Data risk control model construction method based on neural network

Abstract

The invention provides a data risk control model construction method based on a neural network, and relates to the technical field of risk control prediction optimization, and the method comprises the steps: combining transaction behavior data with emotion data, and carrying out the multi-step data collection, processing and model adjustment, so as to more comprehensively capture the transaction behavior characteristics and emotion state changes of a customer; according to the method, the comprehensiveness and accuracy of risk assessment are improved, and dynamic division and accurate prediction of different risk levels are realized through introduction of the risk adjustment model, so that the effect and efficiency of financial transaction risk control are remarkably improved on the basis of the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of risk control prediction and optimization, and specifically to a method for constructing a data risk control model based on a neural network. Background Art

[0002] With the rapid development of financial technology, risk control has become increasingly important in financial transaction management. Traditional financial risk control methods mainly rely on statistical analysis and rule engines to identify and manage potential risks by setting fixed thresholds and rules. These methods can effectively monitor and control risks to a certain extent, but they often have limited ability to identify complex trading behaviors and diverse emotional factors. In recent years, with the breakthroughs in big data and artificial intelligence technologies, risk control models based on machine learning, especially neural networks, have gradually emerged. Through deep learning and feature extraction of massive data, such models can more accurately identify abnormal trading behaviors and potential risks, thereby improving the accuracy and response speed of risk prediction. At the same time, combined with natural language processing and sentiment analysis technology, it can evaluate customer behavior and emotional state in a wider range of dimensions, providing financial institutions with a more comprehensive basis for risk assessment. In the prior art, the publication number is CN117273193A, and the name is a method for predicting the risk of waterlogging in power distribution equipment based on neural networks, which includes constructing rainfall data containing multiple features as the rainfall input of the waterlogging simulation model; collecting and preprocessing surface data, and using a two-dimensional hydrodynamic model to simulate urban waterlogging to obtain rainfall-waterlogging distribution samples; constructing an LSTM-CNN hybrid neural network model, training the model to determine the mapping relationship between input and output, and establishing a rainfall waterlogging prediction model; dividing the risk level of waterlogging and power outage of power distribution equipment according to the nature of the power distribution equipment; based on the rainfall prediction data, the LSTM-CNN hybrid neural network model is used to realize the real-time forecast of the risk of waterlogging and power outage of power distribution equipment. The constructed hybrid neural network model can realize high-precision dynamic prediction of rainfall to waterlogging in different time periods and spatial distributions; based on factors such as equipment characteristics and waterlogging depth and terrain, it can realize individualized waterlogging risk assessment for each power distribution equipment, and realize the refinement of risk management and control.

[0003] Although the risk control model based on neural network has shown significant advantages in improving risk identification capabilities, the existing technology still has shortcomings in integrating trading behavior and sentiment data; traditional models focus more on the trading data itself, and lack effective quantification and analysis methods for customers' emotional changes and mood fluctuations during the transaction process, resulting in information missing and prediction bias in dynamic risk assessment; in addition, when constructing risk level classification, existing models often lack comprehensive consideration of trading behavior trends and sentiment variation coefficients, making it difficult to achieve accurate adjustment and dynamic optimization of risk levels.

[0004] The above information disclosed in the above Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0005] The purpose of the present invention is to provide a method for constructing a data risk control model based on a neural network to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: A method for constructing a data risk control model based on a neural network, the specific steps include: Collect in real time the customer’s transaction behavior data during the previous monitoring period, as well as sentiment data associated with each financial transaction; Convert the processed transaction behavior data and sentiment data in the previous monitoring period respectively; Build a data risk control model based on neural network, use the transaction behavior data and sentiment data of the previous monitoring period as input features, and output the initial customer risk level value in the current monitoring period; Based on the converted transaction behavior data, calculate the transaction behavior trend coefficient of the customer in the previous monitoring period; Based on the converted sentiment data, calculate the coefficient of variation of the customer's sentiment in the previous monitoring period; Constructing a risk adjustment model for adjusting the transaction behavior data and sentiment data collected during the current monitoring period; Input the adjusted transaction behavior data and sentiment data into the data risk control model to output the predicted value of the customer risk level within the current monitoring period; The predicted customer risk level value is compared and analyzed with the initial customer risk level value, and finally an assessment result of the customer's financial transaction risk level is generated.

[0007] Compared with the prior art, the beneficial effects of the present invention are: by combining transaction behavior data with emotional data, through multi-step data collection, processing and model adjustment, it is possible to more comprehensively capture the customer's transaction behavior characteristics and emotional state changes; this method not only improves the comprehensiveness and accuracy of risk assessment, but also through the introduction of risk adjustment models, realizes the dynamic division and accurate prediction of different risk levels, thereby significantly improving the effectiveness and efficiency of financial transaction risk control on the basis of existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the overall method flow of the present invention. DETAILED DESCRIPTION

[0009] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0010] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. Embodiment 1:

[0011] See also Figure 1 , the present invention provides a technical solution: A method for constructing a data risk control model based on a neural network, which performs risk control management on a customer's financial transactions, and the specific steps include: Step S1: collect the transaction behavior data of customers in the previous monitoring period and the sentiment data related to each financial transaction from multiple data sources in real time, and clean and standardize these data to obtain a standardized data set; data cleaning includes: removing outliers, filling missing values ​​and anonymizing data to ensure data accuracy and compliance; Step S2: converting the processed transaction behavior data in the previous monitoring period into transaction frequency, transaction amount change rate and high-risk transaction ratio; Step S3: using natural language processing and speech analysis technology, converting the processed emotion data in the previous monitoring period into emotion scores and emotion fluctuation indexes; Step S4: construct a data risk control model based on a neural network, use the transaction behavior data and sentiment data of the previous monitoring period as input features, and output the initial customer risk level value in the current monitoring period; Divide the initial customer risk level value output by the data risk control model into predefined risk level ranges; Step S5: Calculate a transaction behavior trend coefficient based on the transaction behavior data of the customer in the previous monitoring period, and the transaction behavior trend coefficient is used to evaluate the transaction trend of the customer in the current monitoring period; Step S6: Analyze the emotion score and emotion fluctuation index of the previous monitoring period to calculate the emotion variation coefficient of the customer, and use the emotion variation coefficient to evaluate the degree of change of the customer's emotion state in the current monitoring period; Step S7: Obtain the sentiment variation coefficient and the transaction behavior trend coefficient calculated according to the previous monitoring period, and perform analysis and processing to construct a risk adjustment model; Use the constructed risk adjustment model to adjust the transaction behavior data and sentiment data collected during the current monitoring period; Step S8: Input the adjusted transaction behavior data and sentiment data into the data risk control model to output a predicted value of the customer risk level within the current monitoring period; The predicted customer risk level value is compared with the initial customer risk level value to determine whether they belong to the same risk level range. Based on this comparison and analysis result, the customer's financial transaction risk level assessment result is finally generated.

[0012] To further illustrate, transaction behavior and sentiment data are collected from multiple data sources and standardized data sets are obtained, including: Multiple data sources include but are not limited to the following: Transaction database: stores customers’ historical transaction records, including transaction amount, time, location and transaction type; Social media platforms: Twitter, Facebook, Weibo, to collect sentiment data related to customers; Customer service records: email and phone follow-up records, used to analyze customer sentiment and satisfaction; Market data platform: provides real-time market trends, news information and economic indicators; 1.1) Interface development: Develop RESTful APIs using Python and Flask framework to ensure secure and real-time data extraction from transaction databases and sentiment analysis platforms; 1.2) Stream processing framework: Deploy Apache Kafka and create topics to receive and transmit data streams to achieve efficient distribution and processing of data; 1.3) Data format conversion: Use JSON format to unify data formats and ensure compatibility of different data sources; The above uses mature APIs and Kafka tools to ensure the continuity, real-time and accuracy of data collection; Data cleaning, including: 2.1) Outlier detection and removal: used to identify and remove outliers in the data set to improve data quality. The specific operations are as follows: Calculate mean and standard deviation: Calculate the mean for each data field and standard deviation ; Calculate the Z-Score: For each data point X, calculate ; Remove outliers: If , then mark the data point as abnormal and remove it; 2.2) Missing value filling: used to fill in missing items in the data set. The specific operations are as follows: Identifying missing values: Detecting missing values ​​in a dataset using Pandas; Linear interpolation: For each missing data point, sum the data adjacent to the missing data point and take the average value, and use the average value as the missing data point for interpolation; 2.3) Data anonymization: used to protect data privacy. The specific operations are as follows: Select data fields: determine the fields that need to be anonymized, including but not limited to customer ID; SHA-256 encryption: Use Python's Hashlib library to encrypt these fields using the SHA-256 algorithm; Data standardization, including: 3.1) Data normalization: used to standardize data to a uniform range. The specific operations are as follows: Determine minimum and maximum values: For each data field, calculate its minimum value and maximum value ; Apply the normalization formula: Normalize the data using the formula .

[0013] Further explanation: the processed transaction behavior data in the previous monitoring period is converted into transaction frequency, transaction amount change rate and high-risk transaction ratio, including: Define the previous monitoring period as T, where T is the length of the previous monitoring period; and obtain the transaction behavior data of the current customer from the standardized data set; 4.1) In the previous monitoring period, the standardized transaction frequency is defined as ; The specific definition of the transaction frequency calculation formula before standardization is: ;in, is the total number of transactions in the previous monitoring period; The value range is [0,1]; The closer the value is to 0, the fewer financial transactions there are. The closer the value is to 1, the more financial transactions there are; 4.2) During the previous monitoring period, the calculation formula for defining the transaction amount change rate is:

[0014] in, is the rate of change of transaction amount during the previous monitoring period. is the amount of the ith financial transaction during the previous monitoring period; is the amount of the i-1th financial transaction during the previous monitoring period; The value range is [0,1]; the closer the value is to 0, the more stable the transaction amount is; The closer the value is to 1, the more unstable the transaction amount is; 4.3) Predefine risk criteria for marking high-risk transactions, including transaction amount threshold, transaction frequency threshold, and geographic location abnormality threshold. If any of the current customer's transaction amount, transaction frequency, and geographic location exceeds the corresponding threshold, the transaction will be marked as a high-risk transaction; The transaction amount threshold is recorded as ;Exceed of financial transactions were marked as high-risk transactions. The way to obtain is: The total number of transactions Divide into at least three intervals according to the transaction order, and the corresponding transaction amount in each interval shall not exceed the set ; and calculate the corresponding transaction amount average value in each interval, and select the largest value among these transaction amount average values; and according to the transaction amount change rate, perform superposition calculation on the selected maximum transaction amount average value to predict the upper limit of the transaction amount fluctuation, and then use the upper limit of the transaction amount fluctuation as ; The calculation formula is as follows:

[0015] in, is the average of the selected maximum transaction amounts; Transaction frequency threshold: Multiple transactions occurring intensively during the previous monitoring period are considered abnormal; Let the transaction frequency threshold be ;Exceed of financial transactions were marked as high-risk transactions. The way to obtain is: Divide the previous monitoring period into at least three intervals, and the corresponding transaction frequency in each interval does not exceed the set ; and calculate the corresponding transaction frequency in each interval, and select the largest value among these transaction frequencies; the selected largest transaction frequency value is used as ; Geographic location abnormality threshold: the transaction location of the customer's financial transaction behavior in the previous monitoring period; The geographic location anomaly threshold is denoted as ; in, represents the range threshold of the customer in the jth financial transaction activity area, represents the range threshold of the customer in the Jth financial transaction activity area, and J represents the total number of financial transaction activity areas of the customer; the financial transaction activity area is defined as: taking the geographical location of the customer when conducting financial transactions as the center point, calculating the maximum activity trajectory distance between the customer and the center point, using the center point and the maximum activity trajectory distance as the radius to form a circular area, and the circular area is the financial transaction activity area; Rule engine development: Use the rule engine Drools to automatically identify high-risk transactions; Real-time monitoring and alerting: Through the stream processing framework Apache Kafka, transaction data is monitored in real time and alerts are triggered based on set criteria; During the previous monitoring period, the calculation formula for defining the proportion of high-risk transactions is:

[0016] in, is the proportion of high-risk transactions in the previous monitoring period, is the number of high-risk transactions during the previous monitoring period; set up The value range is [0,1]; The closer the value is to 0, the lower the risk of financial transactions. The closer the value is to 1, the higher the risk of the financial transaction.

[0017] Further explanation: using natural language processing and speech analysis technology, the processed sentiment data is converted into sentiment scores and mood swing indexes, including: 5.1) Conversion calculation for sentiment score: Use the BERT (Bidirectional Encoder Representations from Transformers) model for sentiment analysis; the sentiment analysis steps are as follows: Text preprocessing: clean up the raw text data and remove stop words and punctuation marks; use the natural language processing toolkit NLTK and spaCy to clean up the text data; remove unnecessary punctuation marks and common stop words to reduce noise; Word segmentation: Use BERT's word segmenter to decompose the text into independent words or subword units; BERT uses the WordPiece word segmentation algorithm; Word embedding generation: Use the embedding layer of the BERT model to convert the segmented text into a vector representation; Model loading: Use HuggingFace's Transformers library to load the pre-trained BERT model and fine-tune it to suit specific sentiment analysis tasks; Feature extraction: Get features through the output of the encoder part of the BERT model. This is to get the hidden state of each word in the last layer of the model, and then apply average pooling or maximum pooling to get a feature vector for the entire sequence; and then map the feature vector to the range of [0,1] through the Sigmoid function; Sentiment classification model: Use feature vectors to determine sentiment polarity through a fully connected layer. In this embodiment, the sentiment polarities are set to positive, negative, and neutral. The specific implementation plan is as follows: Data preparation: collect text data of customers’ transaction behaviors in each transaction for model training and verification; Model environment: Use HuggingFace's Transformers library to build a BERT model environment; Training and fine-tuning: Use the optimizer AdamW to select the appropriate learning rate and apply dropout to prevent overfitting; the training data set needs to be divided into training set, validation set and test set; Result evaluation: Evaluate the fine-tuned model through indicators such as accuracy and recall, and adjust hyperparameters to optimize model performance; The calculation formula for defining the sentiment score is ,in is the weight corresponding to the h-th category result obtained by the output layer of the BERT model in the i-th financial transaction in the previous monitoring period, is the feature vector extracted from the h-th category result in the i-th financial transaction in the previous monitoring period, and The value is mapped to the range of [0,1] by the Sigmoid function; is the total number of category results in the ith financial transaction during the previous monitoring period; is the sentiment score of the ith financial transaction in the previous monitoring period; The valid value range of is [0,1]; The closer the value is to 0, the more negative the customer's emotions are. The closer the value is to 1, the more positive the customer's emotions are; The output probability of the sentiment classification model is obtained, which indicates the contribution of each feature to the classification result; the output probability is the softmax output; 5.2) Conversion calculation of the emotional volatility index: Quantify the customer's emotional fluctuations. The specific operations are: Mel frequency cepstral coefficients are used to extract the emotional features of the customer's voice in each financial transaction to obtain MFCC features; the following Mel frequency cepstral coefficients are referred to as MFCC (Mel Frequency Cepstral Coefficients); Use Python's Librosa library to extract speech features; pre-set frame length, frame shift, number of Mel filters and related parameters; Signal preprocessing: Apply pre-emphasis filter to pre-emphasize the speech signal and enhance the high-frequency components; Frame segmentation: Segment the speech signal into short frames to capture local features; Windowing: Apply a Hamming Window to each frame to reduce spectrum leakage; Fourier transform: Apply fast Fourier transform (FFT) to each windowed frame and convert it into a spectrum. Mel filtering: The spectrum is converted to Mel scale through a Mel filter bank; Logarithmic compression: Logarithmic compression of the filter bank output to simulate human ear perception; Discrete Cosine Transform (DCT): Use DCT to convert the logarithmic Mel spectrum into cepstral coefficients, take the first 12 to 13 coefficients as features, and get MFCC features; The extracted MFCC features are used as input and combined with the sentiment classifier for sentiment analysis; The calculation formula for defining the mood swing index is:

[0018] in, is the mood swing index during the previous monitoring period. represents the sentiment score of the i-1th financial transaction number; The value range is [0,1]; The closer the value is to 0, the more stable the sentiment is, indicating a lower risk in financial transactions; The closer the value is to 1, the more unstable the sentiment is, indicating that the risk of financial transactions is increasing.

[0019] To further explain, building a data risk control model based on neural network specifically includes: The neural network selects a multilayer perceptron model, which includes an input layer, a hidden layer, and an output layer; it is used to process structured financial data and sentiment data; For the input layer: determine the input features including transaction frequency, transaction amount change rate, high-risk transaction ratio, sentiment score and sentiment fluctuation index; For the hidden layer: set two hidden layers, each containing 64 nodes; The ReLU activation function is used to capture complex patterns and nonlinear relationships between features; For the output layer: output the initial customer risk level value of the current monitoring period; use the Sigmoid activation function to compress the initial customer risk level value to the range of [0,1], and record the initial customer risk level value in the current monitoring period as , where T+1 represents the current monitoring period; The training steps of the multilayer perceptron model include: 6.1) Data preparation: Collect and preprocess the transaction behavior data and sentiment data of the previous monitoring period, standardize them to [0,1], and prepare them for training; 6.2) Training configuration: Loss function: Use binary cross entropy loss function; Optimizer: Select Adam optimizer, and set the initial learning rate to 0.001, which can be adjusted according to the training performance of the model; 6.3) Training process: Set the batch size to 32 and train for 100 epochs; The data was divided into 80% training set and 20% validation set for model performance evaluation; Use cross-validation methods to ensure model generalization ability and adjust hyperparameters; For the predefined risk level interval, the transaction behavior data and sentiment data corresponding to the past M1 monitoring time periods are specifically analyzed, and the value of M1 is a positive integer of at least 3; several risk level intervals are predefined, and these risk level intervals are in positive proportion to the degree of risk of the customer's financial transaction; based on the experience and knowledge of the expert group system, several risk level intervals are set in sequence as low risk level interval [0, 0.33], medium risk level interval (0.33, 0.67], high risk level interval (0.67, 1]; The determination of the above-mentioned risk level intervals also includes but is not limited to methods based on cluster analysis and methods based on quantile regression; Methods based on cluster analysis: K-Means clustering: K-Means clustering is performed on the risk scores, and K=3 is set to represent low, medium, and high categories; The boundaries of each interval are determined based on the cluster centers and clustering results.

[0020] Hierarchical Clustering: Using the hierarchical clustering method, data points are gradually merged according to their distances, and three clusters are finally formed.

[0021] Advantages: Able to adapt to data distribution and suitable for complex data patterns.

[0022] Disadvantages: It is necessary to select appropriate clustering algorithms and parameters, and the computational complexity is high.

[0023] For the quantile regression based method: Utilize quantile regression technology to determine specific risk score thresholds for different risk levels to ensure that the number of customers or risk characteristics in each interval meet business needs.

[0024] The implementation plan for determining several risk level intervals is: Data collection and preprocessing: Integrate the transaction behavior data and sentiment data of the past M1 monitoring time periods.

[0025] Feature Engineering: Extract and select key features to calculate the risk score for each customer.

[0026] Risk level interval division: Use statistical distribution or clustering methods to divide the risk level into three intervals: low, medium and high according to the risk score.

[0027] Using the percentile method, the risk score is divided into low risk (0, 0.33), medium risk (0.33, 0.67), and high risk (0.67, 1).

[0028] Verification and adjustment: Ensure the validity and business applicability of the segmentation results and make adjustments when necessary.

[0029] Implementation and Monitoring: Apply the risk level ranges to actual risk assessments and continuously monitor their effectiveness.

[0030] when When the customer's financial transaction risk level is low, the customer can continue to trade normally and only needs to continue monitoring. Monitoring frequency: Routine monitoring once a month is sufficient to ensure normal trading behavior; Customer notification: No special notification measures are required, and normal customer relationship management is maintained; when When the risk level of the customer's financial transaction is medium, the customer needs to be warned and advised to trade prudently. Monitoring frequency: Conduct more detailed monitoring once a week, focusing on unusual trading activities; Client notification: Send warning messages to remind clients to pay attention to their trading behavior and suggest prudent actions; Providing financial guidance or advice to help clients optimize their trading behavior to reduce risk.

[0031] when When the customer's financial transaction risk level is high, further verification measures are required, including but not limited to account restrictions or transaction freezes.

[0032] Monitoring frequency: Real-time monitoring, automated system triggers alarms, and focuses on tracking high-frequency transactions and large transactions.

[0033] Client Notification: Immediately contact the client to verify their identity and transaction intent, and explain the potential risks.

[0034] Account restriction: Temporarily restrict some functions of the account before satisfactory verification or explanation is obtained. The restriction in this embodiment includes setting transaction limits or freezing suspicious transactions.

[0035] Risk Assessment: Arrange for the risk management team to conduct a detailed assessment and decide whether further legal or financial sanctions are required.

[0036] Further explanation: Calculating the trend coefficient of trading behavior includes: The calculation formula for defining the trend coefficient of trading behavior is:

[0037] Among them, a1, a2, and a3 are weight coefficients, satisfying a1+a2+a3=1, and the value range of a1, a2, and a3 is (0,1). It is the trading behavior trend coefficient calculated based on the trading behavior data of the previous monitoring period; In this embodiment, the specific values ​​of a1, a2, and a3 are determined by an expert group through experimental data; The weights a1, a2, and a3 are assigned according to the relative importance of each factor in different scenarios; the allocation strategy of these weights in the two scenarios is as follows: Scenario 1: Trading Frequency Big impact: When the transaction frequency When it is important to the trend of trading behavior and the rate of change relative to other trading amounts and high risk transaction ratio The influence of a1 is relatively large, and the value range of weight a1 is between 0.5 and 0.7. At this time, transaction frequency is the main indicator reflecting the customer's transaction activity, which has a significant impact on the overall risk assessment. In this scenario, the weight a1 is set to 0.6, which means that the importance of transaction frequency is emphasized; the weights a2 and a3 are 0.2 respectively; this allocation means that although both transaction amount changes and high-risk transactions have an impact on risk assessment, the transaction frequency has a greater impact; this weight configuration is suitable for scenarios where the customer group has frequent transactions in the short term and the amount fluctuates slightly; Scenario 2: Trading Frequency Less impact: When the importance of trading frequency F(T) to the trend of trading behavior is low and its influence is small compared with other factors, the value range of weight a1 is between 0.2 and 0.4; in this case, trading frequency is only used as an auxiliary indicator; Set the weight a1 to 0.3, and the weights a2 and a3 to 0.35 each. This allocation is suitable for a scenario where the transaction frequency is stable and the transaction amount and risk fluctuate greatly. In this case, the change in transaction amount and the proportion of high-risk transactions become the main driving factors for risk assessment. Scenario 3: Transaction Amount Change Rate Big impact: When the transaction amount change rate R(T) is crucial to the trend of trading behavior and has a greater impact than other transaction frequencies F(T) and the proportion of high-risk transactions P(T), the weight a2 ranges from 0.5 to 0.7; this indicates that the volatility of the transaction amount is the main risk consideration; In this scenario, the weight a2 is set to 0.6, highlighting the impact of transaction amount volatility; the weights a1 and a3 are each 0.2, meaning that the impact of other factors is relatively minor; this scenario is suitable for situations where the customer's transaction amount changes greatly but the frequency and risk are relatively stable; Scenario 4: High-risk transaction ratio Big impact: When the high-risk transaction ratio P(T) has a greater impact on the trading behavior trend, the value range of weight a3 is between 0.5 and 0.7; at this time, the proportion of high-risk transactions becomes the main risk assessment indicator; The weight a3 is set to 0.6, which means that the effect of the high-risk transaction ratio is amplified, while the weights a1 and a2 are each 0.2; this configuration is suitable for scenarios where the high-risk transaction ratio significantly affects the customer's overall risk level.

[0038] set up The value range of is [0,1], which is used to evaluate the transaction trend of the customer in the current monitoring period; The closer the value is to 1, the more volatility the client's trading pattern exhibits and the more likely they are to engage in high-risk transactions; The closer the value is to 0, the lower the volatility of the client's trading pattern and the less likely they are to engage in high-risk transactions.

[0039] To further explain, the calculation of the customer's emotional coefficient of variation includes: Introducing the contextual factor of the i-th financial transaction to score the sentiment Make corrections, the correction formula is as follows:

[0040] in, is the contextual factor of the ith financial transaction, used to correct the sentiment score, is the corrected sentiment score; Situational Factors The calculation method includes: The situational factors are set according to market volatility data or economic indicators or customer-specific behavior history. The market volatility data in this embodiment is described as follows:

[0041] is the market volatility factor corresponding to the ith financial exchange in the previous monitoring period, reflecting the volatility of the current market environment; is the adjustment coefficient, with a value range of [-1,1], which is used to adjust the influence of situational factors on the sentiment score; Market volatility factor Calculated by the rate of change of current and past market indicators, including but not limited to financial loan rates:

[0042] It is the financial loan interest rate corresponding to the current monitoring period; is the average interest rate on financial loans for financial transactions during the previous monitoring period; Combined with the corrected sentiment score And the emotional fluctuation index I(T) is used to calculate the emotional coefficient of variation:

[0043] in, is the corrected average sentiment score, It is the coefficient of variation of emotion calculated based on the emotion data of the previous monitoring period; b1 and b2 are weight coefficients of the corresponding parameters. The values ​​of b1 and b2 are both in the range of (0,1), and b1+b2=1. They are used to adjust the influence of the sentiment score and the mood fluctuation index on the coefficient of variation. The following is a detailed description and allocation of b1 and b2 in different scenarios: Scenario 1: The emotional fluctuation index I(T) has a greater impact: In this case, when the sentiment volatility index I(T) is extremely important and has a large impact on the predictive effect of financial transaction risks, the weight of b2 is relatively high in the range of [0.6, 0.8], while the weight of b1 is relatively low in the range of [0.2, 0.4]; If the market is in an extremely volatile state, the emotional volatility of the customer group will intensify, posing a greater threat to the stability of the financial market; in this high volatility situation, the emotional volatility index I(T) can better reflect the potential systemic risk; therefore, b2=0.7 is a reasonable choice because it aggravates the impact of I(T) and emphasizes the risk prediction of volatility; In this case, b1=0.3, with a smaller weight, means that the modified sentiment score has a weaker effect on the sentiment coefficient of variation; this allocation emphasizes the focus on sentiment fluctuations, because the impact of fluctuations on risk is considered more direct and significant; Scenario 2: Corrected average sentiment score Big impact: In this case, when the corrected sentiment score average When the accurate description of the customer's emotional state is extremely important and has a great impact, the weight of b1 is relatively high in the range of [0.6, 0.8], while the weight of b2 is relatively low in the range of [0.2, 0.4]; If the market is relatively stable, but individual customers have strong emotional reactions and deviate from the group average; in this scenario, the average sentiment score can be adjusted It is particularly important to capture individual emotional changes; therefore, b1=0.7 is a reasonable choice, which makes the change of emotional score contribute more to the emotional coefficient of variation, thus more accurately reflecting the emotional fluctuations of individual customers; In this case, b2=0.3, which means that the weight is small, which means that the impact of the sentiment volatility index is weakened; this allocation method is suitable for situations where the overall market volatility is not significant, but the individual emotional state is more critical to the prediction of trading risk; The specific values ​​of b1 and b2 are adjusted based on experimental data from the expert group.

[0044] When the coefficient of variation of emotion The closer it is to 0, the more stable the customer's emotional state is and the lower the customer's financial transaction risk is. When the coefficient of variation of emotion The closer it is to 1, the more unstable the customer's emotional state is and the higher the risk of the customer's financial transactions.

[0045] Further explanation: constructing a risk adjustment model, using the constructed risk adjustment model to adjust the transaction behavior data and sentiment data collected during the current monitoring period, specifically includes: The output of the risk adjustment model is defined as , the calculation formula is as follows:

[0046] Among them, c1 and c2 are the weight coefficients of the corresponding parameters, and c1+c2=1, and the value range of c1 and c2 is both in (0,1); when , indicating that the customer's financial transactions are at high risk and that the transaction behavior data and sentiment data collected during the current monitoring period need to be adjusted at the first level; when , indicating that the customer's financial transactions are in a medium-risk state, and secondary adjustments need to be made to the transaction behavior data and sentiment data collected during the current monitoring period; when , indicating that the customer's financial transactions are in a low-risk state, and that the transaction behavior data and sentiment data collected during the current monitoring period need to be adjusted at the third level; Level 1 adjustment, level 2 adjustment, and level 3 adjustment correspond to The values ​​are lowered in sequence, and the following adjustments of varying degrees are implemented on the transaction behavior data and sentiment data collected during the current monitoring period; The following adjustments are made to the transaction behavior data and sentiment data collected during the current monitoring period based on the risk adjustment model: Transaction amount change rate adjustment formula:

[0047] Among them, d1 is the adjustment coefficient for the rate of change of the amount, and the value range of d1 is 0.01≤d1≤0.68, reflecting Corresponding adjustments to the rate of change of transaction amounts when they increase or decrease; is the transaction amount change rate calculated during the current monitoring period. is the adjusted transaction amount change rate; in this embodiment, d1 takes a value of 0.2; Transaction frequency adjustment formula:

[0048] Wherein, d2 is the adjustment coefficient for transaction frequency, and the value range of d2 is 0.13≤d2≤0.36; is the transaction frequency calculated during the current monitoring period; is the adjusted transaction frequency; in this embodiment, d2 takes a value of 0.15; High-risk transaction ratio adjustment formula:

[0049] Wherein, d3 is the adjustment coefficient for the proportion of high-risk transactions, and the value range of d3 is 0.03≤d3≤0.65; in this embodiment, the value of d3 is 0.18; is the proportion of high-risk transactions calculated during the current monitoring period; is the adjusted proportion of high-risk transactions; Sentiment score adjustment formula:

[0050] Wherein, d4 is the adjustment coefficient of the sentiment score, and the value range of d4 is 0.11≤d4≤0.71; in this embodiment, the value of d4 is 0.31; is the sentiment score calculated during the current monitoring period; is the adjusted sentiment score; Mood swing index adjustment formula:

[0051] Wherein, d5 is the adjustment coefficient of the mood fluctuation index, and the value range of d5 is 0.10≤d5≤0.41; in this embodiment, the value of d5 is 0.18; is the mood fluctuation index calculated during the current monitoring period; is the adjusted mood swing index.

[0052] Further explanation: The predicted value of the customer risk level during the current monitoring period and the generated customer financial transaction risk assessment results specifically include: Based on the initial customer risk level value The data risk control model used; The adjusted transaction frequency , Adjusted transaction amount change rate , Adjusted high-risk transaction ratio , adjusted sentiment score and adjusted mood swing index As input features of data risk control models; In the hidden layer, the ReLU activation function is used to capture the nonlinear relationship between complex patterns and features; Use the multi-layer perceptron-based data risk control model to perform forward propagation calculations to output the predicted value of the customer risk level within the current monitoring period ; The output layer uses the Sigmoid activation function. Compress to the interval [0,1]; Will and Calculate the difference value and get ; 7.1) If ,and and If they all belong to the same risk level range, the risk level remains unchanged; 7.2) In ,and and When they do not belong to the same risk level range; when When The risk level is used as the basis for assessing the risk level of the customer's financial transactions; when When The risk level is used as the basis for assessing the risk level of the customer's financial transactions; 7.3) If ,and and If they all belong to the same risk level range, the risk level remains unchanged; 7.4) In ,and and When they do not belong to the same risk level range; Will The risk level is used as the basis for assessing the risk level of the customer's financial transactions.

[0053] In summary, this data risk control model construction method has the following advantages: First, by collecting and integrating customers' transaction behavior data and emotional data in real time, it can fully reflect customers' trading habits and emotional states. Compared with the traditional risk control model that relies only on transaction data, the introduction of emotional data analysis improves the multidimensionality and accuracy of risk assessment. The introduction of emotional data enables the model to not only capture customers' transaction behavior patterns, but also identify potential emotional fluctuations, which is of great significance for predicting customers' possible risky behaviors; emotional fluctuations indicate that customers are more impulsive in financial decision-making, thereby improving the sensitivity and timeliness of risk warnings.

[0054] Secondly, the use of neural networks as the core algorithm fully utilizes its advantages in complex data pattern recognition and deep feature extraction, significantly improving the accuracy of risk prediction. Neural networks can automatically learn and extract potential features in transaction behavior data and sentiment data through multi-level nonlinear transformations, overcoming the limitations of traditional statistical methods in processing high-dimensional, nonlinear data. In addition, the model training process can continuously optimize parameters and adapt to the dynamic changes of data, thereby maintaining high prediction performance and generalization capabilities.

[0055] Third, the proposed multi-step data processing process, including data cleaning, standardization, feature conversion and the construction of a risk adjustment model, effectively solves the problems of data heterogeneity and noise interference. By converting transaction behavior data into features such as transaction frequency, transaction amount change rate and high-risk transaction ratio, and converting sentiment data into sentiment scores and sentiment fluctuation indexes, the interpretability of the data and the stability of the model are further enhanced. At the same time, the introduction of the risk adjustment model enables the risk assessment in different monitoring time periods to be dynamically adjusted to adapt to changes in customer behavior and emotional state, thereby improving the flexibility and adaptability of the model.

[0056] Fourth, the division of risk levels is proportional to the degree of risk of the customer's financial transactions, which makes risk assessment quantifiable and categorized. This design not only facilitates financial institutions to conduct risk grading management and resource allocation, but also provides customers with more personalized risk warnings and suggestions, enhancing customers' risk management awareness and initiative.

[0057] Finally, by comparing and analyzing the initial customer risk level value with the risk-adjusted risk level forecast value, the accuracy and reliability of the risk assessment results are ensured. This comparison mechanism can effectively avoid the assessment bias caused by a single risk indicator, provide a more comprehensive and objective risk assessment result, and improve the robustness and effectiveness of the overall risk control system.

[0058] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0059] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0060] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0061] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A method for constructing a data risk control model based on a neural network, characterized in that: The specific steps include: Collect in real time the customer’s transaction behavior data during the previous monitoring period, as well as sentiment data associated with each financial transaction; Convert the processed transaction behavior data and sentiment data in the previous monitoring period respectively; Build a data risk control model based on neural network, use the transaction behavior data and sentiment data of the previous monitoring period as input features, and output the initial customer risk level value in the current monitoring period; Based on the converted transaction behavior data, calculate the transaction behavior trend coefficient of the customer in the previous monitoring period; Based on the converted sentiment data, calculate the coefficient of variation of the customer's sentiment in the previous monitoring period; Constructing a risk adjustment model for adjusting the transaction behavior data and sentiment data collected during the current monitoring period; Input the adjusted transaction behavior data and sentiment data into the data risk control model to output the predicted value of the customer risk level within the current monitoring period; The predicted customer risk level value is compared and analyzed with the initial customer risk level value, and finally an assessment result of the customer's financial transaction risk level is generated.

2. According to the method for constructing a data risk control model based on a neural network according to claim 1, it is characterized in that: The conversion of transaction behavior data specifically includes: Convert the processed transaction behavior data in the previous monitoring period into transaction frequency, transaction amount change rate and high-risk transaction ratio; Define the previous monitoring period as T; and obtain the current customer's transaction behavior data from the standardized data set; In the previous monitoring period, the standardized transaction frequency is defined as ;set up The value range is [0,1]; The closer the value is to 0, the fewer financial transactions there are. The closer the value is to 1, the more financial transactions there are; During the previous monitoring period, the calculation formula for defining the transaction amount change rate is: in, is the rate of change of transaction amount during the previous monitoring period. is the total number of transactions in the previous monitoring period, is the amount of the ith financial transaction in the previous monitoring period, is the amount of the i-1th financial transaction during the previous monitoring period; The value range is [0,1]; the closer the value is to 0, the more stable the transaction amount; The closer the value is to 1, the more unstable the transaction amount is; Predefined risk standards for marking high-risk transactions, including transaction amount threshold, transaction frequency threshold, and geographic location abnormality threshold. If any of the current customer's transaction amount, transaction frequency, and geographic location exceeds the corresponding threshold, the transaction will be marked as a high-risk transaction. The transaction amount threshold is recorded as ; Let the transaction frequency threshold be ; The geographic location anomaly threshold is recorded as ; in, represents the range threshold of the customer in the jth financial transaction activity area, represents the range threshold of the customer in the Jth financial transaction activity area, and J represents the total number of financial transaction activity areas of the customer; During the previous monitoring period, the calculation formula for defining the proportion of high-risk transactions is: in, is the proportion of high-risk transactions in the previous monitoring period, is the number of high-risk transactions during the previous monitoring period; set up The value range is [0,1]; The closer the value is to 0, the lower the risk of financial transactions. The closer the value is to 1, the higher the risk of the financial transaction.

3. The method for constructing a data risk control model based on a neural network according to claim 2, characterized in that: The conversion of sentiment data includes: Convert the processed sentiment data into sentiment scores and mood swing indexes; For the conversion calculation of sentiment scores: use the BERT model for sentiment analysis; The calculation formula for defining the sentiment score is ,in is the weight corresponding to the h-th category result obtained by the output layer of the BERT model in the i-th financial transaction in the previous monitoring period, is the feature vector extracted from the h-th category result in the i-th financial transaction in the previous monitoring period, and The value is mapped to the range of [0,1] by the Sigmoid function; is the total number of category results in the ith financial transaction during the previous monitoring period; is the sentiment score of the ith financial transaction in the previous monitoring period; The valid value range of is [0,1]; The closer the value is to 0, the more negative the customer's emotions are. The closer the value is to 1, the more positive the customer's emotions are; For the conversion calculation of the emotional fluctuation index: use the Mel frequency cepstral coefficient to extract the voice emotional characteristics of the customer in each financial transaction to obtain the MFCC features; The calculation formula for defining the mood swing index is: in, is the mood swing index during the previous monitoring period. represents the sentiment score of the i-1th financial transaction number; The value range is [0,1]; The closer the value is to 0, the more stable the sentiment is, indicating a lower risk in financial transactions; The closer the value is to 1, the more unstable the sentiment is, indicating that the risk of financial transactions is increasing.

4. The method for constructing a data risk control model based on a neural network according to claim 3, characterized in that: Build a data risk control model based on neural network, including: The neural network selects a multilayer perceptron model, which includes an input layer, a hidden layer, and an output layer; For the input layer: determine the input features including transaction frequency, transaction amount change rate, high-risk transaction ratio, sentiment score and sentiment fluctuation index; For the hidden layer: set two hidden layers, each containing 64 nodes; The ReLU activation function is used to capture complex patterns and nonlinear relationships between features; For the output layer: output the initial customer risk level value of the current monitoring period; use the Sigmoid activation function to compress the initial customer risk level value to the range of [0,1], and record the initial customer risk level value in the current monitoring period as , where T+1 represents the current monitoring period; The initial customer risk level value output by the data risk control model is divided into predefined risk level intervals, specifically: For the predefined risk level interval, the transaction behavior data and sentiment data corresponding to the past M1 monitoring time periods are specifically analyzed, and the value of M1 is a positive integer of at least 3; several risk level intervals are predefined, and several risk level intervals are set in sequence as low risk level interval [0, 0.33], medium risk level interval (0.33, 0.67], and high risk level interval (0.67, 1]; The risk level is divided according to the initial customer risk level value output by the multi-layer perceptron model, which is specifically divided into three main risk levels: when When the customer's financial transaction risk level is low, the customer can continue to trade normally and only needs to continue monitoring. when When the risk level of the customer's financial transaction is medium, the customer needs to be warned; when When the customer's financial transaction risk level is high, further verification measures are required, including but not limited to account restrictions or transaction freezes.

5. The method for constructing a data risk control model based on a neural network according to claim 4, characterized in that: Calculate the customer's transaction behavior trend coefficient in the previous monitoring period, including: The trading behavior trend coefficient is calculated based on the customer's trading behavior data in the previous monitoring period. The trading behavior trend coefficient is used to evaluate the customer's trading trend in the current monitoring period; and the calculation formula of the trading behavior trend coefficient is defined as: Among them, a1, a2, and a3 are weight coefficients, satisfying a1+a2+a3=1, and the value range of a1, a2, and a3 is (0,1). It is the trading behavior trend coefficient calculated based on the trading behavior data of the previous monitoring period; set up The value range of is [0,1], which is used to evaluate the transaction trend of the customer in the current monitoring period; The closer the value is to 1, the more volatility the client's trading pattern exhibits and the more likely they are to engage in high-risk transactions; The closer the value is to 0, the lower the volatility of the client's trading pattern and the less likely they are to engage in high-risk transactions.

6. The method for constructing a data risk control model based on a neural network according to claim 5, characterized in that: Calculate the customer's emotional variation coefficient in the previous monitoring period, including: Combine the sentiment score and mood fluctuation index of the previous monitoring period to calculate the sentiment coefficient of variation of the customer. The sentiment coefficient of variation is used to evaluate the degree of change of the customer's emotional state in the current monitoring period. The context factor of the ith financial transaction is introduced to adjust the sentiment score. Make corrections, the correction formula is as follows: in, is the contextual factor of the ith financial transaction, used to correct the sentiment score, is the corrected sentiment score; Situational Factors The calculation method includes: in is the market volatility factor corresponding to the ith financial exchange in the previous monitoring period, reflecting the volatility of the current market environment; is the adjustment coefficient, with a value range of [-1,1], which is used to adjust the influence of situational factors on the sentiment score; Market volatility factor Calculated by the rate of change of current and past market indicators, including but not limited to financial loan rates: , It is the financial loan interest rate corresponding to the current monitoring period; is the average interest rate on financial loans for financial transactions during the previous monitoring period; Combined with the corrected sentiment score And the emotional fluctuation index I(T) is used to calculate the emotional coefficient of variation: in, is the corrected average sentiment score, It is the coefficient of variation of emotion calculated based on the emotion data of the previous monitoring period; b1 and b2 are weight coefficients of the corresponding parameters. The values ​​of b1 and b2 are both in the range of (0,1), and b1+b2=1; When the coefficient of variation of emotion The closer it is to 0, the more stable the customer's emotional state is and the lower the customer's financial transaction risk is. When the coefficient of variation of emotion The closer it is to 1, the more unstable the customer's emotional state is and the higher the risk of the customer's financial transactions.

7. The method for constructing a data risk control model based on a neural network according to claim 6, characterized in that: Construct a risk adjustment model for adjusting the transaction behavior data and sentiment data collected during the current monitoring period, including: Obtain the sentiment variation coefficient and trading behavior trend coefficient calculated based on the previous monitoring period, and analyze and process them to build a risk adjustment model; The output of the risk adjustment model is defined as , the calculation formula is as follows: Among them, c1 and c2 are the weight coefficients of the corresponding parameters, and c1+c2=1, and the value range of c1 and c2 is both in (0,1); when , indicating that the customer's financial transactions are at high risk and that the transaction behavior data and sentiment data collected during the current monitoring period need to be adjusted at the first level; when , indicating that the customer's financial transactions are in a medium-risk state, and secondary adjustments need to be made to the transaction behavior data and sentiment data collected during the current monitoring period; when , indicating that the customer's financial transactions are in a low-risk state, and that the transaction behavior data and sentiment data collected during the current monitoring period need to be adjusted at the third level; The following adjustments are made to the transaction behavior data and sentiment data collected during the current monitoring period based on the risk adjustment model: Transaction amount change rate adjustment formula: Where d1 is the adjustment coefficient for the rate of change of the amount, and the value range of d1 is 0.01≤d1≤0.

68. is the transaction amount change rate calculated during the current monitoring period. is the adjusted rate of change of transaction amount; Transaction frequency adjustment formula: Wherein, d2 is the adjustment coefficient for transaction frequency, and the value range of d2 is 0.13≤d2≤0.36; is the transaction frequency calculated during the current monitoring period; is the adjusted transaction frequency; High-risk transaction ratio adjustment formula: Wherein, d3 is the adjustment coefficient for the proportion of high-risk transactions, and the value range of d3 is 0.03≤d3≤0.65; is the proportion of high-risk transactions calculated during the current monitoring period; is the adjusted proportion of high-risk transactions; Sentiment score adjustment formula: Among them, d4 is the adjustment coefficient of the sentiment score, and the value range of d4 is 0.11≤d4≤0.71; is the sentiment score calculated during the current monitoring period; is the adjusted sentiment score; Mood swing index adjustment formula: Among them, d5 is the adjustment coefficient of the mood fluctuation index, and the value range of d5 is 0.10≤d5≤0.41; is the mood fluctuation index calculated during the current monitoring period; is the adjusted mood swing index.

8. The method for constructing a data risk control model based on a neural network according to claim 7, characterized in that: Output the predicted value of the customer's risk level during the current monitoring period, and generate the customer's financial transaction risk assessment results, including: The adjusted transaction frequency , Adjusted transaction amount change rate , adjusted high-risk transaction ratio , adjusted sentiment score and adjusted mood swing index As input features of data risk control models; Use the multi-layer perceptron-based data risk control model to perform forward propagation calculations to output the predicted value of the customer risk level within the current monitoring period ; The output layer uses the Sigmoid activation function. Compress to the interval [0,1]; Will and Calculate the difference value and get ; like ,and and If they all belong to the same risk level range, the risk level remains unchanged; exist ,and and When they do not belong to the same risk level range, the following judgment is made; when When The risk level is used as the basis for assessing the risk level of the customer's financial transactions; when When The risk level is used as the basis for assessing the risk level of the customer's financial transactions; like ,and and If they all belong to the same risk level range, the risk level remains unchanged; exist ,and and If they do not belong to the same risk level range, The risk level is used as the basis for assessing the risk level of the customer's financial transactions.

Images