot cope with high-dimensional feature spaces and complex user behavior patterns. Furthermore, how to improve marketing effectiveness while maintaining user experience and avoiding the negative impacts of over-marketing, and how to achieve continuous model learning and adaptive optimization, are all challenges that current technologies cannot effectively address. Therefore, there is an urgent need for a new marketing model that can perceive user status in real time, intelligently generate marketing strategies, and continuously self-optimize. Summary of the Invention

[0005] To address the technical problems of existing consumer finance marketing technologies, such as static user profiles, lack of intelligent strategy formulation, blind channel selection, and inability to continuously optimize models, this invention provides a consumer finance marketing strategy outreach model based on dynamic profiling and reinforcement learning closed-loop optimization. The aim is to achieve precise, personalized, and adaptive marketing strategies by constructing an intelligent closed-loop system of "perception-decision-execution-feedback-optimization".

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A consumer finance marketing strategy outreach model includes: The data acquisition system obtains user clickstream data, page dwell time, search keywords, product browsing paths, transaction records, credit application status, repayment records, credit limit usage, and static attribute data such as age, gender, occupation, and income level through data acquisition agents deployed in mobile apps, mini programs, transaction systems, and CRM systems, and transmits the data in real time through Apache Kafka message queues; The dynamic user profile construction module uses the Apache Flink stream processing engine to clean, deduplicatize, normalize to the [0,1] interval, and fill in missing values ​​for the received data. The user behavior sequence is encoded into a 512-dimensional vector through the Transformer model. The behavior weights are calculated using the multi-head self-attention mechanism of a 6-layer Transformer encoder. The 512-dimensional dynamic interest vector is concatenated with the 128-dimensional static attribute vector, and a 128-dimensional comprehensive user profile vector is generated through a three-layer fully connected neural network with 512, 256, and 128 nodes. The reinforcement learning strategy network models marketing decisions as a Markov decision process. It uses a 128-dimensional user profile vector as the state space and defines an action space that includes product type, marketing script template, reach channel, incentive coupon amount of 0-10000 yuan, and 24 time period selections. The PPO algorithm is used to train the strategy network with three hidden layers containing 256, 512, and 256 nodes. The discrete action probability distribution is output through the Softmax layer, and the continuous action is output through the Gaussian distribution parameters. The strategy execution and feedback collection module parses the strategy network output into execution instructions that include reach time, channel, product, script, and incentive amount. These instructions are then sent to the message platform, SMS gateway, outbound call platform, and advertising platform via API. The module also collects feedback data on push arrival, clicks, page visits, applications, conversions, and complaints through full-link data tracking. The closed-loop optimization module maintains an experience replay pool with a capacity of 1 million records. Every 4 hours, it samples 10,000 of the latest data and 10,000 of the historical data to perform incremental training on the PPO network for 20 epochs. Through A / B testing, users are split into an 80% control group and a 20% experimental group. The conversion rate, ROI, and complaint rate are monitored. When the experimental group significantly outperforms the control group for 7 consecutive days, the new strategy is promoted.

[0007] Furthermore, the multi-head self-attention mechanism of the Transformer model dynamically calculates the attention weight of each behavior to other behaviors in the sequence through three matrix operations: Query, Key, and Value. Each layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer. Residual connections and layer normalization ensure training stability, and position encoding adds temporal information to each behavior vector.

[0008] Furthermore, the reward function of the PPO algorithm is defined as follows: click reward +1, application reward +5, successful conversion reward +10, ignore reward -0.5, unsubscribe reward -3, complaint reward -5. The KL divergence between the new and old strategies is limited by importance sampling and pruning techniques. A value network V(s) is introduced to estimate the state value, and the advantage function A(s,a)=Q(s,a)-V(s) is calculated. The learning rate is initially 0.0003 and decays as the training progresses.

[0009] Furthermore, the strategy execution module includes a frequency control management mechanism, which sets a maximum of 3 marketing outreaches per user per day and no more than 10 per week. It also includes a filtering module for high-risk user exclusion sets, which automatically filters users who refuse to receive marketing or have a complaint record. Additionally, it includes a time-slot control mechanism to avoid making phone calls or sending text messages between 10 p.m. and 8 a.m.

[0010] Furthermore, the closed-loop optimization module adopts the priority experience replay technique, assigns sampling weights to experience samples based on TD error, and uses the elastic weight consolidation EWC technique to add a regularization term to the loss function to limit the variation of important parameters. The regularization strength is dynamically adjusted by estimating the Fisher information matrix.

[0011] Furthermore, the closed-loop optimization module includes a meta-learning mechanism. Through Bayesian optimization, it automatically tunes the learning rate (range 1e-5 to 1e-2), the PPO pruning parameter ε (range 0.1 to 0.3), the reward function weights, the number of neural network layers, and the number of nodes. Meta-optimization is initiated once a week. The relationship between hyperparameters and model performance is modeled through Gaussian processes. The expected improvement EI is used as the acquisition function, and the optimal combination of hyperparameters is found after 20-30 iterations.

[0012] Furthermore, the feedback data includes push notification arrival, push clicks, and message viewing duration at the message level; landing page visits, page dwell time, in-page click behavior, and form filling progress at the page level; product application, approval, credit limit usage, and repayment behavior at the business level; as well as negative feedback such as unsubscribing, complaints, and message deletion. All feedback data is timestamped and associated with specific touchpoints via session IDs, forming a quadruple of state s, action a, reward r, and next state s'.

[0013] Furthermore, the message queue employs a partitioning mechanism and replication strategy to ensure high availability, reduces network transmission overhead through data compression, and maintains millisecond-level latency. A / B testing uses statistical significance testing, and when the p-value is less than 0.05, the experimental group is determined to be significantly better than the control group. Monitoring indicators include reach conversion rate, application conversion rate, final conversion rate, user lifetime value (LTV), and marketing return on investment (ROI).

[0014] This invention constructs a complete intelligent marketing system. The system first acquires user behavior and business data in real time through a multi-source data collection system. Specifically, the system deploys a data collection SDK on user terminals such as mobile apps and mini-programs to capture users' real-time behavioral trajectories, including dynamic data such as clickstreams, page dwell times, and search records. Simultaneously, through interfaces with backend transaction and CRM systems, it obtains business data such as transaction records, credit status, and repayment status, as well as basic user attribute information. All data is transmitted in real time via the Apache Kafka message queue system to ensure data timeliness and integrity.

[0015] In user profile construction, this invention innovatively introduces a Transformer-based temporal modeling technique. The system first utilizes the Apache Flink stream processing engine to preprocess the raw data in real time, including data cleaning, deduplication, standardization, and missing value imputation. Then, the user's behavior sequence is input into the Transformer model, where a multi-head self-attention mechanism with a 6-layer encoder dynamically calculates the contribution weights of different behaviors to the user's current interest state, generating a 512-dimensional dynamic interest vector. This vector is concatenated with a 128-dimensional static attribute vector and then fused using a three-layer fully connected neural network, ultimately generating a 128-dimensional comprehensive user profile vector, achieving a comprehensive and dynamic characterization of the user's state.

[0016] The marketing strategy generation employs an intelligent decision-making mechanism based on reinforcement learning. The system models marketing outreach as a Markov decision process, using user profile vectors as state input and defining an action space encompassing multiple dimensions such as product type, marketing message, outreach channels, incentive amount, and outreach timing. A deep strategy network with three hidden layers is trained using the PPO algorithm, capable of outputting the optimal combination of marketing strategies based on user states. The reward function comprehensively considers both positive and negative user feedback (clicks, applications, conversions) and negative feedback (ignores, unsubscriptions, complaints), as well as marketing costs, achieving a balance between short-term conversion and long-term value.

[0017] Strategy execution and feedback collection form a closed-loop operation system. The system parses the output of the strategy network into specific execution instructions, which are then distributed to various channels via standardized APIs. Simultaneously, it implements safeguards such as frequency control management, filtering modules for high-risk user groups, and time-based control. Through end-to-end data tracking technology, the system collects user feedback data in real time across messages, pages, and business processes, forming comprehensive experience data. This data is stored in a 1 million-entry experience replay pool. The system performs incremental learning every 4 hours, employing prioritized experience replay and elastic weight consolidation techniques to ensure continuous model evolution while avoiding catastrophic forgetting.

[0018] This invention also introduces A / B testing and meta-learning mechanisms to achieve automatic system optimization. By dividing user traffic into control and experimental groups and continuously monitoring key business indicators, the new strategy is automatically promoted when it is significantly better than the old one. Simultaneously, Bayesian optimization techniques are used to automatically tune system hyperparameters, including learning rate, network architecture, and reward weights, ensuring the system always maintains optimal performance.

[0019] This invention, by constructing a closed-loop marketing system based on dynamic profiling and reinforcement learning, transforms marketing strategies from static rules to intelligent decision-making, offering the following significant advantages: First, through the temporal modeling and attention mechanism of the Transformer model, it can capture dynamic changes in user needs in real time, significantly improving the accuracy of marketing timing. Second, the policy network based on reinforcement learning can automatically learn the optimal combination of marketing strategies, achieving personalized matching of products, channels, communication scripts, and incentives, significantly improving marketing conversion rates. Third, through continuous feedback learning and incremental optimization mechanisms, the system can continuously evolve, avoiding model aging and maintaining long-term marketing effectiveness. Fourth, the introduction of multiple guarantee mechanisms effectively balances marketing effectiveness and user experience, reducing user complaint rates. Fifth, meta-learning and A / B testing mechanisms enable automatic system tuning, reducing manual intervention and improving operational efficiency. Practical application shows that the system using this invention can increase the marketing conversion rate from 2.3% to 3.1%, the marketing ROI from 1.5 to 2.1, and reduce the user complaint rate by 40%, fully verifying the technological advancement and practical value of this invention. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall system architecture of the present invention, illustrating the complete system architecture from data acquisition to closed-loop optimization; Figure 2 The flowchart for constructing the dynamic user profile of this invention illustrates the process of generating the user profile. Figure 3 This is a network structure diagram of the reinforcement learning policy of the present invention, which shows the internal structure of the PPO policy network; Figure 4 The flowchart for strategy execution and feedback collection of the present invention describes the complete process of strategy from decision-making to execution. Figure 5 This is a schematic diagram of the closed-loop optimization mechanism of the present invention, showing the four key components of the system's self-optimization. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0023] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0024] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0025] See Figures 1 to 5 As shown This invention provides a consumer finance marketing strategy outreach model and system based on dynamic profiling and reinforcement learning closed-loop optimization. This system acquires multi-dimensional user data in real time to construct dynamic user profiles, generates personalized outreach strategies using a reinforcement learning strategy network, and continuously optimizes through feedback data, forming an intelligent closed-loop system of "perception-decision-execution-feedback-optimization". The specific implementation methods of this invention will be described in detail below.

[0026] Step 1: Specific Implementation of Real-Time Acquisition of Multi-Dimensional User Data and Historical Marketing Feedback Data In this implementation, the data acquisition system achieves comprehensive data collection through data acquisition agents deployed at multiple data source points. Specifically, at the user terminal level, this includes data acquisition SDKs embedded in mobile apps and mini-programs. These SDKs capture real-time user behavior data, including but not limited to dynamic behavior data such as clickstream data, page dwell time, search keywords, product browsing paths, and function usage frequency. At the business backend level, through interfaces with the transaction system and CRM system, real-time user transaction data is obtained, including real-time transaction records, credit application status, repayment records, and credit limit usage. Simultaneously, static attribute data of users, such as age, gender, occupation, income level, and asset status, is extracted from the CRM system. Historical marketing feedback data is extracted from the marketing management platform, including feedback information such as historical push message open rates, click-through rates, conversion rates, rejection rates, and user complaint records. All collected data is transmitted in real-time through a message queue system (such as Apache Kafka) to ensure data timeliness and integrity. The message queue employs a partitioning mechanism and a replication strategy to ensure high availability and reliability of data transmission. At the same time, it reduces network transmission overhead through data compression technology, maintaining millisecond-level latency throughout the entire data acquisition process, thus providing a data foundation for subsequent real-time processing and decision-making.

[0027] Step 2: Specific implementation of building and updating dynamic user profiles in real time based on temporal neural networks and attention mechanisms This step is one of the core components of the entire system, and its implementation process is divided into three key sub-steps. The first is the data preprocessing stage, where the system uses a stream processing engine (such as Apache Flink) to process the raw data received from the message queue in real time. This includes data cleaning (removing outliers and noisy data), data deduplication (identifying duplicate data based on user ID and timestamp), data standardization (normalizing data of different dimensions and mapping them to the [0,1] interval), and missing value imputation (using mean imputation or context-based intelligent imputation strategies).

[0028] Secondly, user behavior sequence modeling based on the Transformer model is a key technology for dynamic profile construction. The system treats a series of user behaviors as a temporal sequence. For example, a user behavior sequence can be represented as [browsing product A, saving product B, applying for a loan C, checking credit limit D, comparing interest rates E]. Each behavior contains multi-dimensional features such as behavior type, occurrence time, duration, and associated objects. The Transformer model uses a multi-head self-attention mechanism to calculate the contribution weight of different behaviors in the sequence to the current user's interest state. In specific implementation, the model first encodes each behavior into a d-dimensional vector (usually d=512), then adds temporal information to each behavior vector through positional encoding, and then performs feature extraction and information fusion through a multi-layer Transformer encoder (usually 6 layers). Each layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer, and residual connections and layer normalization techniques are used to ensure the training stability of the model. The core of the self-attention mechanism lies in dynamically calculating the attention weight of each element in the sequence to other elements through the operation of three matrices: Query, Key, and Value. This captures long-distance dependencies and complex interaction patterns between behaviors, and ultimately outputs a dynamic user interest vector that integrates temporal information and behavioral associations.

[0029] Finally, the system generates a comprehensive user profile vector. It concatenates the dynamic interest vector (512-dimensional) output by the Transformer model with the user's static attribute vector (containing encoded features such as age, gender, and occupation, with a dimension of 128) to form a 640-dimensional joint feature vector. This vector is then fused and dimensionality reduced using a three-layer fully connected neural network (with 512, 256, and 128 hidden layer nodes respectively), ultimately generating a 128-dimensional comprehensive user profile vector. This vector comprehensively depicts the user's current state, preferences, and potential needs, and is updated in real-time through an incremental learning mechanism as new data arrives, ensuring the timeliness and accuracy of the profile.

[0030] Step 3: Specific Implementation of Generating Personalized Outreach Strategies Based on Reinforcement Learning Policy Networks This step models marketing outreach decisions as a Markov Decision Process (MDP), using reinforcement learning to generate intelligent strategies. Within the MDP framework, the system defines the following core elements: the state space S is composed of a 128-dimensional comprehensive user profile vector generated in step 2, fully describing the user's current state; the action space A is a multi-dimensional discrete-continuous hybrid space, including dimensions such as recommended product types (e.g., discrete selection of consumer loans, cash loans, credit card installments), marketing script templates (selected from a pre-set template library), outreach channels (Push notifications, SMS, in-app messages, phone calls, etc.), incentive coupon amounts (continuous values, ranging from 0-10,000 yuan), and outreach timing (selection of 24 time slots within a day); the reward function R(s,a) is defined based on user feedback behavior. Positive feedback such as clicks (reward +1), applications (reward +5), and successful conversions (reward +10) receive positive rewards, while negative feedback such as ignoring (reward -0.5), unsubscribing (reward -3), and complaining (reward -5) receive negative rewards. Marketing costs are also considered as a penalty in the reward function.

[0031] The policy network is trained using the Proximal Policy Optimization (PPO) algorithm. PPO ensures training stability and avoids drastic fluctuations in policy performance by limiting the magnitude of policy updates. Specifically, the policy network π(a|s) employs a deep neural network architecture. The input layer receives a 128-dimensional user state vector, which undergoes feature extraction through three hidden layers (256, 512, and 256 nodes respectively, with ReLU activation). Finally, multiple output heads output probability distributions or values ​​for different action dimensions. For discrete actions (such as product type and channel selection), a Softmax layer is used to output the probability distribution; for continuous actions (such as incentive coupon amount), the mean and standard deviation of a Gaussian distribution are used as outputs, thus achieving modeling of the mixed action space.

[0032] During training, the system collects experience data through continuous interaction with the environment (i.e., real users). The four-tuple (state s, action a, reward r, next state s') generated from each interaction is stored in the experience replay pool. The PPO algorithm updates the policy network parameters by optimizing the following objective function: maximizing the expected cumulative reward while limiting the KL divergence between the old and new policies through importance sampling and pruning techniques to ensure the smoothness of policy updates. Simultaneously, a value network V(s) is introduced to estimate state values, used to calculate the advantage function A(s,a)=Q(s,a)-V(s), further improving the accuracy and sample efficiency of policy gradient estimation. The entire training process employs mini-batch stochastic gradient descent, with an adaptive learning rate adjustment strategy. The initial learning rate is 0.0003, gradually decreasing as training progresses.

[0033] Step 4: Specific implementation of the outreach strategy and real-time collection of user feedback data After receiving the action vector output by the policy network, the strategy decision engine parses it into specific execution instructions. For example, a typical execution instruction might be: "At 2 PM, push a consumer loan product X with a credit limit of 5,000 yuan to user A via the APP Push channel, using the script template Y (e.g., 'Exclusive offer: 5,000 yuan credit limit available immediately, daily interest as low as 0.02%')." These instructions are distributed to various channel outreach systems through standardized API interfaces, including the message middle platform (responsible for Push and in-app message push), SMS gateway (responsible for SMS sending), outbound call platform (responsible for telemarketing), and advertising platform (responsible for displaying advertisements).

[0034] To ensure accurate outreach, the system implements multiple safeguards: First, frequency control management, which maintains user-level outreach frequency records to avoid excessive disruption to the same user, for example, setting each user to receive a maximum of 3 marketing outreaches per day and no more than 10 per week; second, a filtering module for high-risk user groups, which automatically filters out users who explicitly refuse to receive marketing information or have a history of complaints; and third, time-based control, which avoids making phone calls or text messages during rest periods (such as 10 pm to 8 am) based on regulatory requirements and user experience considerations.

[0035] User feedback data is collected through end-to-end data tracking technology, deploying data collection points across various outreach channels and user interfaces to capture user feedback behavior in real time. This includes: message-level feedback (push notification arrival, push notification clicks, message viewing duration), page-level feedback (landing page visits, page dwell time, in-page clicks, form completion progress), business-level feedback (product application, approval, credit limit usage, repayment behavior), and negative feedback (unsubscription, complaints, message deletion). All feedback data is timestamped and includes a session ID, enabling precise association with specific outreach actions and forming complete (state s, action a, reward r, next state s') experience data, which is written to the experience replay pool in real time, providing data support for continuous model optimization.

[0036] Step 5: Implementing a closed-loop operation based on a strategy model that iteratively optimizes the feedback data. The closed-loop optimization mechanism is the key to the self-evolution of this invention, and it mainly includes two core components: a model continuous learning mechanism and an effect evaluation and meta-optimization system.

[0037] For continuous model learning, the system employs an incremental learning strategy to avoid the "catastrophic forgetting" problem. Specifically, the system maintains an experience replay pool with a capacity of 1 million records and uses Prioritized Experience Replay (PRIR) to assign different sampling weights to experience samples based on the TD error, allowing the model to learn more frequently from samples with larger prediction errors. Every 4 hours, the system samples a mixed batch of 10,000 recent data records and 10,000 historical data records from the PRIR pool for incremental training of the PPO policy network. The training process lasts for 20 epochs, with each epoch containing 100 mini-batch updates. To maintain the retention of historical knowledge, the system uses Elastic Weight Consolidation (EWC) technology. This involves adding a regularization term to the loss function to limit the variation of important parameters. The regularization strength is dynamically adjusted based on the importance of the parameters to the historical task (estimated via the Fisher information matrix).

[0038] The performance evaluation and meta-optimization system uses an A / B testing platform to verify the effectiveness of the new strategy. The system distributes user traffic according to a set ratio (typically 80% in the control group using the old strategy and 20% in the experimental group using the new strategy), continuously monitoring key business metrics for both groups, including reach conversion rate (clicks / reaches), application conversion rate (applications / clicks), final conversion rate (successful loan disbursements / applications), lifetime value (LTV, the profit contributed by a user over their entire lifecycle), return on investment (ROI = (revenue from marketing - marketing costs) / marketing costs), and user experience metrics (complaint rate, cancellation rate). When the experimental group's overall performance significantly outperforms the control group (through statistical significance testing, p-value < 0.05) for more than 7 days, the system automatically promotes the new strategy to all users.

[0039] The meta-learning mechanism automatically tunes the system's hyperparameters using Bayesian optimization techniques. These parameters include the learning rate (search range 1e-5 to 1e-2), the PPO pruning parameter ε (search range 0.1 to 0.3), the weights of various components in the reward function (relative proportions of conversion rewards, cost penalties, user experience, etc.), and neural network architecture parameters (number of hidden layers, number of nodes per layer). The system initiates a meta-optimization process weekly, modeling the relationship between hyperparameters and model performance using a Gaussian process. Expected Improvement (EI) is used as the acquisition function to guide the hyperparameter search. After 20-30 iterations, the optimal hyperparameter combination is found, and the model is then retrained with the new hyperparameter configuration. Example

[0040] To better illustrate the implementation effects of the present invention, a specific application example is provided below to explain in detail the application of the present invention in a real consumer finance marketing scenario.

[0041] A consumer finance company has deployed the marketing strategy outreach model system described in this invention to improve the marketing effectiveness of its consumer loan products. The company has 5 million registered users and approximately 500,000 daily active users, offering a variety of financial products including installment payments, cash loans, and credit card installment plans.

[0042] After system deployment, the marketing outreach process for a typical user, Mr. Zhang (User ID: USER_123456), was as follows: The system first obtained Mr. Zhang's real-time data through the data collection agent, including that he had browsed travel product pages 5 times in the past 7 days, spent more than 2 hours in travel forums, searched for keywords such as "European travel guide," and checked his credit card limit 3 times without applying. Simultaneously, his static attributes were obtained: 35 years old, male, IT professional, monthly income of 15,000 yuan, existing credit card limit of 30,000 yuan, and a historical on-time repayment rate of 100%.

[0043] Based on this data, the Transformer model identified that Mr. Zhang has a strong recent desire to spend on travel, and the dynamic interest vector shows that his interest weight for travel-related financial products reaches 0.78. The comprehensive user profile vector further reflects that he is a potential high-quality customer with good credit, stable income, and moderate price sensitivity.

[0044] Based on this user profile, the PPO strategy network output the optimal outreach strategy: at 6 PM on Friday (rush hour), push a marketing message via app push offering "Exclusive loan for European travel, limit of 8000 yuan, interest-free installments," along with a 2% interest rate coupon. After the system executed this strategy, Mr. Zhang clicked to view the message within 5 minutes of receiving it, stayed on the landing page for 3 minutes, filled out the loan application form, and was successfully approved. He ultimately used the 6000 yuan limit to purchase airfare and hotel accommodations for a European trip.

[0045] This successful outreach generated positive feedback (reward value +10), and the relevant experience data was stored in the experience replay pool. In subsequent model updates, the system learned the effectiveness of the strategy of "pushing highly relevant product + discount combinations to users with clear purchasing intentions during their active periods." After three months of continuous operation and optimization, the company's overall marketing conversion rate increased from 2.3% to 3.1%, marketing ROI increased from 1.5 to 2.1, and user complaint rate decreased by 40%, fully validating the effectiveness and practical value of the invention.

[0046] As can be seen from the above specific implementation methods and embodiments, the consumer finance marketing strategy outreach model provided by the present invention can realize the dynamic updating of user profiles, the intelligent generation of marketing strategies, and the continuous optimization of the model, forming a complete intelligent marketing closed-loop system. It effectively solves the problems of static strategies, blind channel selection, and model forgetting that exist in traditional marketing models, and has important application value and promotion prospects.

[0047] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0048] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A consumer financial marketing strategy reach model, characterized in that, include: The data acquisition system obtains user clickstream data, page dwell time, search keywords, product browsing paths, transaction records, credit application status, repayment records, credit limit usage, and static attribute data such as age, gender, occupation, and income level through data acquisition agents deployed in mobile apps, mini programs, transaction systems, and CRM systems, and transmits the data in real time through Apache Kafka message queues; The dynamic user profile construction module uses the Apache Flink stream processing engine to clean, deduplicatize, normalize to the [0,1] interval, and fill in missing values ​​for the received data. The user behavior sequence is encoded into a 512-dimensional vector through the Transformer model. The behavior weights are calculated using the multi-head self-attention mechanism of a 6-layer Transformer encoder. The 512-dimensional dynamic interest vector is concatenated with the 128-dimensional static attribute vector, and a 128-dimensional comprehensive user profile vector is generated through a three-layer fully connected neural network with 512, 256, and 128 nodes. The reinforcement learning strategy network models marketing decisions as a Markov decision process. It uses a 128-dimensional user profile vector as the state space and defines an action space that includes product type, marketing script template, reach channel, incentive coupon amount of 0-10000 yuan, and 24 time period selections. The PPO algorithm is used to train the strategy network with three hidden layers containing 256, 512, and 256 nodes. The discrete action probability distribution is output through the Softmax layer, and the continuous action is output through the Gaussian distribution parameters. The strategy execution and feedback collection module parses the strategy network output into execution instructions that include reach time, channel, product, script, and incentive amount. These instructions are then sent to the message platform, SMS gateway, outbound call platform, and advertising platform via API. The module also collects feedback data on push arrival, clicks, page visits, applications, conversions, and complaints through full-link data tracking. The closed-loop optimization module maintains an experience replay pool with a capacity of 1 million records. Every 4 hours, it samples 10,000 of the latest data and 10,000 of the historical data to perform incremental training on the PPO network for 20 epochs. Through A / B testing, users are split into an 80% control group and a 20% experimental group. The conversion rate, ROI, and complaint rate are monitored. When the experimental group significantly outperforms the control group for 7 consecutive days, the new strategy is promoted.

2. The consumer financial marketing strategy reach model of claim 1, wherein, The multi-head self-attention mechanism of the Transformer model dynamically calculates the attention weight of each behavior to other behaviors in the sequence through three matrix operations: Query, Key, and Value. Each layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer. Residual connections and layer normalization ensure training stability, and position encoding adds temporal information to each behavior vector.

3. The consumer financial marketing strategy reach model of claim 1, wherein, The reward function of the PPO algorithm is defined as: click reward +1, application reward +5, successful conversion reward +10, ignore reward-0.5, unsubscribe reward-3, complaint reward-5, the KL divergence of new and old strategies is limited by importance sampling and clipping technology, the value network V(s) is introduced to estimate the state value, the advantage function A(s,a)=Q(s,a)-V(s) is calculated, and the learning rate initial value is 0.0003 and decays with the training process.

4. The consumer financial marketing strategy reach model of claim 1, wherein, The policy execution module includes a frequency control management mechanism, which sets that each user can receive marketing touch at most 3 times a day and no more than 10 times a week, a filtering module containing a high-risk exclusion user set, which automatically filters users who refuse to receive marketing or have complaint records, and a time period control mechanism to avoid phone or SMS touch from 10pm to 8am.

5. The consumer financial marketing strategy reach model of claim 1, wherein, The closed-loop optimization module uses the priority experience replay technology to assign sampling weights to experience samples according to TD error, and uses the elastic weight consolidation (EWC) technology to add a regularization term to the loss function to limit the variation amplitude of important parameters, and the regularization strength is dynamically adjusted through the Fisher information matrix estimation.

6. The consumer financial marketing strategy reach model of claim 1 or 5, wherein, The closed-loop optimization module includes a meta-learning mechanism that automatically tunes the learning rate in the range of 1e-5 to 1e-2, the PPO clipping parameter in the range of 0.1 to 0.3, the reward function weight, the number of neural network layers and nodes through Bayesian optimization, and starts meta-optimization once a week. Through Gaussian process modeling of the relationship between hyperparameters and model performance, use expected improvement (EI) as the acquisition function, and find the optimal hyperparameter combination after 20-30 iterations.

7. The consumer financial marketing strategy reach model of claim 1, wherein, The feedback data includes message level push arrival, push click, message viewing time, page level landing page access, page dwell time, in-page click behavior, form filling progress, business level product application, approval, credit usage, repayment behavior, and negative feedback such as unsubscribing, complaining, and deleting messages. All feedback data is associated with specific touch actions with timestamps and session IDs to form a four-tuple of state s, action a, reward r, and next state s'.

8. The consumer financial marketing strategy reach model of claim 1, wherein, The message queue uses partitioning and replication strategies to ensure high availability, reduces network transmission overhead through data compression, and maintains millisecond-level latency; the A / B test passes the statistical significance test, and when the p-value is less than 0.05, it is determined that the experimental group is significantly better than the control group. The monitoring indicators include touch conversion rate, application conversion rate, final conversion rate, user lifetime value (LTV), and marketing return on investment (ROI).

Images