Personalized reward distribution method and system based on deep reinforcement learning

Abstract

The invention relates to a personalized reward distribution method and system based on deep reinforcement learning, and the method comprises the steps: collecting the feature data and reward pool information of a user through a distributed data collection frame in response to a request of participating in a reward distribution activity initiated by the user through a client, and carrying out the standardization processing to generate structured input data; processing the structured input data by using a deep reinforcement learning model, generating a personalized reward distribution probability, and ensuring that the distribution cost meets a preset budget constraint through a constraint optimization algorithm; and according to the personalized reward distribution probability, reward distribution requested by the user is executed by adopting a randomization algorithm, and a distribution result is returned. According to the method, personalized reward distribution is realized through deep reinforcement learning, and the user participation degree and satisfaction degree are remarkably improved; the real-time performance and the dynamic adaptability are ensured by a distributed data processing and online learning mechanism; the constraint optimization algorithm effectively controls the distribution cost and is suitable for large-scale application.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a personalized reward distribution method and system based on deep reinforcement learning. Background Art

[0002] In user marketing activities, reward distribution systems are widely used to increase user engagement and brand loyalty. Traditional reward distribution methods mostly use fixed probability or random distribution strategies, which makes it difficult to dynamically adjust the distribution plan according to the user's personalized characteristics, resulting in poor user experience and waste of resources. In recent years, data-driven personalization technologies have gradually emerged, such as recommendation systems based on user profiles, but their application in reward distribution still has limitations, such as the lack of real-time optimization mechanisms and cost control capabilities. In existing technologies, some methods attempt to introduce machine learning, but they mostly rely on static models and have difficulty adapting to dynamic changes in user behavior. In addition, when processing multi-dimensional user data such as behavior sequences and preference information, traditional methods have insufficient data processing efficiency and model accuracy, and lack an effective budget constraint mechanism.

[0003] To address the above issues, there is an urgent need for a technical solution that can integrate multi-dimensional user data, dynamically generate personalized allocation probabilities, and achieve real-time optimization under budget constraints to improve the accuracy and efficiency of reward allocation and meet the needs of modern marketing scenarios. Summary of the Invention

[0004] The purpose of the present invention is to provide a personalized reward distribution method and system based on deep reinforcement learning to solve the problem that the existing technology model has poor longitude and cannot control the budget.

[0005] To achieve one of the above-mentioned objectives, an embodiment of the present invention provides a personalized reward distribution method based on deep reinforcement learning, the method comprising:

[0006] In response to a user's request to participate in a reward distribution activity through the client, the user's feature data and reward pool information are collected through a distributed data collection framework, and standardized to generate structured input data;

[0007] Use deep reinforcement learning models to process structured input data, generate personalized reward allocation probabilities, and use constrained optimization algorithms to ensure that the allocation costs meet the preset budget constraints;

[0008] Based on the personalized reward allocation probability, a randomized algorithm is used to execute the reward allocation requested by the user and return the allocation result.

[0009] As a further improvement of an embodiment of the present invention, the method further includes that collecting the user's feature data and reward pool information through the distributed data collection framework includes:

[0010] The distributed data collection framework collects user feature data and reward pool information in real time through streaming data processing technology, where:

[0011] The user characteristic data includes demographic characteristics, device information, behavioral sequence data, consumption records and preference information;

[0012] The reward pool information includes the value of the reward items, the distribution popularity and the user's preference information for the reward items;

[0013] The data collection further includes data cleaning of the user feature data and the reward pool information, wherein the data cleaning removes invalid data through an anomaly detection algorithm.

[0014] As a further improvement of an embodiment of the present invention, the method further includes that the performing standardization processing to generate structured input data includes:

[0015] The standardization process includes applying a normalization algorithm to convert the user feature data and the reward pool information after data cleaning into a unified dimension;

[0016] Using embedding technology to encode the user feature data converted to a unified dimension and the non-numerical data in the reward pool information into a numerical vector;

[0017] The numerical vectors described in the distributed data processing framework are batch processed and format converted to generate a structured feature matrix suitable for deep neural network input.

[0018] As a further improvement of an embodiment of the present invention, the method further includes that the processing of structured input data using a deep reinforcement learning model to generate personalized reward allocation probabilities includes:

[0019] Encoding the structured input data into a state vector using an embedding technique;

[0020] The state vector is processed by a multi-layer deep neural network to generate a non-normalized weight for each reward item j ;

[0021] The probability of allocating each reward item is calculated based on the non-normalized weight and reward value, and the formula is:

[0022]

[0023] in, is the probability of assigning the jth reward, is the target expected cost, is the price of the kth prize, k is the prize index, and n is the total number of prizes.

[0024] As a further improvement of an embodiment of the present invention, the method further includes that ensuring that the allocation cost satisfies the preset budget constraint through the constraint optimization algorithm includes:

[0025] Applying a constrained loss function To ensure that the total expected cost and the probability of a single item do not exceed the preset upper limit, the constraint loss function is expressed as:

[0026]

[0027] in, is the total expected cost of all reward items, The maximum allowable allocation probability for a single reward project, is the constraint strength coefficient.

[0028] As a further improvement of an embodiment of the present invention, the method further includes that the using of a randomized algorithm to execute the reward distribution requested by the user and returning the distribution result includes:

[0029] Generate random numbers using a pseudo-random number generation algorithm ;

[0030] Based on the distribution probability of each reward item , calculate the sum of the winning probabilities of all reward items ;

[0031] By adding the random number Based on the probability of distribution of each reward item The probability distribution intervals are divided and compared to determine whether the user wins the prize and what kind of reward item he or she receives, specifically:

[0032] like , it is determined that the user has not received the reward;

[0033] like , it is determined that the user obtains the jth reward item;

[0034] The allocation results are stored in the distributed database through the application program interface.

[0035] As a further improvement of an embodiment of the present invention, the method further includes collecting user feedback data on the allocation results, and optimizing the deep reinforcement learning model through an online learning algorithm to update the reward allocation probability generation strategy;

[0036] The optimization of the deep reinforcement learning model by the online learning algorithm includes:

[0037] Collect user feedback data through the client interface; the user feedback data includes satisfaction scores, text evaluations and interactive behavior data;

[0038] Use natural language processing algorithms to perform sentiment analysis on the text reviews and generate instant rewards based on the satisfaction scores , as a reward return signal;

[0039] Use online learning algorithm based on comprehensive loss function Update the deep neural network parameters; the comprehensive loss function is expressed as:

[0040]

[0041] in, is the weight hyperparameter, is the value loss function, expressed as:

[0042]

[0043] in, For expectations, is the loss function, is the target value, is the value function estimate in state s, expressed as:

[0044]

[0045] Where i is the time step index, is the discount factor, is the immediate reward at time t+i, is the state at time t.

[0046] To achieve one of the above-mentioned objectives of the invention, an embodiment of the present invention further provides a personalized reward distribution system based on deep reinforcement learning, the system comprising a data acquisition module, a probability generation module and a reward distribution module;

[0047] The data collection module is used to respond to a user's request to participate in a reward distribution activity through a client, collect the user's feature data and reward pool information through a distributed data collection framework, and perform standardization processing to generate structured input data;

[0048] The probability generation module is used to process structured input data using a deep reinforcement learning model to generate personalized reward allocation probabilities and ensure that the allocation cost meets the preset budget constraints through a constrained optimization algorithm;

[0049] The reward allocation module is used to execute the reward allocation requested by the user using a randomized algorithm according to the personalized reward allocation probability and return the allocation result.

[0050] In order to achieve one of the above-mentioned purposes of the invention, an embodiment of the present invention also provides an electronic device, including a memory and a processor, characterized in that the memory stores a computer program that can be run on the processor, and when the program is executed on the processor, the steps in the personalized reward distribution method based on deep reinforcement learning as described above are implemented.

[0051] To achieve one of the above-mentioned objectives of the invention, an embodiment of the present invention further provides a storage medium, wherein the storage medium stores a computer program, and is characterized in that when the computer program is executed by a processor, the steps in the personalized reward allocation method based on deep reinforcement learning as described above are implemented.

[0052] Compared with the existing technology, the present invention provides a personalized reward distribution method and system based on deep reinforcement learning, which realizes personalized reward distribution through deep reinforcement learning, significantly improving user participation and satisfaction; distributed data processing and online learning mechanisms ensure real-time and dynamic adaptability; the constrained optimization algorithm effectively controls distribution costs and is suitable for large-scale applications; the closed-loop process integrates data collection, probability generation, distribution execution and model optimization, enhancing system efficiency and marketing effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is an overall flow chart of the personalized reward distribution method based on deep reinforcement learning described in the present invention.

[0054] Figure 2 Schematic diagram of the architecture of the personalized reward distribution system based on deep reinforcement learning described in the present invention. DETAILED DESCRIPTION

[0055] The present invention will be described in detail below with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional changes made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.

[0056] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not to be construed as limiting the present invention.

[0057] In the first embodiment of the present invention, the present invention provides a personalized reward distribution method based on deep reinforcement learning, such as Figure 1 Shown, including,

[0058] S1: In response to a user's request to participate in a reward distribution activity through the client, the user's feature data and reward pool information are collected through a distributed data collection framework, and standardized to generate structured input data;

[0059] S2: Uses a deep reinforcement learning model to process structured input data, generates personalized reward allocation probabilities, and uses a constrained optimization algorithm to ensure that the allocation cost meets the preset budget constraints;

[0060] S3: Based on the personalized reward allocation probability, a randomized algorithm is used to execute the reward allocation requested by the user and return the allocation result.

[0061] In a specific embodiment of the present invention, the user's feature data and reward pool information are collected through a distributed data collection framework, specifically,

[0062] The distributed data collection framework collects user feature data and reward pool information in real time through streaming data processing technology, where:

[0063] The user characteristic data includes demographic characteristics, device information, behavioral sequence data, consumption records and preference information;

[0064] The reward pool information includes the value of the reward items, the distribution popularity and the user's preference information for the reward items;

[0065] The data collection further includes data cleaning of the user feature data and the reward pool information, wherein the data cleaning removes invalid data through an anomaly detection algorithm.

[0066] It should be noted that the distributed data collection framework utilizes streaming data processing technology, enabling real-time collection of user profile data and reward pool information through high-throughput, low-latency message queues. Specifically, the framework is deployed on a distributed server cluster and comprises multiple data collection nodes. Each node interacts with clients (e.g., mobile apps, web pages) via an application programming interface (API) to capture user behavior and preference data in real time. This collection process occurs in response to a user's request to participate in a reward distribution activity through the client, such as by clicking a "Join Activity" button or accessing an activity page.

[0067] Furthermore, user characteristic data includes the following dimensions: Demographic characteristics: including user age, gender, geographic location, occupation and other information, obtained through user registration information or third-party authorization data. Device information: including device model, operating system version, screen resolution, etc., collected through the client SDK. Behavioral sequence data: including the number of user clicks, page dwell time, sharing behavior, etc., recording user interaction sequences through event tracking technology (such as embedding). Consumption records: including historical order amounts, purchase categories, browsing records, etc., extracted from the e-commerce platform database. Preference information: including user preferences for reward types (such as physical rewards, discount coupons) and time sensitivity, obtained through user ratings or selection behaviors on the client.

[0068] Furthermore, reward pool information includes the following: Reward item value: This refers to the monetary value or equivalent cost of each reward item, such as the face value of a cash voucher or the market price of a physical reward, extracted from the reward pool database. Distribution popularity: This refers to the popularity of a reward item, calculated by counting the frequency with which users click or select a reward item. User preference information for reward items: This includes user ratings or rankings of reward items, such as user priority preferences for rewards of equal value collected through the client interface.

[0069] Furthermore, data collection also includes data cleaning of the user feature data and the reward pool information, which is specifically implemented as follows: Anomaly detection algorithm: Statistical anomaly detection methods and machine learning algorithms are used to remove invalid data. For example, behavioral data with abnormally high click frequencies (exceeding a preset threshold) are eliminated, or invalid coordinate values ​​in geographic location data (such as those outside a reasonable range) are removed. Data deduplication and integrity check: The collected data is deduplicated through a distributed data processing framework, and data integrity is checked to ensure that key fields (such as user ID and reward project ID) are not empty. The cleaned data is transmitted to the data storage module via a distributed message queue, providing high-quality input for subsequent standardization processing and deep reinforcement learning models.

[0070] In a specific embodiment of the present invention, the normalization process is performed to generate structured input data, specifically,

[0071] The standardization process includes applying a normalization algorithm to convert the user feature data and the reward pool information after data cleaning into a unified dimension;

[0072] Using embedding technology to encode the user feature data converted to a unified dimension and the non-numerical data in the reward pool information into a numerical vector;

[0073] The numerical vectors described in the distributed data processing framework are batch processed and format converted to generate a structured feature matrix suitable for deep neural network input.

[0074] It should be noted that the normalization process aims to convert the cleaned user feature data and reward pool information into a unified format suitable for deep neural network input, ensuring data consistency and model computational efficiency. This process is executed on a distributed server cluster using a distributed data processing framework and consists of three sub-steps: normalization, embedding encoding, and batch processing, to generate a structured feature matrix.

[0075] Furthermore, a normalization algorithm is applied to convert the cleaned user feature data and reward pool information into a unified dimension to eliminate dimensional differences between different data dimensions. Preferably, a min-max or Z-score normalization algorithm is used to process numerical data (e.g., the number of clicks in a user behavior sequence or the value of a reward item) to ensure that the data values ​​are distributed within a uniform range, thereby eliminating the impact of dimensional differences on deep neural network training.

[0076] Furthermore, non-numeric data in the user feature data and reward pool information (such as gender and reward item categories in demographics) is encoded using embedding techniques and converted into numerical vectors. Specifically, a pre-trained embedding model or a custom embedding layer is used to map categorical variables into a high-dimensional vector space, generating a low-dimensional dense vector representation that preserves the semantic relationships of the data and adapts to the input requirements of the neural network.

[0077] Furthermore, a distributed data processing framework is used to batch process and convert the numerical vectors. Specifically, the distributed data processing framework aggregates the numerical vectors in batches through parallel computing, unifies the data format, and organizes the data according to the input requirements of the deep neural network to generate a structured feature matrix. The structured feature matrix contains a fusion representation of user features and reward pool information, arranged in a fixed dimension to ensure compatibility with the input layer of the deep reinforcement learning model.

[0078] In one embodiment of the present invention, a deep reinforcement learning model is used to process structured input data to generate personalized reward distribution probabilities, specifically,

[0079] Encoding the structured input data into a state vector using an embedding technique;

[0080] The state vector is processed by a multi-layer deep neural network to generate a non-normalized weight for each reward item j ;

[0081] The probability of allocating each reward item is calculated based on the non-normalized weight and reward value, and the formula is:

[0082]

[0083] in, is the probability of assigning the jth reward, is the target expected cost, is the price of the kth prize, k is the prize index, and n is the total number of prizes.

[0084] It should be noted that embedding technology is used to encode the structured input data into a state vector. Specifically, a pre-trained embedding model is used to map user feature data and reward pool information into a high-dimensional continuous vector space to capture the semantic relationships and feature correlations between the data. This embedding process is implemented through a distributed computing framework to ensure efficient processing of high-dimensional data.

[0085] Furthermore, the state vector is processed by a multi-layer deep neural network to generate a non-normalized weight for each reward item, wherein the deep neural network adopts a multi-layer perceptron or recurrent neural network architecture, including an input layer, multiple hidden layers and an output layer, processes nonlinear relationships through an activation function, and outputs a weight vector corresponding to the number of reward items.

[0086] Furthermore, according to the non-normalized weight and reward value, the formula Calculate the probability of each reward item being allocated. This probability calculation is performed by distributed computing nodes to ensure real-time and accuracy. The generated probability table is stored in the in-memory database for subsequent use by the reward allocation module.

[0087] In a specific embodiment of the present invention, a constraint optimization algorithm is used to ensure that the allocation cost meets the preset budget constraint, specifically,

[0088] Applying a constrained loss function To ensure that the total expected cost and the probability of a single item do not exceed the preset upper limit, the constraint loss function is expressed as:

[0089]

[0090] in, is the total expected cost of all reward items, The maximum allowable allocation probability for a single reward project, is the constraint strength coefficient.

[0091] It should be noted that the constraint loss function is applied to optimize the deep reinforcement learning model to ensure that the total expected cost is equal to the preset budget and the probability of single project allocation does not exceed the maximum allowable probability limit. The first term of the constraint loss function is The deviation of the total expected cost from the budget is penalized by the squared error. The second term A non-negative penalty is imposed on cases where the probability exceeds the upper limit by the rectified linear unit function.

[0092] Furthermore, the optimization process is implemented through a distributed computing framework, using a gradient descent algorithm to adjust the parameters of the deep neural network to minimize the constrained loss function. The optimized probability distribution is then passed to the reward allocation module via an in-memory database, ensuring precise control of the allocation cost and the validity of the probability.

[0093] In a specific embodiment of the present invention, a randomized algorithm is used to execute the reward distribution requested by the user and return the distribution result, specifically,

[0094] Generate random numbers using a pseudo-random number generation algorithm ;

[0095] Based on the distribution probability of each reward item , calculate the sum of the winning probabilities of all reward items ;

[0096] By adding the random number Based on the probability of distribution of each reward item The probability distribution intervals are divided and compared to determine whether the user wins the prize and what kind of reward item he or she receives, specifically:

[0097] like , it is determined that the user has not received the reward;

[0098] like , it is determined that the user obtains the jth reward item;

[0099] The allocation results are stored in the distributed database through the application program interface.

[0100] It should be noted that uniformly distributed random numbers are generated in a distributed computing environment through a high-performance pseudo-random number generator to ensure the fairness and randomness of the allocation process.

[0101] Furthermore, the sum of the probabilities of all reward items is calculated Calculations are completed by reading a pre-generated probability list from an in-memory database. The user's reward and the specific reward item are determined by comparing the random number with the cumulative probability interval. This comparison process is implemented using an efficient interval query algorithm to ensure real-time performance. The allocation results (including user ID, reward item ID, allocation timestamp, etc.) are stored in a distributed database via an application programming interface to support subsequent feedback data collection and system auditing. The allocation results are also returned to the user through the client interface.

[0102] In one embodiment of the present invention, user feedback data on the allocation results is collected, and a deep reinforcement learning model is optimized through an online learning algorithm to update the reward allocation probability generation strategy;

[0103] Specifically, the optimization of the deep reinforcement learning model by the online learning algorithm includes:

[0104] Collect user feedback data through the client interface; the user feedback data includes satisfaction scores, text evaluations and interactive behavior data;

[0105] Use natural language processing algorithms to perform sentiment analysis on the text reviews and generate instant rewards based on the satisfaction scores , as a reward return signal;

[0106] Use online learning algorithm based on comprehensive loss function Update the deep neural network parameters; the comprehensive loss function is expressed as:

[0107]

[0108] in, is the weight hyperparameter, is the value loss function, expressed as:

[0109]

[0110] in, For expectations, is the loss function, is the target value, is the value function estimate in state s, expressed as:

[0111]

[0112] Where i is the time step index, is the discount factor, is the immediate reward at time t+i, is the state at time t.

[0113] It should be noted that the client interface uses a real-time data transmission protocol to obtain user feedback data on reward distribution results from the client (such as a mobile application or web page). The user feedback data includes satisfaction scores (quantifying user satisfaction in numerical form), text evaluations (user text feedback on reward projects) and interactive behavior data (such as page dwell time and sharing behavior).

[0114] Furthermore, a natural language processing algorithm is used to perform sentiment analysis on the text reviews, extracting sentiment characteristics. This is then combined with the satisfaction scores using a weighted fusion algorithm to generate an immediate reward, which serves as the reward signal for the deep reinforcement learning model. An online learning algorithm is used to update the parameters of the deep neural network based on a comprehensive loss function, adjusting the balance between value loss and constraint loss.

[0115] Furthermore, online learning is implemented through a distributed deep learning framework, and the gradient descent algorithm is used to update the neural network parameters in real time. The optimized model parameters are stored in the model server for subsequent probability generation.

[0116] In the second embodiment of the present invention, the present invention provides a personalized reward distribution system based on deep reinforcement learning, such as Figure 2 As shown, the system includes a data acquisition module 1, a probability generation module 2 and a reward distribution module 3;

[0117] The data collection module 1 is used to respond to a user's request to participate in a reward distribution activity through a client, collect the user's feature data and reward pool information through a distributed data collection framework, and perform standardization processing to generate structured input data;

[0118] The probability generation module 2 is used to process structured input data using a deep reinforcement learning model to generate personalized reward allocation probabilities, and ensure that the allocation cost meets the preset budget constraints through a constrained optimization algorithm;

[0119] The reward allocation module 3 is used to execute the reward allocation requested by the user using a randomized algorithm according to the personalized reward allocation probability and return the allocation result.

[0120] In a third embodiment of the present invention, the present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the program is executed on the processor, the steps of the personalized reward allocation method based on deep reinforcement learning as described above are implemented.

[0121] In a fourth embodiment of the present invention, the present invention provides a storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the personalized reward allocation method based on deep reinforcement learning as described above.

[0122] In summary, the present invention provides a personalized reward distribution method and system based on deep reinforcement learning, which realizes personalized reward distribution through deep reinforcement learning, significantly improving user participation and satisfaction; distributed data processing and online learning mechanisms ensure real-time and dynamic adaptability; the constrained optimization algorithm effectively controls distribution costs and is suitable for large-scale applications; the closed-loop process integrates data collection, probability generation, distribution execution and model optimization, enhancing system efficiency and marketing effectiveness.

[0123] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working process of the modules described above can refer to the corresponding process in the aforementioned method implementation, and will not be repeated here.

[0124] Modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical modules, i.e., they may be located in one place or distributed across multiple network modules. Some or all of these modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0125] In addition, the functional modules in various embodiments of the present application may be integrated into a single processing module, or each module may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0126] The above-mentioned integrated modules implemented in the form of software functional modules can be stored in a computer-readable storage medium. The above-mentioned software functional modules are stored in a storage medium and include a number of instructions for causing a computer system (which may be a personal computer, server, or network system, etc.) or a processor to execute some of the steps of the methods described in various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A personalized reward distribution method based on deep reinforcement learning, characterized by: include, In response to a user's request to participate in a reward distribution activity through the client, the user's feature data and reward pool information are collected through a distributed data collection framework, and standardized to generate structured input data; Use deep reinforcement learning models to process structured input data, generate personalized reward allocation probabilities, and use constrained optimization algorithms to ensure that the allocation costs meet the preset budget constraints; Based on the personalized reward allocation probability, a randomized algorithm is used to execute the reward allocation requested by the user and return the allocation result.

2. The personalized reward distribution method based on deep reinforcement learning according to claim 1, characterized in that: The user's feature data and reward pool information collected through the distributed data collection framework include: The distributed data collection framework collects user feature data and reward pool information in real time through streaming data processing technology, where: The user characteristic data includes demographic characteristics, device information, behavioral sequence data, consumption records and preference information; The reward pool information includes the value of the reward items, the distribution popularity and the user's preference information for the reward items; The data collection further includes data cleaning of the user feature data and the reward pool information, wherein the data cleaning removes invalid data through an anomaly detection algorithm.

3. The personalized reward distribution method based on deep reinforcement learning according to claim 2, characterized in that: The standardization process to generate structured input data includes: The standardization process includes applying a normalization algorithm to convert the user feature data and the reward pool information after data cleaning into a unified dimension; Using embedding technology to encode the user feature data converted to a unified dimension and the non-numerical data in the reward pool information into a numerical vector; The numerical vectors described in the distributed data processing framework are batch processed and format converted to generate a structured feature matrix suitable for deep neural network input.

4. The personalized reward distribution method based on deep reinforcement learning according to claim 3, characterized in that: The method of using a deep reinforcement learning model to process structured input data and generate personalized reward distribution probabilities includes: Encoding the structured input data into a state vector using an embedding technique; The state vector is processed by a multi-layer deep neural network to generate a non-normalized weight for each reward item j ; The probability of allocating each reward item is calculated based on the non-normalized weight and reward value, and the formula is: ； in, is the probability of assigning the jth reward, is the target expected cost, is the price of the kth prize, k is the prize index, and n is the total number of prizes.

5. The personalized reward distribution method based on deep reinforcement learning according to claim 4, characterized in that: The method of ensuring that the allocation cost meets the preset budget constraint by using the constraint optimization algorithm includes: Applying a constrained loss function To ensure that the total expected cost and the probability of a single item do not exceed the preset upper limit, the constraint loss function is expressed as: ； in, is the total expected cost of all reward items, The maximum allowable allocation probability for a single reward project, is the constraint strength coefficient.

6. The personalized reward distribution method based on deep reinforcement learning according to claim 5, characterized in that: The method of using a randomized algorithm to execute the reward distribution requested by the user and returning the distribution result includes: Generate random numbers using a pseudo-random number generation algorithm ; Based on the distribution probability of each reward item , calculate the sum of the winning probabilities of all reward items ; By adding the random number Based on the probability of distribution of each reward item The probability distribution intervals are divided and compared to determine whether the user wins the prize and what kind of reward item he or she receives, specifically: like , it is determined that the user has not received the reward; like , it is determined that the user obtains the jth reward item; The allocation results are stored in the distributed database through the application program interface.

7. The personalized reward distribution method based on deep reinforcement learning according to claim 1, characterized in that: It also includes collecting user feedback data on the allocation results and optimizing the deep reinforcement learning model through online learning algorithms to update the reward allocation probability generation strategy; The optimization of the deep reinforcement learning model by the online learning algorithm includes: Collect user feedback data through the client interface; The user feedback data includes satisfaction scores, text evaluations and interactive behavior data; Use natural language processing algorithms to perform sentiment analysis on the text reviews and generate instant rewards based on the satisfaction scores , as a reward return signal; Use online learning algorithm based on comprehensive loss function Update the deep neural network parameters; the comprehensive loss function is expressed as: ； in, is the weight hyperparameter, is the value loss function, expressed as: ； in, For expectations, is the loss function, is the target value, is the value function estimate in state s, expressed as: ； Where i is the time step index, is the discount factor, is the immediate reward at time t+i, is the state at time t.

8. A personalized reward distribution system based on deep reinforcement learning, characterized by: Includes data acquisition module, probability generation module and reward distribution module; The data collection module is used to respond to a user's request to participate in a reward distribution activity through a client, collect the user's feature data and reward pool information through a distributed data collection framework, and perform standardization processing to generate structured input data; The probability generation module is used to process structured input data using a deep reinforcement learning model to generate personalized reward allocation probabilities and ensure that the allocation cost meets the preset budget constraints through a constrained optimization algorithm; The reward allocation module is used to execute the reward allocation requested by the user using a randomized algorithm according to the personalized reward allocation probability and return the allocation result.

9. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program that can be run on the processor, and when the program is executed on the processor, the steps in the personalized reward distribution method based on deep reinforcement learning as described in any one of claims 1 to 7 are implemented.

10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the personalized reward distribution method based on deep reinforcement learning are implemented.

Images