An integrated marketing management system

Abstract

The application relates to an integrated promotion marketing management system, in particular to the promotion management field, and the scheme generates a distributed user interest vector by fusing multi-party data under privacy protection through a user interest representation module, combines a promotion staff ability portrait vector and a commodity attraction vector refined by a business portrait construction module, and then calculates a privacy protection matching degree under the premise of data non-output by using a safe matching calculation module, so that a high-potential clue list is finally generated by intelligent fusion of a clue generation feedback module and a feedback closed loop is formed; data islands are effectively broken, the conversion efficiency and precision of promotion are significantly improved, and through continuous feedback data backflow, the self-optimization of the business portrait and the interest representation is driven, a virtuous cycle of getting smarter and smarter is formed, and the pain points of traditional promotion blindness and inefficiency are fundamentally solved.

Description

Technical Field

[0001] This invention relates to the field of promotion management, and more specifically, to an integrated promotion and marketing management system. Background Technology

[0002] With the deep integration of mobile internet and local life services, integrated promotion and marketing management systems have become the core hub connecting platforms with merchants, promoters, and consumers. Within the business ecosystem built by this system, individual and merchant promoters widely promote products and services through their unique promotion codes on social networks and in offline scenarios, forming a multi-level distribution network. Platforms typically accumulate a large amount of online user behavior data, such as browsing history, search keywords, and online order history, forming the basis for user interest profiles. Meanwhile, offline partner merchants joining the system independently hold anonymized transaction records of their in-store consumers, including valuable preference information such as product categories, frequency of purchases, and average order value. Ideally, if online behavior and offline transaction data can be deeply integrated, the system can accurately match each promoter with the products they are best at promoting and high-potential target consumer groups, thereby completely changing the traditional extensive promotion model that relies on personal connections and experience, and greatly improving the conversion efficiency and promotion network vitality of the "one-referral win-win" mechanism. However, in current business scenarios, the data of online platforms and offline merchants are naturally heterogeneous and isolated, forming solid data silos.

[0003] Existing technologies struggle to support the aforementioned demand for precise cross-domain matching, primarily due to the fundamental conflict between data utilization and privacy protection. Mainstream recommendation technologies, such as collaborative filtering algorithms, heavily rely on mining and analyzing centralized, rich user behavior data across all domains, which cannot be directly applied in decentralized environments with strict privacy constraints. While attempts exist for cross-channel matching based on device identifiers like phone numbers, these face compliance risks under increasingly stringent regulations and cannot address identification issues in logged-out scenarios, nor do they address the non-identifiable data held by offline merchants, which is transaction-centric. In recent years, privacy-preserving computation technologies like federated learning have offered new solutions to the data silo problem, allowing for joint modeling without data leaving the local environment. However, in the specific scenario of integrated marketing promotion, directly applying a general federated learning framework is not feasible. The system still faces significant challenges: First, the data scale and feature space of the participating parties, namely the platform and numerous merchants, vary greatly, and the merchant-side data may be extremely sparse, resulting in poor predictive performance of the jointly trained global model on the individual merchant side. Second, existing solutions typically aim to optimize a globally unified recommendation model, failing to address the highly personalized task of "finding potential consumers for specific promoters," and ignoring the unique "promotional capability profile" reflected in the promoter's own historical success experience. Therefore, there is currently a lack of feasible technical solutions that can effectively integrate heterogeneous and sparse online and offline data while strictly adhering to data privacy regulations, and ultimately generate personalized promotional leads for each unique promoter in the system. This leads to blind promotional activities, high conversion costs, and severely restricts the healthy development of the distribution network and the enthusiasm of promoters. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an integrated promotion and marketing management system. It solves the problems mentioned in the background art through a user interest characterization module, a business profile construction module, a secure matching calculation module, and a lead generation and feedback module.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically including: User interest representation module: Under the premise that the original data does not leave the local machine, the online platform and offline merchants are coordinated to conduct joint training. Using the user interaction behavior data of the online platform and the de-identified transaction data of the offline merchants, the distributed machine learning process that integrates knowledge transfer and parameter security aggregation is used to generate corresponding distributed user interest vectors. Business profile building module: Analyzes the promotion behavior records and transaction order data stored in the backend, performs data mining and feature extraction operations, and generates a promoter ability profile vector to represent the promoter's historical promotion ability tendency, and a product attraction vector to represent the product's historical audience and promotion conversion characteristics. Secure matching calculation module: When distributed user interest vectors are stored locally while promoter capability profile vectors and product attraction vectors are stored centrally on the platform, privacy-preserving calculation technology is used to calculate the privacy-preserving matching degree between the specified promoter capability profile vector, product attraction vector and each distributed user interest vector. Lead generation and feedback module: Based on the multi-source matching degree fusion rules that integrate direct matching, indirect product association matching, and collaborative filtering matching, it calculates a comprehensive matching score for a designated promoter and generates a list of high-potential leads containing group tags and product suggestions based on the comprehensive matching score. The feedback data after the promoter executes based on the high-potential lead list is sent back to update the corresponding promoter capability profile vector and product attraction vector in the business profile building module, and is used as training data to participate in the subsequent joint training initiated by the user interest representation module.

[0006] In a preferred embodiment, the coordinated joint training process in the user interest representation module, including knowledge transfer and secure parameter aggregation, includes: The online platform uses local user interaction data to train a teacher model and constructs a set of public anchor items that do not involve user privacy. The teacher model processes the set of common anchor items to generate a soft-label prediction distribution that contains general knowledge representations; Each offline merchant initializes a student model locally and receives the common anchor item set and soft tag predicted distribution; Each student model is trained on its local desensitized transaction data. Meanwhile, the output distribution it generates for the set of public anchor items needs to be aligned with the received soft-label prediction distribution through a total loss function that combines task loss and knowledge distillation loss, so as to transfer the general knowledge from the platform to the local model. Knowledge distillation loss is used to measure the consistency between the student model and the teacher model in their cognition of a set of common anchor items.

[0007] In a preferred embodiment, the parameter secure aggregation process specifically includes: In each round of joint training, after completing the local student model training, each offline merchant encrypts the parameters of the student model and uploads them to the online platform. The online platform calculates the encrypted global aggregate parameters by weighting the parameters of all received encrypted student models based on the scale of local anonymized transaction data from each offline merchant and a preset dynamically adjusted weighting factor. The online platform securely distributes the decrypted global aggregation parameters to each offline merchant to update their local student models and start the next round of training. After multiple iterations, offline merchants use the updated final student model to process the anonymized transaction data of local consumers and generate corresponding distributed user interest vectors.

[0008] In a preferred embodiment, the process of generating a promoter capability profile vector by performing data mining and feature extraction operations in the business profile construction module includes: By linking promotional activity records with transaction order data, all successful promotional conversion records are filtered out to form a valid promotional sample set; For each record in the effective promotion sample set, a comprehensive weight is calculated. This comprehensive weight is calculated based on the time decay factor determined by the time difference between the time of the promotion and the current time, the amount factor determined by the amount of the transaction brought by the promotion, and the preset efficiency coefficient of the channel used in the promotion. For a given promoter, all of its valid promotion records are extracted, and the distributed user interest vectors associated with the corresponding consumers in the records are weighted and averaged according to the comprehensive weight of each record. The resulting weighted average vector is the promoter capability profile vector of that promoter.

[0009] In a preferred embodiment, the process of generating a commodity gravity vector by performing data mining and feature extraction operations includes: For a given product, extract all records of successful promotions of that product from the valid promotion sample set; Based on the comprehensive weight of each record, the weighted average of two types of feature vectors is calculated: the first type is the weighted average of the distributed user interest vectors of the consumer group corresponding to the successful promotion of the product, which yields the historical audience feature vector. The second type is the weighted average of the promoter capability profile vectors of the promoters corresponding to the successful promotion of the product, which yields the feature vector of the successful promoter. The historical audience feature vector and the successful promoter feature vector are linearly transformed by a preset weight matrix and then added together. After being processed by a non-linear activation function, the final output vector is the product attraction vector of the product.

[0010] In a preferred embodiment, the process of calculating the privacy-preserving matching degree using privacy-preserving computing technology in the secure matching calculation module includes: The online platform generates a pair of keys for homomorphic encryption, including a public key and a private key, and securely distributes the public key to offline merchants that store distributed user interest vectors. When it is necessary to calculate the matching degree, the online platform uses a public key to homomorphically encrypt the specified promoter capability profile vector and the related product attraction vector, generating an encrypted promoter capability profile vector and an encrypted product attraction vector, and sends them as a secure query request to the specified offline merchant. Upon receiving a security query request, each offline merchant locally performs a secure similarity calculation function that supports both ciphertext and plaintext operations on each distributed user interest vector and the received encrypted promoter capability profile vector and encrypted product attraction vector, respectively. This calculates the corresponding promoter-user matching score and product-user matching score in the encrypted state. These encrypted matching scores are then appended with an anonymized user identifier and returned to the online platform.

[0011] In a preferred embodiment, after obtaining the encrypted matching score, the following further steps are performed: The online platform uses its private key to decrypt the encrypted matching scores returned from various offline merchants in batches, obtaining the plaintext set of promoter-user matching scores and the set of product-user matching scores; Subsequently, the online platform performs cross-batch standardization processing on the decrypted matching scores from different offline merchants to generate standardized privacy-protected matching scores with consistent comparability. Finally, the standardized privacy-preserving matching score is associated with the corresponding anonymized user identifier and the product identifier corresponding to the product gravity vector associated with the matching score, forming a structured list of matching score results.

[0012] In a preferred embodiment, the process of calculating the comprehensive matching score based on the multi-source matching degree fusion rule in the clue generation feedback module includes: Receive the standardized privacy-protected matching score for a specified promoter, and combine it with the promoter's capability profile vector and the product's attraction vector; the multi-source matching score fusion rule is as follows: First, the promoter-user matching score in the standardized privacy-protected matching score is used as an adjustment weight and weighted with the promoter's capability profile vector as a direct matching signal; Secondly, from the set of historical products that the promoter is good at promoting, which is associated with the promoter's ability profile vector, find the product with the highest product-user matching degree in the current user's standardized privacy protection matching degree, and extract the product's product gravity vector. The highest product-user matching degree is used as the adjustment weight and weighted with it as an indirect product association matching signal. Furthermore, a graph network is constructed based on the historical interaction relationship between promoters and users, and high-order implicit association features are extracted through graph collaborative filtering algorithm as collaborative filtering matching signals; Finally, the direct matching signal, indirect product association matching signal, and collaborative filtering matching signal are dynamically fused and nonlinearly transformed through a learnable attention mechanism network, and a scalar value is finally output as the comprehensive matching score between the promoter and the user.

[0013] In a preferred embodiment, the process of generating a list of high-potential leads based on a comprehensive matching score includes: The overall matching scores of all candidate users are statistically analyzed, and their mean and standard deviation are calculated. A dynamic threshold is set, which is the product of the mean, standard deviation and a preset coefficient. Users with a comprehensive matching score higher than the dynamic threshold are selected to form a high-potential user set. For this high-potential user set, the corresponding distributed user interest vector is analyzed by clustering algorithm, the important interest features of each cluster are extracted, and combined with the user's geographical location attributes, descriptive group labels are generated. Simultaneously, the high-frequency products associated with this high-potential user set during the generation of indirect product association matching signals are analyzed to form a product suggestion list; Encapsulate group tags, product suggestion lists, and corresponding anonymized user identifiers to generate a high-potential lead; integrate multiple such high-potential leads to form a high-potential lead list.

[0014] In a preferred embodiment, the process of sending back feedback data for updating includes: Collect user interaction and order conversion data generated after promoters execute high-potential lead lists, and associate them with the corresponding promoter ID, product ID, anonymized user ID, and lead ID used when generating the high-potential lead to form structured feedback data; Feedback data is transmitted and utilized in two ways: First, the successful conversion records in the feedback data, along with their associated promoter and product identifiers, are sent to the business profile building module in real time; The business profile building module uses this successful conversion record as a new effective promotion sample, recalculates and updates the corresponding promoter capability profile vector and product attraction vector with a higher weight, and realizes rapid iteration of the business profile. The second approach involves accumulating feedback data over a period of time, especially successful conversion orders and their associated anonymized transaction data, and sending them as new training samples to the user interest representation module. When initiating the next round of joint training, the user interest representation module incorporates these newly added samples into the training process to optimize the teacher and student models, thereby achieving a long-term, gradual improvement in user interest representation capabilities.

[0015] The beneficial effects of this invention are as follows: A distributed user interest vector is generated by fusing multi-party data under privacy protection through a user interest representation module. This vector is then combined with the promoter capability profile vector and product attraction vector extracted by the business profile construction module. A privacy-protected matching degree is calculated using a secure matching calculation module without data leakage. Finally, a high-potential lead list is generated through intelligent fusion via a lead generation and feedback module, forming a feedback loop. This effectively breaks down data silos, significantly improves the conversion efficiency and accuracy of promotions, and drives the self-optimization of business profiles and interest representations through continuous feedback data flow, forming a virtuous cycle of increasing intelligence with use. This fundamentally solves the pain points of blind and inefficient traditional promotion methods. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] Example 1 This embodiment provides, for example Figure 1-2 The integrated promotion and marketing management system shown here specifically includes: User interest representation module: Under the premise that the original data does not leave the local machine, the online platform and offline merchants are coordinated to conduct joint training. Using the user interaction behavior data of the online platform and the de-identified transaction data of the offline merchants, the distributed machine learning process that integrates knowledge transfer and parameter security aggregation is used to generate corresponding distributed user interest vectors. Business profile building module: Analyzes the promotion behavior records and transaction order data stored in the backend, performs data mining and feature extraction operations, and generates a promoter ability profile vector to represent the promoter's historical promotion ability tendency, and a product attraction vector to represent the product's historical audience and promotion conversion characteristics. Secure matching calculation module: When distributed user interest vectors are stored locally while promoter capability profile vectors and product attraction vectors are stored centrally on the platform, privacy-preserving calculation technology is used to calculate the privacy-preserving matching degree between the specified promoter capability profile vector, product attraction vector and each distributed user interest vector. Lead generation and feedback module: Based on the multi-source matching degree fusion rules that integrate direct matching, indirect product association matching, and collaborative filtering matching, it calculates a comprehensive matching score for a designated promoter and generates a list of high-potential leads containing group tags and product suggestions based on the comprehensive matching score. The feedback data after the promoter executes based on the high-potential lead list is sent back to update the corresponding promoter capability profile vector and product attraction vector in the business profile building module, and is used as training data to participate in the subsequent joint training initiated by the user interest representation module.

[0021] In this embodiment, it is specifically necessary to explain the coordinated joint training in the user interest representation module, whose knowledge transfer and parameter secure aggregation process includes: The online platform trains a teacher model using local user interaction data. The training process of the teacher model is as follows: through a deep neural network architecture, feature extraction and pattern learning are performed on the user's historical browsing sequence, search keyword set, and order records. Finally, a predictive model that can map user behavior to a general interest latent space is output. A set of public anchor items that does not involve user privacy is constructed. The method for constructing the set of public anchor items is: randomly sampling or uniformly extracting a certain number of product items from the platform's full product library according to the category distribution to form a list of public items that does not contain any user personal information. This list serves as the benchmark for knowledge alignment. The teacher model processes the set of common anchor items to generate a soft-label prediction distribution containing general knowledge representations. The specific processing steps are as follows: the features of each item in the set of common anchor items are input into the trained teacher model, and the model outputs the probability distribution of the user's possible interest in the item. The set of probability distributions of all items constitutes the soft-label prediction distribution, which reflects the general interest patterns contained in the platform-side data. Each offline merchant initializes a student model locally. The student model uses a neural network that is compatible with the teacher model but may have a simplified structure. Its parameters are randomly initialized or transferred from some parameters of the teacher model. It also receives a set of common anchor items and a soft-label predicted distribution. Each student model is trained on its local anonymized transaction data. The training process involves using historical consumer transaction records from the local anonymized transaction data as input features and the purchased goods or categories as prediction targets for supervised learning. Simultaneously, the output distribution generated by the student model for the common anchor item set needs to be aligned with the received soft-label prediction distribution through a total loss function that integrates task loss and knowledge distillation loss. The calculation process of the total loss function is as follows: first, the prediction error of the student model on its local anonymized transaction data is calculated as the task loss; second, the difference between the predicted probability distribution generated by the student model for the common anchor item set and the soft-label prediction distribution generated by the received teacher model is calculated as the knowledge distillation loss; finally, the total loss function value is the sum of the task loss value and the knowledge distillation loss value multiplied by a preset distillation intensity coefficient. This distillation intensity coefficient is used to control the strength of knowledge transfer, with a typical value range between 0.1 and 10, to transfer general knowledge from the platform to the local model. Knowledge distillation loss is used to measure the consistency between the student model and the teacher model in their cognition of a common anchor set of items. Specifically, the consistency measure is achieved by calculating the KL divergence between two probability distributions, that is, calculating the relative entropy between the student model's output distribution and the teacher model's soft label prediction distribution. The smaller this value, the more consistent their cognition is. The specific process of parameter safe aggregation includes: In each round of joint training, after completing the training of the student model locally, each offline merchant encrypts the parameters of the student model and uploads them to the online platform. The encryption process uses a homomorphic encryption algorithm to encrypt the model parameters, ensuring that the parameters exist in ciphertext form during transmission and aggregation, and that the original parameter values ​​cannot be parsed by the platform or other merchants. The online platform calculates a weighted average of all received encrypted student model parameters based on the scale of anonymized transaction data from each offline merchant and a preset dynamically adjusted weight factor, resulting in encrypted global aggregate parameters. The specific process of weighted average calculation is as follows: First, a base weight is determined for each offline merchant, proportional to the number of records in their anonymized transaction data. Second, a dynamically adjusted weight factor is introduced for each offline merchant. This factor is calculated based on the stability of their local model performance improvement over multiple training rounds or their deviation from the global model, used to amplify or reduce their base weight to balance their contribution to the global model. Finally, the aggregate weight of each offline merchant is... The weight is the product of the base weight and the dynamically adjusted weight factor. The encrypted model parameters of all offline merchants are summed according to this weight, and then divided by the sum of all weights to obtain the encrypted global aggregation parameters. The dynamically adjusted weight factor is used to adjust the balance between the general consensus and local characteristics in the aggregation process. The dynamic adjustment logic is as follows: if the local task loss of a local merchant's local student model has continued to decrease significantly in recent training rounds, its dynamically adjusted weight factor is increased to enhance the influence of its local characteristics in the global model; if its local task loss fluctuates greatly or differs too much from the global model parameters, its dynamically adjusted weight factor is decreased to make it more inclined to move towards the global consensus. The online platform securely distributes the decrypted global aggregation parameters to each offline merchant to update their local student model. The update process involves each offline merchant using the received decrypted global aggregation parameters to completely replace or mix their original local student model parameters in a certain proportion, and then starting the next round of training. After multiple iterations, offline merchants use the updated final student model to process the anonymized transaction data of local consumers and generate corresponding distributed user interest vectors. The generation process is as follows: input the features of the consumer's anonymized transaction data into the final student model, extract the output vector of a specific hidden layer in the middle of the model, or encode the output of the last layer of the model in a specific way. This vector is used as the distributed user interest vector of the consumer and is used for subsequent matching calculations.

[0022] In this embodiment, it is specifically necessary to explain that the process of generating the promoter capability profile vector by performing data mining and feature extraction operations in the business profile construction module includes: By linking promotional activity records with transaction order data, all successful promotional conversion records are filtered out to form a valid promotional sample set. The criteria for determining a successful promotional conversion record are: the order status in the transaction order data is "completed" or "paid", and the order can be accurately linked to the specific promotional action in the promotional activity record through the order number or user identifier, ensuring that each transaction can be traced back to the promoter who initiated the promotion and the promotional channel used. To effectively promote each record in the sample set, a comprehensive weight is calculated. This comprehensive weight is based on the time decay factor determined by the time difference between the transaction time of the promotion and the current time, the amount factor determined by the transaction amount generated by the promotion, and the preset efficiency coefficient of the channel used in the promotion. The specific calculation process of the comprehensive weight is as follows: First, the time decay factor is calculated using an exponential decay function. The time difference (in days) is multiplied by a negative decay coefficient, and the result is used as the exponent of the natural constant e. This decay coefficient is used to control the decay rate of historical data, and its typical value range is from 0.01 to... The first step is to calculate the amount factor, which is then processed using a function that adds one to the transaction amount and takes the natural logarithm to smooth out the impact of amount differences. Finally, the results of the time decay factor calculation, the amount factor calculation, and the preset channel efficiency coefficient are added together. The channel efficiency coefficient is preset according to the type of promotion channel; for example, the coefficient for private domain community sharing channels is set to 1.2, and the coefficient for public domain content platform sharing channels is set to 0.8. The sum is then multiplied by a channel adjustment coefficient, which is used to balance the influence of channels and is usually set between 0.5 and 1. The final value is the comprehensive weight of the record. For a given promoter, all of their valid promotion records are extracted. Based on the overall weight of each record, the distributed user interest vectors associated with the corresponding consumers in the record are weighted and averaged. The resulting weighted average vector is the promoter's promoter capability profile vector. The specific calculation process for the weighted average is as follows: the overall weight of each valid promotion record of the promoter is multiplied by the distributed user interest vector of the corresponding consumer in the record obtained from the user interest representation module to obtain a set of weighted vectors. Then, all weighted vectors are summed, and the sum is divided by the sum of all overall weights. The final vector represents the interest preference characteristics of the consumer group that the promoter has historically been most adept at reaching. The process of generating a product gravity vector by performing data mining and feature extraction operations includes: For a given product, extract all records of successful promotions of that product from the valid promotion sample set; that is, filter out all records in the valid promotion sample set whose product identifier matches the identifier of the specified product. Based on the comprehensive weight of each record, a weighted average of two types of feature vectors is calculated: the first type is the weighted average of the distributed user interest vectors of the consumer group corresponding to the successful promotion of the product, which yields the historical audience feature vector; the weighted average calculation process is similar to the weighted average process in generating the promoter capability profile vector, that is, multiplying the comprehensive weight of each record by the distributed user interest vector of the consumer in that record, summing them and dividing by the sum of the comprehensive weights, the resulting vector reflects the common interests of the consumer group attracted by the product; The second type is the weighted average of the promoter capability profile vectors of the promoters who successfully promoted the product, which yields the feature vector of the successful promoters. The weighted average calculation process is as follows: multiply the comprehensive weight of each record by the promoter capability profile vector corresponding to the promoter who performed the promotion in that record, sum them up and divide by the sum of the comprehensive weights. The resulting vector reflects the common capability characteristics of promoters who can successfully promote the product. The historical audience feature vector and the successful promoter feature vector are linearly transformed and added together using a preset weight matrix. Then, the result is processed by a non-linear activation function, and the final output vector is the product's attraction vector. The specific process of linear transformation and addition is as follows: the historical audience feature vector is multiplied by a preset first weight matrix to obtain the first transformed vector; the successful promoter feature vector is multiplied by a preset second weight matrix to obtain the second transformed vector; the first transformed vector, the second transformed vector, and a preset bias vector are added together. The non-linear activation function processing is as follows: the above addition result is input into a non-linear activation function, such as the hyperbolic tangent function, which performs a non-linear mapping on each element of the vector, restricting the output value to between negative one and positive one. The final vector is the product attraction vector, which integrates the product's audience attributes and promotional suitability attributes, and is used for subsequent matching calculations.

[0023] In this embodiment, it is specifically necessary to explain that the process of calculating the privacy-preserving matching degree using privacy-preserving computing technology in the security matching calculation module includes: The online platform generates a pair of keys for homomorphic encryption, including a public key and a private key, and securely distributes the public key to offline merchants that store distributed user interest vectors. The homomorphic encryption algorithm can be selected to satisfy the additive homomorphic property, such as the Paillier encryption algorithm, which enables direct addition and scalar multiplication operations in the ciphertext state, providing a foundation for subsequent secure computation. When it is necessary to calculate the matching degree, the online platform uses the public key to homomorphically encrypt the specified promoter capability profile vector and the related product attraction vector, generating the encrypted promoter capability profile vector and the encrypted product attraction vector, and sends them to the specified offline merchant as a secure query request; the encryption process is as follows: each numerical element of the vector is encrypted independently using the public key to generate a corresponding ciphertext element, and all ciphertext elements are arranged in the original order to form the encrypted vector. Upon receiving a secure query request, each offline merchant locally performs a secure similarity calculation function that supports ciphertext and plaintext operations on each distributed user interest vector, along with the received encrypted promoter capability profile vector and encrypted product attraction vector. This calculates the corresponding promoter-user matching score and product-user matching score in the encrypted state. The specific execution process of the secure similarity calculation function is as follows: First, it calculates the homomorphic dot product between the encrypted query vector and the local plaintext user vector. This operation is achieved by multiplying each ciphertext element of the encrypted vector with the corresponding element of the plaintext vector, and then homomorphically summing the ciphertext results of all products. Second, it calculates a secure normalization factor, which is obtained by homomorphically multiplying the ciphertext of the squared modulus of the encrypted query vector (pre-calculated by the platform and sent with the query) with the squared modulus of the local plaintext user vector. Finally, it performs a homomorphic division operation on the ciphertext result obtained from the homomorphic dot product, the ciphertext of the secure normalization factor (adding a very small constant, such as 10 to the power of -8), and the final output is the matching score in the encrypted state. These encrypted matching scores are then appended with an anonymized user identifier and returned to the online platform. The anonymization process uses a one-way hash function to irreversibly transform the original user identifier, ensuring that the platform cannot deduce the real identity. After obtaining the encrypted match score, the following steps are performed: The online platform uses its private key to decrypt the encrypted matching scores returned from various offline merchants in batches, obtaining the plaintext set of promoter-user matching scores and the set of product-user matching scores; the batch decryption process involves decrypting each encrypted score ciphertext independently using the private key to recover the original floating-point score value. Subsequently, the online platform performs cross-batch standardization on the decrypted matching scores from different offline merchants. This process includes calculating the global mean and standard deviation of all scores of the same type (e.g., all promoter-user matching scores), and then standardizing each score based on the global mean and standard deviation. The specific standardization process is as follows: for each score that needs to be standardized, subtract the global mean of that type of score, and then divide by the global standard deviation of that type of score, thereby transforming the score distribution into a standard normal distribution with a mean of zero and a standard deviation of one. This process effectively eliminates the impact of differences in the local data distribution of different merchants on the numerical scale of the matching scores, thus eliminating the score scale differences that may be caused by different data sources and generating standardized privacy-protected matching scores with consistent comparability. The standardized matching scores can be directly used for horizontal comparison and ranking across data sources. Finally, the standardized privacy-preserving matching score is associated with the corresponding anonymized user identifier and the product identifier corresponding to the product gravity vector associated with the matching score, forming a structured matching score result list. The structured result list is usually organized in the form of a database table or array. Each record contains a standardized matching score value, anonymized user identifier string, product identifier string, and matching type field, which can be directly queried and used by the subsequent lead generation module.

[0024] In this embodiment, it is specifically necessary to explain that the process of calculating the comprehensive matching score based on the multi-source matching degree fusion rule in the clue generation feedback module includes: The system receives standardized privacy-preserving match scores for a specified promoter and combines them with the promoter's capability profile vector and the product's attraction vector. The multi-source match score fusion rule is as follows: First, the promoter-user match score in the standardized privacy-preserving match score is used as an adjustment weight and weighted with the promoter's capability profile vector to serve as a direct match signal. This weighting operation involves multiplying the promoter-user match score with the value of each dimension of the promoter's capability profile vector to obtain a new vector that has been scaled by the match score. This vector reflects the interest tendencies of the promoter directly related to the current user. Secondly, from the set of historically promoted products associated with the promoter's capability profile vector, identify the product with the highest product-user matching degree in the standardized privacy protection matching degree of the current user, and extract the product gravity vector of the product. The highest product-user matching degree is used as an adjustment weight to weight it, as an indirect product association matching signal. This step involves traversing all the products successfully promoted by the promoter in the past, finding the product record with the highest matching degree of the current user, and weighting the product gravity vector of the product with the matching degree score to obtain a feature vector representing "products that the user may like and that the promoter is good at promoting". Secondly, a graph network is constructed based on the historical interaction relationship between promoters and users. High-order implicit association features are extracted through the graph collaborative filtering algorithm as collaborative filtering matching signals. The graph network is constructed with promoters and users as nodes and the number of historical successful promotions as edge weights. The graph collaborative filtering algorithm adopts the message passing mechanism of graph neural network to perform multi-round feature aggregation on the nodes in the graph, and finally extracts a dense feature vector that reflects the high-order similarity relationship between promoters and users. Finally, the direct matching signal, indirect product association matching signal, and collaborative filtering matching signal are dynamically fused and nonlinearly transformed through a learnable attention mechanism network. The final output is a scalar value as the comprehensive matching score between the promoter and the user. The specific structure of the attention mechanism network is a multilayer perceptron. Its input is a fused vector formed by concatenating the direct matching signal vector, the indirect product association matching signal vector, and the collaborative filtering matching signal vector. The multilayer perceptron first performs a linear transformation on the concatenated vector through a fully connected layer, then passes it through a nonlinear activation function, and finally through another fully connected layer to map the high-dimensional features into a scalar value. During training, the network learns how to assign appropriate attention weights to the three different matching signals, thereby achieving adaptive optimal fusion. The process of generating a list of high-potential leads based on the overall matching score includes: The overall matching scores of all candidate users are statistically analyzed, and their mean and standard deviation are calculated. That is, the sum of the overall matching scores between all promoters and each user is divided by the total number of users to obtain the mean. Then, the sum of the squares of the differences between each score and the mean is calculated, divided by the total number of users, and the square root is taken to obtain the standard deviation. A dynamic threshold is set, which is the product of the average score plus the standard deviation and a preset coefficient. Users with a comprehensive matching score higher than the dynamic threshold are selected to form a high-potential user set. The preset coefficient is used to control the strictness of the selection. Its value range is usually between 0.5 and 1.5. For example, setting it to 0.8 means selecting users whose scores are 0.8 standard deviations higher than the average score, thus achieving adaptive and non-fixed number of selections. For this set of high-potential users, clustering algorithms are used to analyze their corresponding distributed user interest vectors, extracting key interest features for each cluster and combining them with users' geographic location attributes to generate descriptive group labels. The clustering algorithm employs K-means clustering or density-based clustering to divide high-potential users into multiple groups based on their interest vectors. For each group, the most representative interest labels are identified by calculating the mean or mode of its members' interest vectors across various dimensions, such as "digital enthusiast" or "foodie." Simultaneously, frequently occurring geographic location codes are parsed from the anonymized user identifiers and converted into regional descriptions such as "Business District A" or "High-tech Zone." These regional descriptions are then combined with core interest labels to form group labels such as "digital enthusiasts in Business District A." Simultaneously, the high-frequency products associated with this high-potential user set during the generation of indirect product association matching signals are analyzed to form a product suggestion list; that is, the highest matching product corresponding to each high-potential user is counted when generating indirect product association matching signals, and then these products are sorted in descending order of frequency of occurrence, and the top five to ten products are selected to form a product suggestion list recommended to promoters. Encapsulate the group tags, product suggestion list, and corresponding anonymized user identifiers to generate a high-potential lead; integrate multiple such high-potential leads to form a high-potential lead list. Each lead contains a group tag string, a product identifier list, an associated anonymous user identifier list, and a unique identifier for the lead in its data structure. All leads are sorted by generation time or priority to form a list for promoters to display on the front end. The process of sending back feedback data for updating includes: We collect user interaction and order conversion data generated after promoters execute high-potential lead lists, and associate this data with the corresponding promoter ID, product ID, anonymized user ID, and the lead ID used to generate the high-potential lead to form structured feedback data. The user interaction data includes exposure to promotional content, click behavior, and occurrence time, while the order conversion data includes order ID, transaction amount, and completion time. All data is associated with the original group tags and product suggestions through the lead ID. Feedback data is transmitted and utilized in two ways: First, successful conversion records in the feedback data, along with their associated promoter and product identifiers, are sent to the business profile building module in real time. The business profile building module uses this successful conversion record as a new valid promotion sample, recalculates and updates the corresponding promoter capability profile vector and product attraction vector with a higher weight, and realizes rapid iteration of the business profile. The higher weight is achieved by assigning an initial weight to the new record based on its transaction time (latest) and transaction amount. This weight is significantly higher than the average weight of historical samples, so that the new sample has a greater influence in the weighted average calculation, thereby quickly correcting the profile vector. The second approach involves accumulating feedback data over a period of time, especially successful conversion orders and their associated anonymized transaction data, as new training samples and sending them to the user interest representation module. The accumulation period is usually set to one day or one week, and all new successful conversion order data and their corresponding anonymized transaction records from offline merchants (such as purchase category and amount) within the period are packaged into a batch of new training sample sets. When initiating the next round of joint training, the user interest representation module incorporates these newly added samples into the training process to optimize the teacher and student models, thereby achieving a long-term, gradual improvement in user interest representation capabilities. Specifically, the online platform adds user interaction behavior data from the newly added samples to its local training set for fine-tuning the teacher model; offline merchants add anonymized transaction data from their local consumers from the newly added samples to their local training set for training the student model in the next round of joint training. In this way, the positive feedback of the promotional effect is continuously encoded into the system's underlying user interest representation, forming a complete reinforcement loop from decision-making to representation.

[0025] In a specific deployment of this system, the following business rules can be combined to work together to further enhance the vitality of the promotion network and user stickiness, wherein; The individual promotion mechanism includes: after an individual promoter successfully invites a new merchant to join the platform through their exclusive promotion code, they can immediately receive a certain percentage of the new merchant's sales revenue as an immediate reward; thereafter, as long as the merchant continues to generate transactions on the platform, the individual promoter can receive a corresponding percentage of the commission from each transaction, achieving "one-time promotion, continuous income"; the individual promoter's promotion targets are limited to merchants, and their income comes entirely from the sales revenue of the merchants they directly recruit, without involving the performance sharing of their subordinate promoters; when operating alone, the individual promotion mechanism can stimulate the promotion enthusiasm of a large number of individual users, forming a wide-ranging ground promotion network and rapidly expanding the platform's merchant base; however, since individual promoters lack team leverage, their income growth depends on continuous investment in their own promotion capabilities; The merchant promotion mechanism includes: Merchant promoters (i.e., merchants already registered on the platform) enjoy the same direct customer acquisition benefits as individual promoters, and also have the privilege of developing subordinate individual promoters. Specifically: Merchant promoters can directly invite new merchants to join and enjoy a certain percentage of the new merchants' sales revenue as commission. Simultaneously, merchant promoters can also develop individual promoters. When a new merchant is successfully recruited by a recruited individual promoter, both the merchant promoter and the recruited individual promoter share a certain percentage of the new merchant's sales revenue as commission. However, it should be noted that this commission relationship is limited to one level of commission; that is, when a merchant promoter's subordinate individual promoter develops new promoters, the merchant promoter will no longer receive commission. When operating independently, the merchant promotion mechanism can fully utilize merchants' existing business networks and industry resources to achieve high-quality mutual recommendations between merchants. Furthermore, by developing individual promoters, merchants can build their own promotion teams, creating a leverage effect and accelerating revenue growth. The differences between individual promotion and merchant promotion include: individual promoters can only earn income through their own direct user acquisition, with a simple income structure but low barriers to entry and flexible participation; merchant promoters, on the other hand, have both direct user acquisition income and team management income, and can multiply their income by developing individual promoters, but require higher operational capabilities and resource investment from the merchant; the individual promotion mechanism focuses on broad coverage and is suitable for mobilizing individual users on a large scale; the merchant promotion mechanism focuses on deep penetration and is suitable for building a stable promotional hierarchy network. The synergistic effect of combining individual and merchant promotion includes: when the individual and merchant promotion mechanisms operate in tandem, a three-dimensional promotion network of "individual outreach, in-depth merchant outreach, and tiered incentives" is formed; individual promoters are responsible for broadly reaching potential merchants, rapidly expanding the platform's merchant base; merchant promoters, leveraging their industry experience and networks, precisely invite high-value merchants and extend the promotion network downwards by developing individual promoters; after the two join forces, individual promoters provide merchant promoters with potential promotion team members, while merchant promoters provide individual promoters with a more stable source of income and growth path; this synergy makes the promotion network both broad and deep, significantly improving overall promotion efficiency and conversion rates; The upgraded points-based lottery mechanism includes: a built-in multi-stage upgraded points-based lottery rule; users earn points for every certain amount of spending, and when points accumulate to a preset threshold, they can participate in the corresponding level of lottery; the specific rules are as follows: when the cumulative number of participants and the average spending per person both reach the preset conditions of the first stage, participants can use the number of points specified for that stage to participate in one lottery, and any remaining points will automatically be transferred to the next stage; subsequent stages follow the same pattern, each stage has independent thresholds for the number of participants, spending, and points consumed, and the higher the stage, the higher the threshold and the points consumed; when all stages of the lottery are completed, the level is automatically reset to the first stage, and any remaining points are transferred to the first stage to continue the cycle; each stage of the lottery uses a random number matching mechanism to ensure fairness; this lottery rule is linked to group tags and product suggestions in the clue generation feedback module, which can dynamically adjust the prize pool for different user groups to improve user participation; The overall synergistic effect of promotion and lotteries includes: individual and merchant promotion mechanisms bring a continuous increase in merchant supply and user traffic to the platform, while the upgraded lottery mechanism effectively promotes user consumption and points consumption, forming a positive cycle: promoters earn commission income through promotion, which incentivizes them to actively expand new merchants and new users; the arrival of new merchants brings more consumption scenarios, users generate points through consumption and then participate in lotteries, and the lottery activities, in turn, stimulate user consumption, thereby increasing merchant turnover and indirectly increasing promoters' commission income; at the same time, the system, through user interest representation modules and secure matching calculation modules, can accurately identify high-potential promoters and high-value users, and push lottery activities to the most likely responding groups, realizing intelligent synergy between promotion and lotteries, and ultimately achieving a win-win ecological closed loop for the platform, merchants, promoters, and users.

[0026] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0027] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0028] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0029] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0030] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0031] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make further changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. An integrated promotion and marketing management system, characterized in that, Specifically, it includes: User interest representation module: Under the premise that the original data does not leave the local machine, the online platform and offline merchants are coordinated to conduct joint training. Using the user interaction behavior data of the online platform and the de-identified transaction data of the offline merchants, the distributed machine learning process that integrates knowledge transfer and parameter security aggregation is used to generate corresponding distributed user interest vectors. Business profile building module: Analyzes the promotion behavior records and transaction order data stored in the backend, performs data mining and feature extraction operations, and generates a promoter ability profile vector to represent the promoter's historical promotion ability tendency, and a product attraction vector to represent the product's historical audience and promotion conversion characteristics. Secure matching calculation module: When distributed user interest vectors are stored locally while promoter capability profile vectors and product attraction vectors are stored centrally on the platform, privacy-preserving calculation technology is used to calculate the privacy-preserving matching degree between the specified promoter capability profile vector, product attraction vector and each distributed user interest vector. Lead generation and feedback module: Based on the multi-source matching degree fusion rules that integrate direct matching, indirect product association matching, and collaborative filtering matching, it calculates a comprehensive matching score for a designated promoter and generates a list of high-potential leads containing group tags and product suggestions based on the comprehensive matching score. The feedback data after the promoter executes based on the high-potential lead list is sent back to update the corresponding promoter capability profile vector and product attraction vector in the business profile building module, and is used as training data to participate in the subsequent joint training initiated by the user interest representation module.

2. The integrated promotion and marketing management system according to claim 1, characterized in that: In the user interest representation module, the coordinated joint training process, including knowledge transfer and secure parameter aggregation, includes: The online platform uses local user interaction data to train a teacher model and constructs a set of public anchor items that do not involve user privacy. The teacher model processes the set of common anchor items to generate a soft-label prediction distribution that contains general knowledge representations; Each offline merchant initializes a student model locally and receives the common anchor item set and soft tag predicted distribution; Each student model is trained on its local desensitized transaction data. Meanwhile, the output distribution it generates for the set of public anchor items needs to be aligned with the received soft-label prediction distribution through a total loss function that combines task loss and knowledge distillation loss, so as to transfer the general knowledge from the platform to the local model. Knowledge distillation loss is used to measure the consistency between the student model and the teacher model in their cognition of a set of common anchor items.

3. The integrated promotion and marketing management system according to claim 2, characterized in that: The specific process of parameter safe aggregation includes: In each round of joint training, after completing the local student model training, each offline merchant encrypts the parameters of the student model and uploads them to the online platform. The online platform calculates the encrypted global aggregate parameters by weighting the parameters of all received encrypted student models based on the scale of local anonymized transaction data from each offline merchant and a preset dynamically adjusted weighting factor. The online platform securely distributes the decrypted global aggregation parameters to each offline merchant to update their local student models and start the next round of training. After multiple iterations, offline merchants use the updated final student model to process the anonymized transaction data of local consumers and generate corresponding distributed user interest vectors.

4. The integrated promotion and marketing management system according to claim 3, characterized in that: In the business profile building module, the process of generating a promoter capability profile vector by performing data mining and feature extraction operations includes: By linking promotional activity records with transaction order data, all successful promotional conversion records are filtered out to form a valid promotional sample set; For each record in the effective promotion sample set, a comprehensive weight is calculated. This comprehensive weight is calculated based on the time decay factor determined by the time difference between the time of the promotion and the current time, the amount factor determined by the amount of the transaction brought by the promotion, and the preset efficiency coefficient of the channel used in the promotion. For a given promoter, all of its valid promotion records are extracted, and the distributed user interest vectors associated with the corresponding consumers in the records are weighted and averaged according to the comprehensive weight of each record. The resulting weighted average vector is the promoter capability profile vector of that promoter.

5. The integrated promotion and marketing management system according to claim 4, characterized in that: The process of generating a product gravity vector by performing data mining and feature extraction operations includes: For a given product, extract all records of successful promotions of that product from the valid promotion sample set; Based on the comprehensive weight of each record, the weighted average of two types of feature vectors is calculated: the first type is the weighted average of the distributed user interest vectors of the consumer group corresponding to the successful promotion of the product, which yields the historical audience feature vector. The second type is the weighted average of the promoter capability profile vectors of the promoters corresponding to the successful promotion of the product, which yields the feature vector of the successful promoter. The historical audience feature vector and the successful promoter feature vector are linearly transformed by a preset weight matrix and then added together. After being processed by a non-linear activation function, the final output vector is the product attraction vector of the product.

6. The integrated promotion and marketing management system according to claim 5, characterized in that: In the secure matching calculation module, the process of calculating the privacy-preserving matching degree using privacy-preserving computing technology includes: The online platform generates a pair of keys for homomorphic encryption, including a public key and a private key, and securely distributes the public key to offline merchants that store distributed user interest vectors. When it is necessary to calculate the matching degree, the online platform uses a public key to homomorphically encrypt the specified promoter capability profile vector and the related product attraction vector, generating an encrypted promoter capability profile vector and an encrypted product attraction vector, and sends them as a secure query request to the specified offline merchant. Upon receiving a security query request, each offline merchant locally performs a secure similarity calculation function that supports both ciphertext and plaintext operations on each distributed user interest vector and the received encrypted promoter capability profile vector and encrypted product attraction vector, respectively. This calculates the corresponding promoter-user matching score and product-user matching score in the encrypted state. These encrypted matching scores are then appended with an anonymized user identifier and returned to the online platform.

7. The integrated promotion and marketing management system according to claim 6, characterized in that: After obtaining the encrypted match score, the following steps are performed: The online platform uses its private key to decrypt the encrypted matching scores returned from various offline merchants in batches, obtaining the plaintext set of promoter-user matching scores and the set of product-user matching scores; Subsequently, the online platform performs cross-batch standardization processing on the decrypted matching scores from different offline merchants to generate standardized privacy-protected matching scores with consistent comparability. Finally, the standardized privacy-preserving matching score is associated with the corresponding anonymized user identifier and the product identifier corresponding to the product gravity vector associated with the matching score, forming a structured list of matching score results.

8. The integrated promotion and marketing management system according to claim 7, characterized in that: In the clue generation and feedback module, the process of calculating the comprehensive matching score based on the multi-source matching degree fusion rule includes: Receive the standardized privacy-protected matching score for a specified promoter, and combine it with the promoter's capability profile vector and the product's attraction vector; the multi-source matching score fusion rule is as follows: First, the promoter-user matching score in the standardized privacy-protected matching score is used as an adjustment weight and weighted with the promoter's capability profile vector as a direct matching signal; Secondly, from the set of historical products that the promoter is good at promoting, which is associated with the promoter's ability profile vector, find the product with the highest product-user matching degree in the current user's standardized privacy protection matching degree, and extract the product's product gravity vector. The highest product-user matching degree is used as the adjustment weight and weighted with it as an indirect product association matching signal. Furthermore, a graph network is constructed based on the historical interaction relationship between promoters and users, and high-order implicit association features are extracted through graph collaborative filtering algorithm as collaborative filtering matching signals; Finally, the direct matching signal, indirect product association matching signal, and collaborative filtering matching signal are dynamically fused and nonlinearly transformed through a learnable attention mechanism network, and a scalar value is finally output as the comprehensive matching score between the promoter and the user.

9. The integrated promotion and marketing management system according to claim 8, characterized in that: The process of generating a list of high-potential leads based on the overall matching score includes: The overall matching scores of all candidate users are statistically analyzed, and their mean and standard deviation are calculated. A dynamic threshold is set, which is the product of the mean, standard deviation and a preset coefficient. Users with a comprehensive matching score higher than the dynamic threshold are selected to form a high-potential user set. For this high-potential user set, the corresponding distributed user interest vector is analyzed by clustering algorithm, the important interest features of each cluster are extracted, and combined with the user's geographical location attributes, descriptive group labels are generated. Simultaneously, the high-frequency products associated with this high-potential user set during the generation of indirect product association matching signals are analyzed to form a product suggestion list; Encapsulate group tags, product suggestion lists, and corresponding anonymized user identifiers to generate a high-potential lead; integrate multiple such high-potential leads to form a high-potential lead list.

10. The integrated promotion and marketing management system according to claim 9, characterized in that: The process of sending back feedback data for updating includes: Collect user interaction and order conversion data generated after promoters execute high-potential lead lists, and associate them with the corresponding promoter ID, product ID, anonymized user ID, and lead ID used when generating the high-potential lead to form structured feedback data; Feedback data is transmitted and utilized in two ways: First, the successful conversion records in the feedback data, along with their associated promoter and product identifiers, are sent to the business profile building module in real time; The business profile building module uses this successful conversion record as a new effective promotion sample, recalculates and updates the corresponding promoter capability profile vector and product attraction vector with a higher weight, and realizes rapid iteration of the business profile. The second approach involves accumulating feedback data over a period of time, especially successful conversion orders and their associated anonymized transaction data, and sending them as new training samples to the user interest representation module. When initiating the next round of joint training, the user interest representation module incorporates these newly added samples into the training process to optimize the teacher and student models, thereby achieving a long-term, gradual improvement in user interest representation capabilities.

Images