Multi-scene marketing business risk digital monitoring system and method

Abstract

The invention discloses a multi-scene marketing business risk digital monitoring system and a multi-scene marketing business risk digital monitoring method, belongs to the technical field of marketing management, and solves the problems that an existing system depends on data distribution condition screening of time intervals, pays attention to analysis of time dimensions and lacks comprehensive analysis of multi-dimensional data when screening abnormal business objects. The method comprises the following steps: acquiring real-time marketing data of multiple scenes, preprocessing the real-time marketing data, setting a marketing dataset as a local dataset by a parallel chain node, pre-identifying the marketing dataset based on a pre-trained risk identification sub-model, taking a scene risk set as input by the parallel chain node, performing anomaly detection on the scene risk set, and identifying the scene risk set according to the anomaly detection result. Predicting marketing risk factors, and calculating a comprehensive risk value; according to the invention, through real-time acquisition of multi-scene data and multi-dimensional data preprocessing, it can be ensured that a space-time risk judgment model can identify and predict potential risks in a multi-scene marketing scene more comprehensively and accurately, and the comprehensiveness and accuracy of risk identification are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of marketing management, and particularly relates to a digital monitoring system and method for multi-scenario marketing business risks. Background Art

[0002] With the rapid development of digital technologies, the scenarios of enterprise marketing business have become increasingly complex, covering various channels such as online e-commerce platforms, offline stores, social media promotion, and mobile applications. In these multi-scenario marketing activities, enterprises face many risks, such as fraud risks, credit risks, market risks, operational risks, etc. Traditional manual monitoring and simple data analysis methods have been difficult to meet the requirements of monitoring the risks of complex and changeable marketing business.

[0003] Chinese Patent CN117611211A discloses an intelligent marketing management system, which monitors and analyzes conventional scenarios in the intelligent marketing system through an intelligent analysis terminal, obtains the business lines of any corresponding business objects, and performs data parsing with the help of a parsing model. The parsing model mainly depends on the time points of each performance point in the business line, and according to the time intervals between adjacent performance points, screens the time intervals in the business line according to the data distribution of the time intervals to obtain the activity duration range; then determines all deviant business objects according to the span duration and the activity duration range; however, when the existing system's intelligent analysis terminal screens abnormal business objects, it depends on the data distribution of time intervals for screening, mainly focuses on the analysis of the time dimension, lacks the comprehensive analysis of multi-dimensional data (such as space, user behavior, marketing activity characteristics, etc.), the identification of deviant business objects is relatively simple, and it is impossible to perform risk identification and prediction for multi-dimensions and multi-scenarios. In view of the above problems, we propose a digital monitoring system and method for multi-scenario marketing business risks. Summary of the Invention

[0004] The purpose of the present invention is to provide a digital monitoring system and method for multi-scenario marketing business risks in view of the deficiencies of the prior art, and solve the problem that the existing system depends on the data distribution of time intervals for screening when screening abnormal business objects, mainly focuses on the analysis of the time dimension, and lacks the comprehensive analysis of multi-dimensional data.

[0005] The present invention is implemented as follows. The digital monitoring method for multi-scenario marketing business risks includes: Obtain real-time marketing data of multiple scenarios, preprocess the real-time marketing data to obtain a marketing data set, and send it to the parallel chain node allocated by the blockchain platform; The parallel chain node sets the marketing data set as a local data set, trains a risk identification sub-model based on the fuzzy clustering algorithm using the local data set, and pre-identifies the marketing data set based on the pre-trained risk identification sub-model to pre-identify scenario risk data and generate a scenario risk set; The blockchain platform captures historical marketing data and trains a spatio-temporal risk judgment model that combines the spatio-temporal federated learning algorithm and the risk identification sub-model based on the historical marketing data; The parallel chain node takes the scenario risk set as the input, performs anomaly detection on the scenario risk set, predicts the marketing risk factors, calculates the comprehensive risk value in combination with the marketing risk factors, and uploads the comprehensive risk value to the blockchain platform, and digitally visualizes the comprehensive risk value corresponding to the scenario.

[0006] Preferably, the method for preprocessing the real-time marketing data includes: Loading the real-time marketing data and performing missing value and outlier cleaning on the real-time marketing data; Based on the source of the real-time marketing data, performing business scenario type marking on the data to obtain the real-time marketing data containing scenario type labels; Using the principal component analysis method to weight the marketing tasks corresponding to the business scenarios to obtain the weighted real-time marketing data. When multiple groups of marketing tasks correspond to the business scenario, linear combination weighting is performed on the weights of the multiple groups of marketing tasks; Loading the marked and weighted real-time marketing data, performing spectral analysis on the real-time marketing data to determine the main frequency distribution of the real-time marketing data, identifying the position and bandwidth of the noise signal in the frequency domain, and calculating the noise interference coefficient based on the adaptive threshold algorithm; Judging whether the noise interference coefficient in the real-time marketing data exceeds the preset noise frequency domain value, where the noise frequency domain value is determined by the Stein unbiased risk estimation criterion; If it exceeds the preset noise frequency domain value, perform wavelet denoising on the real-time marketing data based on the db6 wavelet denoising function; If it does not exceed the preset noise frequency domain value, perform wavelet denoising on the real-time marketing data based on the db1 wavelet denoising function; Integrate the real-time marketing data after wavelet denoising processing, extract the key feature points of the real-time marketing data based on the AMPD algorithm combined with variational mode decomposition, and integrate at least one group of key feature points to form a marketing data set.

[0007] Preferably, the noise interference coefficient is calculated by the following formula: (1) (2) (3) Where represents the noise interference coefficient, is the number of noise signals, is the signal length, represents the average bandwidth of the noise signal, are the maximum and minimum values of the power spectral density of the noise signal within the signal length, respectively, represents the current noise signal of the signal bandwidth, represents the current noise signal of the power spectral density; The noise frequency domain value is determined by the following formula: (4) where, represents the noise frequency domain value, represents the initial value of the noise frequency domain value, represents the sign function, which is used to judge the positive and negative attributes of the input signal of, represents the noise level in the Stein unbiased risk estimation criterion.

[0008] Preferably, the method for training the risk identification sub-model based on the fuzzy clustering algorithm using the local data set includes: Taking the convolutional neural network model as the initial model of the risk identification sub-model, and pre-constructing the risk identification sub-model; Loading the pre-constructed risk identification sub-model, and setting the sub-model iteration times, maximum iteration times, learning rate, and hyperparameters; Obtaining the local data set, extracting the training data in the local data set based on the K-fold cross-validation method, and pre-training the risk identification sub-model with the training data to obtain the initial learning rate and initial hyperparameters of the risk identification sub-model; Loading the local data set, dividing the local data set into a training subset and a validation subset, iteratively training the risk identification sub-model based on transfer learning until convergence, and outputting the converged risk identification sub-model; Among them, when training the risk identification sub-model, the learning rate is fine-tuned using cosine decay, and the learning rate at the th iteration is (5) where, represents the initial learning rate of the risk identification sub-model, is the maximum iteration times of the sub-model; Obtaining the validation subset, validating the performance of the risk identification sub-model using the validation subset, outputting the F1 score of the evaluation index of the risk identification sub-model, and judging whether the F1 score of the evaluation index of the risk identification sub-model exceeds the preset score threshold. If it exceeds the preset score threshold, output the converged risk identification sub-model.

[0009] Preferably, the risk identification sub-model consists of an input layer, a convolutional layer, a pooling layer, and an output layer. The convolutional layer consists of three convolutions, including a first convolution, a second convolution, and a third convolution. Residual connection layers are introduced after the first convolution, the second convolution, and the third convolution respectively. The pooling layer is frozen, and a squeeze-and-excitation module is used to replace the pooling layer, and a fuzzy clustering algorithm is introduced into the squeeze-and-excitation module.

[0010] Preferably, the spatio-temporal risk judgment model based on historical marketing data training combined with the spatio-temporal federated learning algorithm and the risk identification sub-model includes: Pre-construct a spatio-temporal risk judgment model. When pre-constructing the spatio-temporal risk judgment model, use the spatio-temporal federated learning framework as the initial model of the spatio-temporal risk judgment model. The spatio-temporal federated learning framework consists of a graph generator, a graph neural network, and a federated learning module. Replace the graph generator with the risk identification sub-model, introduce a relational attention mechanism into the graph neural network, introduce a closed-form solution algorithm with explicit time dependence and a multi-layer perceptron MLP into the federated learning module, and introduce an output layer after the federated learning module; Identify the model hyperparameters of the risk identification sub-model corresponding to the parallel chain nodes. The federated learning module dynamically aggregates the model hyperparameters of the risk identification sub-model based on the spatio-temporal federated averaging algorithm to obtain the global hyperparameters of the spatio-temporal risk judgment model; Among them, the global hyperparameters of the spatio-temporal risk judgment model are represented as follows: (6) Among them, represents the global hyperparameters of the spatio-temporal risk judgment model, is the model hyperparameter of the th risk identification sub-model, represents the local dataset of the parallel chain nodes, are the time weighting coefficient and the space weighting coefficient of the parallel chain nodes respectively, are the time feature and the space feature of the parallel chain nodes respectively, is the number of spatio-temporal risk judgment models, represents the number of data features in the local dataset; Obtain historical marketing data, preprocess the historical marketing data, divide the preprocessed historical marketing data into a global training set and a global test set, load the pre-constructed spatio-temporal risk judgment model, and preset the activation function, loss function, and iterative training rounds of the spatio-temporal risk judgment model; Use the global training set to iteratively train the spatio-temporal risk judgment model until convergence, and output the converged spatio-temporal risk judgment model. During training, use backpropagation training and prevent overfitting through regularization techniques; Load the global test set. Using the global test set as input, execute the spatio-temporal risk judgment model, and output the global test result. Determine whether the global test result meets the preset global accuracy threshold. If the global test result meets the preset global accuracy threshold, output the converged spatio-temporal risk judgment model; If the global test result does not meet the preset global accuracy threshold, use the adaptive moment estimation optimizer to iteratively optimize the global hyperparameters of the spatio-temporal risk judgment model, and use the global training set to iteratively train the spatio-temporal risk judgment model until convergence.

[0011] Preferably, the method for performing anomaly detection on the scenario risk set includes: Load the scenario risk set, identify the key feature points in the scenario risk set, map the key feature points to the graph topology nodes, and encode them as the node connection edges of the graph topology based on the time relationship between the key feature points; The graph neural network identifies the graph topology, aggregates the associated node information of the graph topology nodes based on the relational attention mechanism, calculates the mean, standard deviation, maximum value, and minimum value of the associated node features to form an associated feature vector. The multi-layer perceptron MLP uses the associated feature vector to update the graph topology nodes to obtain a node update vector; The federated learning module performs anomaly detection on the node update vector based on the time weighting coefficient and space weighting coefficient of the corresponding parallel chain nodes of the node, and quantitatively calculates the marketing risk factor of the parallel chain nodes; Load the marketing risk factors of at least one group of parallel chain nodes, calculate the analytical solution of the marketing risk factors based on the closed-form solution algorithm of explicit time dependence, use the analytical solution of the marketing risk factors as the comprehensive risk value, and upload the comprehensive risk value to the blockchain platform.

[0012] Preferably, when quantitatively calculating the marketing risk factor of the parallel chain node, the marketing risk factor quantization decision formula is expressed as: (7) where, represents the quantization value of the marketing risk factor of the parallel chain node, are the node update vector, the mean value of the associated node features, and the standard deviation respectively, represents the marketing task weight corresponding to the node update vector in the parallel chain node, represent the time feature component and space feature component of the node update vector in the parallel chain node respectively; The comprehensive risk value is calculated by the following formula: (8) where, represents the comprehensive risk value, is the activation function of the spatio-temporal risk judgment model, represents the bias term, Represents the quantization value of the marketing risk factor of the parallel chain node.

[0013] On the other hand, the present invention also provides a digital monitoring system for multi-scenario marketing business risks. The multi-scenario marketing business risk digital monitoring system includes: A data acquisition module, which is used to acquire real-time marketing data of multiple scenarios, preprocess the real-time marketing data to obtain a marketing data set, and send it to the parallel chain node allocated by the blockchain platform; A parallel chain node, which is used to set the marketing data set as a local data set, train a risk identification sub-model based on the fuzzy clustering algorithm using the local data set, pre-identify the marketing data set based on the pre-trained risk identification sub-model to pre-identify scenario risk data, and generate a scenario risk set; A blockchain platform, which captures historical marketing data and trains a spatio-temporal risk judgment model that combines the spatio-temporal federated learning algorithm and the risk identification sub-model based on the historical marketing data; A comprehensive risk judgment module, which takes the scenario risk set as input, performs anomaly detection on the scenario risk set, predicts the marketing risk factor, calculates the comprehensive risk value in combination with the marketing risk factor, and uploads the comprehensive risk value to the blockchain platform, and digitally visualizes and presents the comprehensive risk value corresponding to the scenario.

[0014] Preferably, the data acquisition module includes: A data marking unit, which is used to load real-time marketing data, perform missing value and outlier cleaning processing on the real-time marketing data, and mark the data based on the source of the real-time marketing data with the business scenario type to obtain real-time marketing data containing scenario type labels; A task weighting unit, which uses the principal component analysis method to weight the marketing tasks corresponding to the business scenario to obtain the weighted real-time marketing data. When multiple groups of marketing tasks correspond to the business scenario, linear combination weighting processing is performed on the weights of the multiple groups of marketing tasks; An interference coefficient calculation unit, which is used to load the marked and weighted real-time marketing data, perform spectrum analysis processing on the real-time marketing data, determine the main frequency distribution of the real-time marketing data, identify the position and bandwidth of the noise signal in the frequency domain, and calculate the noise interference coefficient based on the adaptive threshold algorithm; A wavelet denoising unit, which is used to judge whether the noise interference coefficient in the real-time marketing data exceeds a preset noise frequency domain value. The noise frequency domain value is determined by the Stein unbiased risk estimation criterion. If it exceeds the preset noise frequency domain value, wavelet denoising processing is performed on the real-time marketing data based on the db6 wavelet denoising function. If it does not exceed the preset noise frequency domain value, wavelet denoising processing is performed on the real-time marketing data based on the db1 wavelet denoising function; A feature point extraction unit is used to integrate the real-time marketing data after wavelet denoising processing, extract the key feature points of the real-time marketing data based on the AMPD algorithm combined with variational mode decomposition, and integrate at least one set of key feature points to form a marketing data set.

[0015] Compared with the prior art, the embodiments of the present application mainly have the following beneficial effects: In the embodiments of the present invention, by using the blockchain platform and the parallel chain nodes, the problems of distributed storage and sharing of multi-scenario data are solved, the phenomenon of data islands is avoided, the credibility and security of the data are improved, and through the real-time collection of multi-scenario data and multi-dimensional data preprocessing, it can be ensured that the spatio-temporal risk judgment model can more comprehensively and accurately identify and predict the potential risks in the multi-scenario marketing scenarios, achieve full coverage of multi-scenario risks, and improve the comprehensiveness and accuracy of risk identification.

[0016] In the embodiments of the present invention, when preprocessing the real-time marketing data, spectrum analysis processing is performed on the real-time marketing data to determine the main frequency distribution of the real-time marketing data, the noise interference coefficient is calculated based on the adaptive threshold algorithm, and the noise frequency domain value is determined through the Stein unbiased risk estimation criterion. Through spectrum analysis and the calculation of the noise interference coefficient, noise can be more accurately identified and removed, and the effective information of the signal can be retained, thereby improving the quality of the signal. According to the noise interference coefficient and the noise frequency domain value determined by the Stein unbiased risk estimation criterion, the most suitable wavelet denoising function can be selected to achieve the best denoising effect. The preprocessed real-time marketing data, due to the reduction of noise and the improvement of signal quality, provides a more accurate and reliable data basis for the training of the risk identification sub-model. This helps to improve the training effect of the model and enables it to more accurately identify the risk patterns in the marketing business.

[0017] In the embodiments of the present invention, a risk identification sub-model is provided. The risk identification sub-model can utilize its powerful capabilities in image recognition and feature extraction by using the convolutional neural network (CNN) as the initial model. Introducing the residual connection layer helps to solve the problem of gradient disappearance in deep networks and improve the training efficiency and performance of the model. By freezing the pooling layer, the model can make full use of these original feature information in subsequent training, avoiding ignoring some potential risk signals due to the pooling operation. The pooling layer is replaced by a squeeze-and-excitation module, and a fuzzy clustering algorithm is introduced into the squeeze-and-excitation module. The squeeze-and-excitation module can adaptively adjust the importance of the feature channels, highlight the features that are more helpful for risk identification, and suppress the unimportant features, thereby improving the computational efficiency and performance of the model. The fuzzy clustering algorithm can cluster similar data points together to form different clusters, helping the model better understand and process the distribution structure of the data.

[0018] In the embodiments of the present invention, the constructed spatio-temporal risk judgment model can comprehensively consider the spatio-temporal characteristics in historical marketing data, and improve the accuracy of risk identification through federated learning and graph neural networks. It can effectively learn from distributed data and make accurate risk predictions while protecting data privacy. The federated learning module dynamically aggregates the model hyperparameters of the risk identification sub-model based on the spatio-temporal federated averaging algorithm. This dynamic aggregation method can comprehensively consider the parameter information of each sub-model to obtain the global hyperparameters of the spatio-temporal risk judgment model. Since the data distributions and risk characteristics of different nodes may vary, dynamic aggregation can effectively integrate these differences, enabling the global hyperparameters to better reflect the characteristics of the overall data, thereby improving the generalization ability and stability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 FIG. is a schematic diagram of the implementation process of the multi-scenario marketing business risk digital monitoring method provided by the present invention.

[0020] Figure 2 FIG. shows a schematic diagram of the implementation process of the real-time marketing data preprocessing method.

[0021] Figure 3 FIG. shows a schematic diagram of the implementation process of the method for training a risk identification sub-model based on a fuzzy clustering algorithm using a local data set.

[0022] Figure 4 FIG. shows a schematic diagram of the implementation process of training a spatio-temporal risk judgment model based on historical marketing data in combination with a spatio-temporal federated learning algorithm and a risk identification sub-model.

[0023] Figure 5 FIG. shows a schematic diagram of the implementation process of the method for detecting anomalies in a scenario risk set.

[0024] Figure 6 FIG. shows a schematic diagram of the structure of a multi-scenario marketing business risk digital monitoring system. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs; the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above drawings are intended to cover non-exclusive inclusion. The terms "first", "second", etc. in the specification and claims of this application or the above drawings are used to distinguish different objects and not to describe a specific order.

[0026] When the intelligent analysis end of the existing system screens abnormal business objects, it depends on the data distribution of the time interval for screening, mainly focuses on the analysis of the time dimension, lacks the comprehensive analysis of multi-dimensional data (such as space, user behavior, marketing activity characteristics, etc.), and the identification of deviant business objects is relatively simple. It is impossible to conduct risk identification and prediction in multi-dimensional and multi-scenario. To solve the above problems, we propose a multi-scenario marketing business risk digital monitoring system and method. Briefly, when the method is implemented, it first obtains real-time marketing data of multiple scenarios, preprocesses the real-time marketing data, and the parallel chain node 200 sets the marketing data set as the local data set. The risk identification sub-model based on the fuzzy clustering algorithm is trained using the local data set. Based on the pre-trained risk identification sub-model, the marketing data set is pre-identified to generate a scenario risk set. Then, the spatio-temporal risk judgment model combining the spatio-temporal federated learning algorithm and the risk identification sub-model is trained based on historical marketing data. The parallel chain node 200 takes the scenario risk set as the input, performs anomaly detection on the scenario risk set, predicts the marketing risk factor, calculates the comprehensive risk value in combination with the marketing risk factor, and uploads the comprehensive risk value to the blockchain platform 300. In the embodiment of the present invention, by using the blockchain platform 300 and the parallel chain node 200, the problems of distributed storage and sharing of multi-scenario data are solved, the phenomenon of data islands is avoided, the credibility and security of the data are improved, and through the real-time collection of multi-scenario data and the preprocessing of multi-dimensional data, it can be ensured that the spatio-temporal risk judgment model can more comprehensively and accurately identify and predict potential risks in multi-scenario marketing scenarios, achieve full coverage of multi-scenario risks, and improve the comprehensiveness and accuracy of risk identification.

[0027] An embodiment of the present invention provides a multi-scenario marketing business risk digital monitoring method, Figure 1 which shows a schematic implementation flow diagram of the multi-scenario marketing business risk digital monitoring method. The multi-scenario marketing business risk digital monitoring method specifically includes: Step S10, obtaining real-time marketing data of multiple scenarios, preprocessing the real-time marketing data to obtain a marketing data set, and sending it to the parallel chain node 200 allocated by the blockchain platform 300; It should be noted that the blockchain platform 300 is communicatively connected to multiple groups of parallel chain nodes 200 in a one - to - n manner. In this embodiment, the parallel chain nodes 200 are a specific type of nodes in the blockchain platform 300. It can be understood that each group of parallel chain nodes 200 corresponds to a group of risk identification sub - models. Therefore, the parameters or weights of the risk identification sub - models can be equivalent to the parameters or weights of the parallel chain nodes 200. The parallel chain nodes 200 are used to process and store data of a specific blockchain (parallel chain). A parallel chain is an independent blockchain that runs in parallel with the main chain and shares the security and consensus mechanism of the main chain. The blockchain platform 300 usually includes a main chain (such as the relay chain of Polkadot), and the main chain is responsible for managing and coordinating multiple parallel chains. The parallel chain shares the security and consensus mechanism through the main chain but can operate independently.

[0028] Step S20: The parallel chain node 200 sets the marketing data set as the local data set, trains a risk identification sub - model based on the fuzzy clustering algorithm using the local data set, pre - identifies the marketing data set based on the pre - trained risk identification sub - model to pre - identify scenario risk data, and generates a scenario risk set. In the embodiment of the present invention, the parallel chain node 200 processes the marketing data set as the local data set, avoiding the privacy risks brought by centralized storage and processing of data. The distributed computing architecture improves the scalability and flexibility of the system, and can better adapt to the needs of large - scale data processing. Pre - identifying the marketing data set based on the pre - trained risk identification sub - model can quickly generate a scenario risk set, providing preliminary results for subsequent in - depth risk analysis. The pre - identification process can significantly reduce the data volume and improve the processing efficiency of the system.

[0029] Step S30: The blockchain platform 300 grabs historical marketing data and trains a spatio - temporal risk judgment model that combines the spatio - temporal federated learning algorithm and the risk identification sub - model based on the historical marketing data. In the embodiment of the present invention, by grabbing historical marketing data, the system can make full use of past experience and patterns, providing richer training data for the risk judgment model. The analysis of historical data helps to discover long - term trends and periodic risks.

[0030] Step S40: The parallel chain node 200 takes the scenario risk set as the input, performs anomaly detection on the scenario risk set, predicts marketing risk factors, calculates a comprehensive risk value in combination with the marketing risk factors, and uploads the comprehensive risk value to the blockchain platform 300, and digitally visualizes the comprehensive risk value corresponding to the scenario.

[0031] In the embodiments of the present invention, the blockchain platform 300 and the parallel chain nodes 200 are utilized to solve the problems of distributed storage and sharing of multi-scenario data, avoid the data island phenomenon, enhance the credibility and security of data, and through the real-time collection of multi-scenario data and multi-dimensional data preprocessing, it is possible to ensure that the spatio-temporal risk judgment model can more comprehensively and accurately identify and predict potential risks in multi-scenario marketing scenarios, achieve full coverage of multi-scenario risks, and improve the comprehensiveness and accuracy of risk identification.

[0032] The embodiments of the present invention provide a method for preprocessing real-time marketing data. Figure 2 The schematic diagram of the implementation process of the method for preprocessing real-time marketing data is shown. The method for preprocessing real-time marketing data specifically includes: Step S101, load the real-time marketing data, and perform missing value and outlier cleaning processing on the real-time marketing data. The data after cleaning is more stable and reliable, enabling the model trained based on these data to have better robustness and be able to better adapt to different business scenarios and data changes. In this embodiment, the real-time marketing data includes but is not limited to social media data, website analysis data, sales data, customer feedback data, advertising placement data, customer behavior data, real-time event data, and geographical location data.

[0033] Step S102, based on the source of the real-time marketing data, mark the data with the business scenario type to obtain the real-time marketing data containing the scenario type tags. It should be noted that by marking the business scenario type, the system can identify and distinguish the data characteristics in different scenarios, providing support for subsequent scenario-based analysis and risk identification. This enables the system to better adapt to multi-scenario marketing operations, improve the comprehensiveness and accuracy of risk monitoring. The marked data can be classified and managed according to the scenario type, facilitating subsequent targeted processing and analysis, and improving the efficiency and effect of data processing. The marketing data acquisition sources include but are not limited to social media platforms, website statistics tools, monitoring devices, and online customer service. The data involves user traffic, behavior paths, page residence times, etc., and is applicable to website optimization and user behavior analysis scenarios. In the embodiments of the present invention, a dynamic label management system is used to generate and update labels according to real-time data. For example, by combining user behavior data and business type data, labels such as "high-potential customers" and "high-value wealth customers" can be generated and these labels can be updated in real time to reflect the latest behaviors and preferences of users.

[0034] Step S103, use the principal component analysis method to assign weights to the marketing tasks corresponding to the business scenario to obtain the weighted real-time marketing data. When the business scenario corresponds to multiple groups of marketing tasks, perform linear combination weighting processing on the weights of the multiple groups of marketing tasks. In the embodiments of the present invention, the principal component analysis method can effectively evaluate the importance of each marketing task and assign reasonable weights to each task. This helps to highlight important features, reduce the interference of noise features, and improve the interpretability and accuracy of the model. When weighting the real-time marketing data corresponding to the business scenario using the principal component analysis method, first calculate the covariance matrix of the real-time marketing data to understand the relationship between variables. Then calculate the eigenvalues and eigenvectors of the covariance matrix. The eigenvalues represent the importance of each principal component. Select the main principal components according to the magnitudes of the eigenvalues. Usually, select the principal components with a cumulative variance contribution rate reaching a certain proportion (such as 75%). Finally, calculate the weights of each marketing task according to the results of the principal component analysis. When performing linear combination weighting on multiple groups of marketing task weights, first determine the weights of each group according to the results of the principal component analysis, and then perform a linear combination of the weights of each group of marketing tasks.

[0035] Step S104: Load the marked and weighted real-time marketing data, perform spectral analysis processing on the real-time marketing data to determine the main frequency distribution of the real-time marketing data, identify the position and bandwidth of the noise signal in the frequency domain, and calculate the noise interference coefficient based on the adaptive threshold algorithm. In the embodiments of the present invention, through spectral analysis, the position and bandwidth of the noise signal in the frequency domain can be accurately identified, providing a precise target for subsequent noise processing. This helps to more effectively remove noise and improve the purity of the data. The noise interference coefficient is an index used to quantify the influence degree of the noise signal on the real-time marketing data. It analyzes the characteristics of the signal in the frequency domain, identifies the position and bandwidth of the noise signal, and is calculated based on the adaptive threshold algorithm. Specifically, the noise interference coefficient reflects the intensity and distribution of the noise signal in the frequency domain and can help us evaluate the impact of noise on the signal quality.

[0036] It should be noted that the noise interference coefficient provides a basis for selecting an appropriate wavelet denoising function. According to the magnitude of the noise interference coefficient, different strength denoising functions can be selected to process the signal. For example, when the noise interference coefficient is relatively high, a stronger denoising function (such as the db6 wavelet) may be needed to remove the noise; while when the noise interference coefficient is relatively low, a milder denoising function (such as the db1 wavelet) can be selected to avoid signal distortion caused by over-denoising.

[0037] In the embodiments of the present invention, the noise interference coefficient is calculated by the following formula: (1) (2) (3) Where represents the noise interference coefficient, is the number of noise signals, is the signal length, represents the average bandwidth of the noise signal, are respectively the maximum and minimum values of the power spectral density of the noise signal within the signal length, represents the current noise signal of the signal bandwidth, represents the current noise signal of the power spectral density. The calculation formula of the noise interference coefficient provides a method to quantify the influence degree of noise on the signal. These parameters comprehensively reflect the characteristics of the noise signal in the time domain and frequency domain, making the calculated noise interference coefficient able to comprehensively evaluate the influence of noise on the signal.

[0038] Step S105, determine whether the noise interference coefficient in the real-time marketing data exceeds a preset noise frequency domain value, where the noise frequency domain value is determined by the Stein unbiased risk estimation criterion; The noise frequency domain value is determined by the following formula: (4) where, represents the noise frequency domain value, represents the initial value of the noise frequency domain value, represents the sign function, which is used to judge the positive and negative attributes of the input signal of, represents the noise level in the Stein unbiased risk estimation criterion.

[0039] It should be noted that the noise frequency domain value refers to a threshold value used to quantify the influence degree of the noise signal on the original signal in frequency domain analysis. It is usually used to determine which parts of the signal are considered noise and which parts are useful signals. In signal processing, the noise frequency domain value can help identify and remove noise, thereby improving the quality of the signal. The Stein unbiased risk estimation criterion is a statistical method used to estimate the risk (such as mean square error) in the signal denoising process. The SURE criterion is very useful in denoising algorithms because it provides a risk estimation method that does not require information on the statistical characteristics of the true signal and noise. The SURE criterion does not depend on prior knowledge of the noise (such as the variance of the noise), so it is applicable to various different noise environments and has good self-adaptability.

[0040] Step S106, if it exceeds the preset noise frequency domain value, perform wavelet denoising processing on the real-time marketing data based on the db6 wavelet denoising function; Step S107, if it does not exceed the preset noise frequency domain value, perform wavelet denoising processing on the real-time marketing data based on the db1 wavelet denoising function. Using the db1 wavelet denoising function for processing can perform lightweight denoising processing and avoid signal distortion caused by overprocessing. The db1 wavelet has high computational efficiency and is suitable for processing low-frequency noise; Step S108: Integrate the real-time marketing data after wavelet denoising processing, extract the key feature points of the real-time marketing data based on the AMPD algorithm combined with variational mode decomposition, and integrate at least one set of key feature points to form a marketing data set. By combining the AMPD algorithm with variational mode decomposition, the key feature points in the data can be effectively extracted, redundant information can be removed, and important features can be highlighted. This helps to improve the training efficiency and prediction performance of the model.

[0041] In the embodiment of the present invention, when preprocessing the real-time marketing data, spectrum analysis processing is performed on the real-time marketing data to determine the main frequency distribution of the real-time marketing data, the noise interference coefficient is calculated based on the adaptive threshold algorithm, and the noise frequency domain value is determined through the Stein unbiased risk estimation criterion. Through spectrum analysis and the calculation of the noise interference coefficient, noise can be more accurately identified and removed, and the effective information of the signal can be retained, thereby improving the quality of the signal. According to the noise interference coefficient and the noise frequency domain value determined by the Stein unbiased risk estimation criterion, the most suitable wavelet denoising function can be selected to achieve the best denoising effect. The preprocessed real-time marketing data, due to reduced noise and improved signal quality, provides a more accurate and reliable data basis for the training of the risk identification sub-model. This helps to improve the training effect of the model and enables it to more accurately identify the risk patterns in marketing operations.

[0042] The embodiment of the present invention provides a method for training a risk identification sub-model based on a fuzzy clustering algorithm using a local data set. Figure 3 The figure shows a schematic implementation flow chart of a method for training a risk identification sub-model based on a fuzzy clustering algorithm using a local data set. The method for training a risk identification sub-model based on a fuzzy clustering algorithm using a local data set specifically includes: Step S201: Use a convolutional neural network model as the initial model of the risk identification sub-model to pre-construct the risk identification sub-model. The risk identification sub-model consists of an input layer, a convolution layer, a pooling layer and an output layer. The convolution layer consists of three layers of convolution. The convolution layer includes the first convolution, the second convolution and the third convolution. The residual connection layer is introduced after the first convolution, the second convolution and the third convolution respectively, and the pooling layer is frozen. By freezing the pooling layer, the model can make full use of the original feature information in subsequent training to avoid ignoring some potential risk signals due to the pooling operation. The compression excitation module is used to replace the pooling layer, and the fuzzy clustering algorithm is introduced in the compression excitation module. The compression excitation module can adaptively adjust the importance of the feature channel, highlight the features that are more helpful for risk identification, and suppress unimportant features, thereby improving the computational efficiency and performance of the model. The fuzzy clustering algorithm can cluster similar data points together to form different clusters, helping the model to better understand and process the distribution structure of the data. The introduction of the fuzzy clustering algorithm in the risk identification sub-model can make more reasonable grouping and analysis of the input data, and dig out the potential risk categories and patterns in the data. This combination method can make the model more focused on key risk features and improve the accuracy and reliability of risk identification.

[0043] In an embodiment of the present invention, by using a convolutional neural network (CNN) as an initial model, its powerful capabilities in image recognition and feature extraction can be utilized. The introduction of a residual connection layer helps solve the gradient vanishing problem in deep networks and improves the training efficiency and performance of the model. The gradient may become very small (gradient vanishing) or very large (gradient explosion) during back propagation, making the network difficult to train. The residual connection layer reduces the risk of gradient vanishing by directly connecting the input to the output so that the gradient can flow directly through these connections. It also helps to stabilize the size of the gradient and prevent gradient explosion. For example, in some complex risk identification tasks, the model needs to be able to capture long-term dependencies and deep patterns. A deeper network structure combined with a residual connection layer can better achieve this goal. Replacing the pooling layer with a compressed excitation module and introducing a fuzzy clustering algorithm therein can enhance the model's adaptability to data distribution and improve the accuracy of feature selection.

[0044] Step S202, loading the pre-built risk identification sub-model, setting the sub-model iteration rounds, maximum number of iterations, learning rate, and hyperparameters; In an embodiment of the present invention, the sub-model iteration rounds can be 60 times, the maximum number of iterations can be 80-120 times, the learning rate can be 0.001, and the hyperparameters can be the number of convolutional layer filters, the convolution kernel size, and the batch size.

[0045] Step S203: Obtain the local dataset, extract the training data from the local dataset based on the K-fold cross-validation method, and pre-train the risk identification sub-model with the training data to obtain the initial learning rate and initial hyperparameters of the risk identification sub-model. The K-fold cross-validation method is a commonly used model evaluation method that can effectively reduce the impact of data partitioning randomness on model performance evaluation. By dividing the local dataset into K subsets, using one subset as the validation set each time and the remaining subsets as the training set for model training and evaluation, more stable and accurate model performance evaluation results can be obtained. Pre-training enables the model to have a preliminary learning and understanding of the data before formal training, laying a good foundation for subsequent training. Through pre-training on the local dataset, the risk identification sub-model can learn some basic risk characteristics and patterns, thus converging to the optimal solution faster in subsequent transfer learning and iterative training.

[0046] Step S204: Load the local dataset, divide the local dataset into a training subset and a validation subset. In the embodiments of the present invention, the allocation ratio of the training subset to the validation subset can be 4:1. The training subset iteratively trains the risk identification sub-model based on transfer learning until convergence, and outputs the converged risk identification sub-model. Transfer learning is a method that uses existing knowledge and experience to accelerate the training of new models. In the training of the risk identification sub-model, through transfer learning, the knowledge and features learned on other related tasks or datasets can be applied to the current training task, helping the model converge to the optimal solution faster. For example, if there is already a model trained on a similar risk identification task, then the parameters of this model can be used as the initial parameters of the current model, and then fine-tuned on the local dataset. This can greatly reduce the training time of the model, improve the training efficiency, and also improve the performance of the model.

[0047] Among them, when training the risk identification sub-model, the learning rate is fine-tuned using cosine decay. The learning rate at the th iteration is (5) Among them, represents the initial learning rate of the risk identification sub-model, is the maximum number of iterations of the sub-model; Step S205: Obtain the validation subset, verify the performance of the risk identification sub-model using the validation subset, and output the F1 score, an evaluation metric of the risk identification sub-model. It should be noted that the F1 score is an evaluation metric that comprehensively considers precision and recall, and can more comprehensively reflect the performance of the model. Precision refers to the proportion of samples correctly predicted by the model in the total samples, and recall refers to the proportion of positive example samples correctly predicted by the model in all positive example samples. The F1 score is the harmonic mean of precision and recall, which can achieve a balance between the two. In the embodiments of the present invention, the preset score threshold can be 0.75.

[0048] Step S206, determine whether the F1 score of the risk identification sub-model exceeds the preset score threshold; Step S207, if it exceeds the preset score threshold, output the converged risk identification sub-model.

[0049] If it does not exceed the preset score threshold, return to step S204 and continue to iteratively train the risk identification sub-model.

[0050] In the embodiments of the present invention, a risk identification sub-model is provided. By using a convolutional neural network (CNN) as the initial model, the risk identification sub-model can utilize its powerful capabilities in image recognition and feature extraction. Introducing a residual connection layer helps to solve the problem of gradient disappearance in deep networks and improve the training efficiency and performance of the model. By freezing the pooling layer, the model can make full use of these original feature information in subsequent training, avoiding ignoring some potential risk signals due to pooling operations. Replace the pooling layer with a squeeze-and-excitation module, and introduce a fuzzy clustering algorithm in the squeeze-and-excitation module. The squeeze-and-excitation module can adaptively adjust the importance of feature channels, highlight the features more helpful for risk identification, and suppress unimportant features, thereby improving the computational efficiency and performance of the model. The fuzzy clustering algorithm can cluster similar data points together to form different clusters, helping the model better understand and process the distribution structure of the data.

[0051] The embodiments of the present invention provide a spatio-temporal risk judgment model based on historical marketing data training combined with a spatio-temporal federated learning algorithm and a risk identification sub-model. Figure 4 Fig. shows a schematic diagram of the implementation process of the spatio-temporal risk judgment model based on historical marketing data training combined with a spatio-temporal federated learning algorithm and a risk identification sub-model. The spatio-temporal risk judgment model based on historical marketing data training combined with a spatio-temporal federated learning algorithm and a risk identification sub-model specifically includes: Step S301: Pre-build a spatio-temporal risk judgment model. When pre-building the spatio-temporal risk judgment model, use the spatio-temporal federated learning framework as the initial model of the spatio-temporal risk judgment model. The spatio-temporal federated learning framework consists of a graph generator, a graph neural network, and a federated learning module. Replace the graph generator with a risk identification sub-model, introduce a relational attention mechanism into the graph neural network, introduce a closed-form solution algorithm with explicit time dependence and a multi-layer perceptron (MLP) into the federated learning module, and introduce an output layer after the federated learning module. In the embodiment of the present invention, replacing the graph generator with a risk identification sub-model enables the model to focus more on risk-related feature extraction when generating node and edge information. Introducing a relational attention mechanism into the graph neural network allows the model to assign different weights according to the strength and importance of the relationships when processing the relationships between nodes, thereby more accurately capturing the transmission path and impact degree of risks between nodes. Introducing a closed-form solution algorithm with explicit time dependence and a multi-layer perceptron (MLP) into the federated learning module can better process time series data and complex non-linear relationships, improving the sensitivity and prediction ability of the model to risk changes. Finally, introducing an output layer ensures that the model can give a clear risk judgment result.

[0052] Step S302: Identify the model hyperparameters of the risk identification sub-model corresponding to the parallel chain node 200. The federated learning module dynamically aggregates the model hyperparameters of the risk identification sub-model based on the spatio-temporal federated averaging algorithm to obtain the global hyperparameters of the spatio-temporal risk judgment model. In the embodiment of the present invention, the federated learning module dynamically aggregates the model hyperparameters of the risk identification sub-model based on the spatio-temporal federated averaging algorithm. This dynamic aggregation method can comprehensively consider the parameter information of each sub-model to obtain the global hyperparameters of the spatio-temporal risk judgment model. Since the data distributions and risk characteristics of different nodes may vary, dynamic aggregation can effectively integrate these differences, making the global hyperparameters better reflect the characteristics of the overall data, thereby improving the generalization ability and stability of the model.

[0053] Among them, the global hyperparameters of the spatio-temporal risk judgment model are represented as follows: (6) Among them, represents the global hyperparameters of the spatio-temporal risk judgment model, is the model hyperparameters of the th risk identification sub-model, represents the local dataset of the parallel chain node 200, are the time weighting coefficient and space weighting coefficient of the parallel chain node 200 respectively, are the time feature and space feature of the parallel chain node 200 respectively, The number of spatio-temporal risk judgment models, or it can also be the number of parallel chain nodes 200. Indicates the number of data features in the local dataset, which can be 3 - 10. It should be noted that the time weighting coefficient is used to adjust the impact of time features on risk assessment. In different business scenarios, the importance of time features may vary. For example, in some scenarios, recent data may be more important than historical data. Therefore, the time weighting coefficient can be used to enhance the influence of recent data. The space weighting coefficient is used to adjust the impact of space features on risk assessment. In a business with a wide geographical distribution, the characteristics of different regions may have different impacts on risk. The space weighting coefficient can be used to reflect this difference, enabling the model to adjust risk assessment according to different geographical locations.

[0054] Step S303: Obtain historical marketing data, preprocess the historical marketing data, divide the preprocessed historical marketing data into a global training set and a global test set, load the pre-constructed spatio-temporal risk judgment model, and preset the activation function, loss function, and number of iterative training rounds of the spatio-temporal risk judgment model. Step S304: Iteratively train the spatio-temporal risk judgment model using the global training set until convergence, and output the converged spatio-temporal risk judgment model. During training, use backpropagation training and prevent overfitting through regularization techniques. Step S305: Load the global test set, use the global test set as the input, execute the spatio-temporal risk judgment model, and output the global test result. Step S306: Determine whether the global test result meets the preset global accuracy threshold. Step S307: If the global test result meets the preset global accuracy threshold, output the converged spatio-temporal risk judgment model. In the embodiments of the present invention, the global accuracy threshold can be 0.85 - 0.85. Step S308: If the global test result does not meet the preset global accuracy threshold, iteratively optimize the global hyperparameters of the spatio-temporal risk judgment model using the adaptive moment estimation optimizer, return to Step S303, and iteratively train the spatio-temporal risk judgment model using the global training set until convergence.

[0055] In the embodiments of the present invention, the constructed spatio-temporal risk judgment model can comprehensively consider the spatio-temporal characteristics in historical marketing data, and improve the accuracy of risk identification through federated learning and graph neural networks. It can effectively learn from distributed data and make accurate risk predictions while protecting data privacy. The federated learning module dynamically aggregates the model hyperparameters of the risk identification sub-model based on the spatio-temporal federated average algorithm. This dynamic aggregation method can comprehensively consider the parameter information of each sub-model to obtain the global hyperparameters of the spatio-temporal risk judgment model. Since the data distributions and risk characteristics of different nodes may vary, dynamic aggregation can effectively integrate these differences, making the global hyperparameters better reflect the characteristics of the overall data, thereby improving the generalization ability and stability of the model.

[0056] The embodiments of the present invention provide a method for anomaly detection of a scenario risk set. Figure 5 The schematic flowchart of the implementation of the method for anomaly detection of a scenario risk set is shown. The method for anomaly detection of a scenario risk set specifically includes: Step S401: Load the scenario risk set, identify the key feature points in the scenario risk set, map the key feature points to the graph topology nodes, and encode them as the node connection edges of the graph topology based on the time relationship between the key feature points. Mapping the key feature points to the graph topology nodes helps to construct a more representative model structure, making subsequent calculations and analyses more efficient. Step S402: The graph neural network identifies the graph topology, aggregates the associated node information of the graph topology nodes based on the relational attention mechanism, calculates the mean, standard deviation, maximum value, and minimum value of the associated node features to form an associated feature vector. The multi-layer perceptron MLP updates the graph topology nodes using the associated feature vector to obtain the node update vector. Using the relational attention mechanism to aggregate the associated node information of the graph topology nodes can pay attention to the importance differences between nodes, thereby improving the accuracy of feature extraction. By calculating the mean, standard deviation, maximum value, and minimum value of the associated node features to form an associated feature vector, the feature distribution of the nodes can be comprehensively described, providing rich information for subsequent node updates. Step S403: The federated learning module performs anomaly detection on the node update vector based on the time weighting coefficient and space weighting coefficient of the node corresponding to the parallel chain node 200, and quantitatively calculates the marketing risk factor of the parallel chain node 200. Introducing the time weighting coefficient and space weighting coefficient can take into account the different importance of different nodes in time and space, making the anomaly detection more accurate. By quantitatively calculating the marketing risk factor of the parallel chain node 200, the risk degree of each node can be quantified, providing a clear judgment basis for anomaly detection. When quantitatively calculating the marketing risk factor of the parallel chain node 200, the marketing risk factor quantification decision formula is expressed as: (7) Among them, represents the quantization value of the marketing risk factor of the parallel chain node 200, which are the node update vector, the mean value of the associated node features, and the standard deviation respectively, represents the marketing task weight corresponding to the node update vector in the parallel chain node 200, respectively represent the time feature component and the space feature component of the node update vector in the parallel chain node 200.

[0057] Step S404: Load at least one set of marketing risk factors of the parallel chain nodes 200, calculate the analytical solution of the marketing risk factors based on the closed-form solution algorithm with explicit time dependence, use the analytical solution of the marketing risk factors as the comprehensive risk value, and upload the comprehensive risk value to the blockchain platform 300.

[0058] In this embodiment, the comprehensive risk value is calculated by the following formula: (8) Among them, represents the comprehensive risk value, is the activation function of the spatio-temporal risk judgment model, represents the bias term, represents the quantization value of the marketing risk factor of the parallel chain node 200.

[0059] In the embodiment of the present invention, the closed-form solution algorithm with explicit time dependence is used to calculate the analytical solution of the marketing risk factors. The calculation speed of the analytical solution is usually faster than that of the numerical solution because the analytical solution is directly given and does not need to be approximated through iteration or simulation. It can efficiently process time series data while ensuring the accuracy of the calculation. Quantifying the marketing risk factors can more accurately evaluate the magnitude and possibility of the risks faced. Through the numerical risk indicators, it can provide data support for calculating the analytical solution of the marketing risk factors based on the closed-form solution algorithm with explicit time dependence.

[0060] The embodiment of the present invention provides a multi-scenario marketing business risk digital monitoring system, Figure 6 shows a schematic structural diagram of the multi-scenario marketing business risk digital monitoring system. The multi-scenario marketing business risk digital monitoring system specifically includes: A data acquisition module 100, configured to acquire real-time marketing data of multiple scenarios, preprocess the real-time marketing data to obtain a marketing data set, and send it to the parallel chain node 200 allocated by the blockchain platform 300; Parallel chain node 200, which is used to set the marketing data set as the local data set, train a risk identification sub-model based on the fuzzy clustering algorithm using the local data set, pre-identify the marketing data set based on the pre-trained risk identification sub-model to pre-identify scenario risk data, and generate a scenario risk set; Blockchain platform 300, which captures historical marketing data and trains a spatio-temporal risk judgment model that combines the spatio-temporal federated learning algorithm and the risk identification sub-model based on the historical marketing data; Comprehensive risk judgment module 400, which takes the scenario risk set as the input, performs anomaly detection on the scenario risk set, predicts marketing risk factors, calculates the comprehensive risk value in combination with the marketing risk factors, uploads the comprehensive risk value to the blockchain platform 300, and digitally visualizes the comprehensive risk value corresponding to the scenario.

[0061] In this embodiment, the data acquisition module 100 includes: Data marking unit 110, which is used to load real-time marketing data, clean the real-time marketing data for missing values and outliers, and mark the data based on the source of the real-time marketing data for the business scenario type to obtain real-time marketing data containing scenario type labels; Task weighting unit 120, which uses the principal component analysis method to weight the marketing tasks corresponding to the business scenario to obtain the weighted real-time marketing data. When multiple groups of marketing tasks correspond to the business scenario, linear combination weighting processing is performed on the weights of the multiple groups of marketing tasks; Interference coefficient calculation unit 130, which is used to load the marked and weighted real-time marketing data, perform spectral analysis processing on the real-time marketing data, determine the main frequency distribution of the real-time marketing data, identify the position and bandwidth of the noise signal in the frequency domain, and calculate the noise interference coefficient based on the adaptive threshold algorithm; Wavelet denoising unit 140, which is used to determine whether the noise interference coefficient in the real-time marketing data exceeds the preset noise frequency domain value. The noise frequency domain value is determined by the Stein unbiased risk estimation criterion. If it exceeds the preset noise frequency domain value, wavelet denoising processing is performed on the real-time marketing data based on the db6 wavelet denoising function. If it does not exceed the preset noise frequency domain value, wavelet denoising processing is performed on the real-time marketing data based on the db1 wavelet denoising function; Feature point extraction unit 150, which is used to integrate the real-time marketing data after wavelet denoising processing, extract the key feature points of the real-time marketing data based on the AMPD algorithm combined with variational mode decomposition, and integrate at least one group of key feature points to form a marketing data set.

[0062] It should be noted that the multi-scenario marketing business risk digital monitoring system provided by the embodiments of the present invention corresponds to the above-mentioned multi-scenario marketing business risk digital monitoring method. For the explanations, examples, beneficial effects, etc. of the relevant content, reference can be made to the corresponding content in the multi-scenario marketing business risk digital monitoring method, which will not be elaborated here.

[0063] In summary, the present invention provides a multi-scenario marketing business risk digital monitoring system and method. In the embodiments of the present invention, by using the blockchain platform 300 and the parallel chain nodes 200, the problems of distributed storage and sharing of multi-scenario data are solved, the phenomenon of data islands is avoided, the credibility and security of data are improved, and through the real-time collection of multi-scenario data and multi-dimensional data preprocessing, it can ensure that the spatio-temporal risk judgment model can more comprehensively and accurately identify and predict potential risks in multi-scenario marketing scenarios, achieve full coverage of multi-scenario risks, and improve the comprehensiveness and accuracy of risk identification.

[0064] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the protection scope of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still, without conflict and without creative efforts, combine, add, delete or make other adjustments to the features in the embodiments of the present invention according to the situation, so as to obtain different technical solutions that essentially do not deviate from the concept of the present invention, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A digital monitoring method for multi-scenario marketing business risks, characterized in that, Including: Obtain real-time marketing data for multiple scenarios, preprocess the real-time marketing data to obtain a marketing data set, and send it to the parallel chain node allocated by the blockchain platform; The parallel chain node sets the marketing data set as the local data set, trains a risk identification sub-model based on the fuzzy clustering algorithm using the local data set, and pre-identifies the marketing data set based on the pre-trained risk identification sub-model to pre-identify scenario risk data and generate a scenario risk set; The blockchain platform grabs historical marketing data and trains a spatio-temporal risk judgment model that combines the spatio-temporal federated learning algorithm and the risk identification sub-model based on the historical marketing data; The parallel chain node takes the scenario risk set as the input, performs anomaly detection on the scenario risk set, predicts the marketing risk factor, calculates the comprehensive risk value in combination with the marketing risk factor, and uploads the comprehensive risk value to the blockchain platform, and digitally visualizes the comprehensive risk value corresponding to the scenario.

2. The digital monitoring method for multi-scenario marketing business risks according to claim 1, characterized in that: The method for preprocessing the real-time marketing data includes: Load the real-time marketing data and perform missing value and outlier cleaning on the real-time marketing data; Based on the source of the real-time marketing data, mark the data with business scenario types to obtain real-time marketing data containing scenario type labels; Use the principal component analysis method to weight the marketing tasks corresponding to the business scenario to obtain the weighted real-time marketing data. When the business scenario corresponds to multiple groups of marketing tasks, perform linear combination weighting on the weights of the multiple groups of marketing tasks; Load the marked and weighted real-time marketing data, perform spectral analysis on the real-time marketing data to determine the main frequency distribution of the real-time marketing data, identify the position and bandwidth of the noise signal in the frequency domain, and calculate the noise interference coefficient based on the adaptive threshold algorithm; Judge whether the noise interference coefficient in the real-time marketing data exceeds the preset noise frequency domain value, where the noise frequency domain value is determined by the Stein unbiased risk estimation criterion; If it exceeds the preset noise frequency domain value, perform wavelet denoising on the real-time marketing data based on the db6 wavelet denoising function; If it does not exceed the preset noise frequency domain value, perform wavelet denoising on the real-time marketing data based on the db1 wavelet denoising function; Integrate the real-time marketing data after wavelet denoising processing, extract the key feature points of the real-time marketing data based on the AMPD algorithm combined with variational mode decomposition, and integrate at least one set of key feature points to form a marketing data set.

3. The digital monitoring method for multi-scenario marketing business risks according to claim 2, wherein: The noise interference coefficient is calculated by the following formula: （1） （2） （3） Among them, represents the noise interference coefficient, is the number of noise signals, is the signal length, represents the average bandwidth of the noise signal, are respectively the maximum and minimum values of the power spectral density of the noise signal within the signal length, represents the current noise signal of the signal bandwidth, represents the current noise signal of the power spectral density; The noise frequency domain value is determined by the following formula: （4） Among them, represents the noise frequency domain value, represents the initial value of the noise frequency domain value, represents the sign function, which is used to judge the positive and negative attributes of the input signal and represents the noise level in the Stein unbiased risk estimation criterion.

4. The digital monitoring method for multi-scenario marketing business risks according to claim 1, wherein: The method for training a risk identification sub-model based on the fuzzy clustering algorithm using the local data set includes: Use the convolutional neural network model as the initial model of the risk identification sub-model and pre-construct the risk identification sub-model; Load the pre-constructed risk identification sub-model and set the sub-model iteration times, maximum iteration times, learning rate, and hyperparameters; Obtain the local data set, extract the training data in the local data set based on the K-fold cross-validation method, and pre-train the risk identification sub-model using the training data to obtain the initial learning rate and initial hyperparameters of the risk identification sub-model; Load the local data set, divide the local data set into a training subset and a validation subset, and iteratively train the risk identification sub-model based on transfer learning until convergence, and output the converged risk identification sub-model; Among them, when training the risk identification sub-model, cosine decay is used to fine-tune the learning rate. The learning rate at the th iteration is: （5） Among them, represents the initial learning rate of the risk identification sub-model, is the maximum number of iterations for the sub-model; Obtain a validation subset, use the validation subset to verify the performance of the risk identification sub-model, output the F1 score of the evaluation index of the risk identification sub-model, and determine whether the F1 score of the evaluation index of the risk identification sub-model exceeds the preset score threshold. If it exceeds the preset score threshold, output the converged risk identification sub-model.

5. The digital monitoring method for multi-scenario marketing business risks according to claim 4, wherein: The risk identification sub-model consists of an input layer, a convolutional layer, a pooling layer, and an output layer. The convolutional layer consists of three convolutions. The convolutional layer includes the first convolution, the second convolution, and the third convolution. Residual connection layers are introduced after the first convolution, the second convolution, and the third convolution respectively. Freeze the pooling layer, replace the pooling layer with a squeeze-and-excitation module, and introduce a fuzzy clustering algorithm into the squeeze-and-excitation module.

6. The digital monitoring method for multi-scenario marketing business risks according to claim 5, wherein: The spatio-temporal risk judgment model based on historical marketing data training combined with the spatio-temporal federated learning algorithm and the risk identification sub-model includes: Pre-construct a spatio-temporal risk judgment model. When pre-constructing the spatio-temporal risk judgment model, use the spatio-temporal federated learning framework as the initial model of the spatio-temporal risk judgment model. The spatio-temporal federated learning framework consists of a graph generator, a graph neural network, and a federated learning module. Replace the graph generator with the risk identification sub-model, introduce a relational attention mechanism into the graph neural network, introduce a closed-form solution algorithm with explicit time dependence and a multi-layer perceptron (MLP) into the federated learning module, and introduce an output layer after the federated learning module; Identify the model hyperparameters of the risk identification sub-model corresponding to the parallel chain nodes. The federated learning module dynamically aggregates the model hyperparameters of the risk identification sub-model based on the spatio-temporal federated averaging algorithm to obtain the global hyperparameters of the spatio-temporal risk judgment model; Among them, the global hyperparameters of the spatio-temporal risk judgment model are represented as follows: （6） Among them, represents the global hyperparameter of the spatio-temporal risk judgment model, is the model hyperparameter of the th risk identification sub-model, represents the local dataset of the parallel chain node, are the time-weighted coefficient and space-weighted coefficient of the parallel chain node respectively, are the time feature and space feature of the parallel chain node respectively, is the number of spatio-temporal risk judgment models, represents the number of data features in the local dataset; Obtain historical marketing data, preprocess the historical marketing data, divide the preprocessed historical marketing data into a global training set and a global test set, load the pre-constructed spatio-temporal risk judgment model, and preset the activation function, loss function, and number of iterative training rounds of the spatio-temporal risk judgment model; Iteratively train the spatio-temporal risk judgment model with the global training set until convergence, and output the converged spatio-temporal risk judgment model. During training, use backpropagation training and prevent overfitting through regularization techniques; Load the global test set, use the global test set as the input, execute the spatio-temporal risk judgment model, output the global test result, and determine whether the global test result meets the preset global accuracy threshold. If the global test result meets the preset global accuracy threshold, output the converged spatio-temporal risk judgment model; If the global test result does not meet the preset global accuracy threshold, iteratively optimize the global hyperparameters of the spatio-temporal risk judgment model using the adaptive moment estimation optimizer, and iteratively train the spatio-temporal risk judgment model with the global training set until convergence.

7. The digital monitoring method for multi-scenario marketing business risks according to claim 6, characterized in that: The method for anomaly detection of the scenario risk set includes: Load the scenario risk set, identify the key feature points in the scenario risk set, map the key feature points to the graph topology nodes, and encode them as the node connection edges of the graph topology based on the time relationship between the key feature points; The graph neural network identifies the graph topology, aggregates the associated node information of the graph topology nodes based on the relational attention mechanism, calculates the mean, standard deviation, maximum value, and minimum value of the associated node features to form an associated feature vector, and the multi-layer perceptron (MLP) updates the graph topology nodes using the associated feature vector to obtain a node update vector; The federated learning module performs anomaly detection on the node update vector based on the time-weighted coefficient and space-weighted coefficient of the nodes corresponding to the parallel chain nodes, and quantitatively calculates the marketing risk factor of the parallel chain nodes; Load the marketing risk factors of at least one group of parallel chain nodes, calculate the analytical solution of the marketing risk factors based on the closed-form solution algorithm with explicit time dependence, use the analytical solution of the marketing risk factors as the comprehensive risk value, and upload the comprehensive risk value to the blockchain platform.

8. The digital monitoring method for multi-scenario marketing business risks as claimed in claim 7, wherein: When quantitatively calculating the marketing risk factor of the parallel chain nodes, the marketing risk factor quantization decision formula is expressed as: （7） Among them, represents the quantified value of the marketing risk factor of the parallel chain node, which are the node update vector, the mean value of the associated node features, and the standard deviation respectively, represents the marketing task weight corresponding to the node update vector in the parallel chain node, respectively represent the time feature component and the space feature component of the node update vector in the parallel chain node; The comprehensive risk value is calculated by the following formula: （8） Among them, represents the comprehensive risk value, is the activation function of the spatio-temporal risk judgment model, represents the bias term, represents the quantization value of the marketing risk factor of the parallel chain node.

9. A digital monitoring system for multi-scenario marketing business risks, which is used to implement the digital monitoring method for multi-scenario marketing business risks as described in any one of claims 1-8, and is characterized in that: The multi-scenario marketing business risk digital monitoring system includes: A data acquisition module, which is used to acquire real-time marketing data of multiple scenarios, preprocess the real-time marketing data to obtain a marketing data set, and send it to the parallel chain nodes allocated by the blockchain platform; Parallel chain nodes, which are used to set the marketing data set as the local data set, train a risk identification sub-model based on the fuzzy clustering algorithm using the local data set, pre-identify the marketing data set based on the pre-trained risk identification sub-model to pre-identify the scenario risk data, and generate a scenario risk set; A blockchain platform, which captures historical marketing data and trains a spatio-temporal risk judgment model that combines the spatio-temporal federated learning algorithm and the risk identification sub-model based on the historical marketing data; A comprehensive risk judgment module, which takes the scenario risk set as the input, performs anomaly detection on the scenario risk set, predicts the marketing risk factor, calculates the comprehensive risk value in combination with the marketing risk factor, and uploads the comprehensive risk value to the blockchain platform, and digitally visualizes and presents the comprehensive risk value corresponding to the scenario.

10. The multi-scenario marketing business risk digital monitoring system according to claim 9, wherein: The data acquisition module includes: A data marking unit, which is used to load real-time marketing data, clean the missing values and outliers of the real-time marketing data, and mark the data with the business scenario type based on the source of the real-time marketing data to obtain real-time marketing data containing scenario type labels; A task weighting unit, which uses the principal component analysis method to weight the marketing tasks corresponding to the business scenarios to obtain the weighted real-time marketing data. When there are multiple groups of marketing tasks corresponding to the business scenario, linear combination weighting is performed on the weights of the multiple groups of marketing tasks; An interference coefficient calculation unit, which is used to load the marked and weighted real-time marketing data, perform spectral analysis on the real-time marketing data, determine the main frequency distribution of the real-time marketing data, identify the position and bandwidth of the noise signal in the frequency domain, and calculate the noise interference coefficient based on the adaptive threshold algorithm; Wavelet denoising unit, which is used to determine whether the noise interference coefficient in the real-time marketing data exceeds a preset noise frequency domain value. The noise frequency domain value is determined by the Stein unbiased risk estimation criterion. If it exceeds the preset noise frequency domain value, wavelet denoising processing is performed on the real-time marketing data based on the db6 wavelet denoising function. If it does not exceed the preset noise frequency domain value, wavelet denoising processing is performed on the real-time marketing data based on the db1 wavelet denoising function; Feature point extraction unit, which is used to integrate the real-time marketing data after wavelet denoising processing, extract the key feature points of the real-time marketing data based on the AMPD algorithm combined with variational mode decomposition, and integrate at least one group of key feature points to form a marketing data set.

Images