Risk control cross-domain risk prediction method and system combining with transfer learning

Abstract

The application provides a risk control cross-domain risk prediction method and system combined with transfer learning. Firstly, the field difference of source field and target field risk control scenes is measured to generate a field difference measurement matrix. Then, a hierarchical knowledge transfer channel is constructed based on the field difference measurement matrix, the source field risk prediction model parameters are selectively transferred, and a target field initialization risk prediction model is generated. The target field risk sample is used for dynamic adaptive training of the target field initialization risk prediction model to obtain a cross-domain risk prediction model. The target field to be evaluated sample is input into the cross-domain risk prediction model to generate a preliminary risk prediction score. Finally, the preliminary score is calibrated across the fields based on the field difference measurement matrix to generate a final risk prediction result, thereby improving the accuracy of the risk control cross-domain risk prediction.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a cross-domain risk prediction method and system for risk control that combines transfer learning. Background Technology

[0002] In the field of risk control, different business scenarios or industries have their own unique risk characteristics and data distributions, which brings many challenges to risk prediction. Traditional risk control prediction methods usually focus on a single domain, using limited data within that domain for model training and prediction. However, when faced with risk control scenarios in target domains where data is scarce, it is difficult to build accurate and effective risk prediction models due to a lack of sufficient samples. For example, in emerging financial business areas, due to the short operating time and limited accumulated risk data, directly using traditional methods for risk prediction often yields inaccurate results. On the other hand, simply applying a model trained in the source domain (rich in data) directly to the target domain will result in a significant drop in model performance due to the significant domain differences between the source and target domains, such as different data distributions and large differences in risk characteristics, failing to meet the accuracy requirements of risk control prediction in the target domain. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a cross-domain risk prediction method and system that combines transfer learning.

[0004] According to a first aspect of this application, a cross-domain risk prediction method for risk control combining transfer learning is provided, the method comprising: Perform domain difference measurement operations on the source domain risk control scenario and the target domain risk control scenario to generate a domain difference measurement matrix; Based on the domain difference metric matrix, a hierarchical knowledge transfer channel is constructed, and selective transfer processing is performed on the multi-level network parameters of the source domain risk prediction model to generate the target domain initial risk prediction model. A cross-domain risk prediction model is generated by performing dynamic adaptation training on the target domain initialization risk prediction model using a target domain risk sample set. Input the target domain sample to be evaluated into the cross-domain risk prediction model, perform risk feature transfer reasoning operation, and generate a preliminary risk prediction score; Based on the domain difference metric matrix, cross-domain calibration is performed on the preliminary risk prediction score to generate the final risk prediction result.

[0005] According to a second aspect of this application, a risk control cross-domain risk prediction system incorporating transfer learning is provided. The risk control cross-domain risk prediction system incorporating transfer learning includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the risk control cross-domain risk prediction system incorporating transfer learning implements the aforementioned risk control cross-domain risk prediction method incorporating transfer learning.

[0006] According to a third aspect of this application, a computer-readable storage medium is provided, wherein computer-executable instructions are stored therein, and when the computer-executable instructions are executed, the aforementioned risk control cross-domain risk prediction method combining transfer learning is implemented.

[0007] Based on any of the above aspects, the technical effect of this application is as follows: By performing domain difference measurement operations on risk control scenarios in the source and target domains to generate a domain difference measurement matrix, the differences between the two domains are accurately quantified, overcoming the shortcomings of traditional methods that ignore domain differences. Based on this domain difference measurement matrix, a hierarchical knowledge transfer channel is constructed, and the multi-level network parameters of the source domain risk prediction model are selectively transferred to generate an initial risk prediction model for the target domain. This achieves effective transfer of knowledge from the source domain, fully utilizes the rich data resources of the source domain, and solves the problem of model construction difficulties caused by the scarcity of data in the target domain. The initial model is dynamically adapted and trained using a risk sample set from the target domain, enabling the model to better adapt to the data characteristics and risk patterns of the target domain, further improving the model's accuracy and generalization ability. After inputting the target domain samples to be evaluated into the cross-domain risk prediction model to generate preliminary risk prediction scores, cross-domain calibration is then performed based on the domain difference measurement matrix, effectively eliminating the impact of domain differences on the prediction results. Finally, accurate and reliable risk prediction results are generated, significantly improving the accuracy of cross-domain risk prediction in risk control. Attached Figure Description

[0008] Figure 1 A flowchart illustrating the cross-domain risk prediction method for risk control combined with transfer learning provided in an embodiment of this application is shown. Figure 2 This illustration shows a schematic diagram of the component structure of a risk control cross-domain risk prediction system that incorporates transfer learning to implement the above-described risk control cross-domain risk prediction method incorporating transfer learning, as provided in an embodiment of this application. Detailed Implementation

[0009] Figure 1 This paper illustrates a flowchart of a cross-domain risk prediction method and system combining transfer learning, as provided in an embodiment of this application. The detailed steps include: Step S110: Perform domain difference measurement operation on the source domain risk control scenario and the target domain risk control scenario to generate a domain difference measurement matrix.

[0010] In this embodiment, the source domain risk control scenario is the transaction risk control scenario of an e-commerce platform, and the target domain risk control scenario is the financial business risk control scenario of a social platform. The domain difference measurement operation aims to comprehensively capture the differences between the two domains in terms of business rules, user behavior, risk performance, and data distribution. The generated domain difference measurement matrix will comprehensively reflect these differences and present them in a multi-dimensional form.

[0011] Step S111: Extract the domain attribute feature set of the source domain risk control scenario. This domain attribute feature set includes business rule features, user behavior features, risk performance features, and data distribution features.

[0012] In e-commerce platform transaction risk control scenarios, extracting business rule features needs to cover platform-defined transaction verification rules, refund risk control rules, and fraudulent transaction identification rules. For example, transaction verification rules include rules for matching transaction amounts with users' historical average spending and secondary verification rules for transactions logged in from different locations; refund risk control rules involve rules for correlating refund frequency with transaction counts and review process rules for large refunds. These business rule features are extracted in a structured form, with each rule containing attributes such as applicable conditions, execution actions, and trigger thresholds.

[0013] When extracting user behavior characteristics, it is necessary to collect various behavioral data of users on e-commerce platforms, including browsing behavior, adding to cart behavior, ordering behavior, and payment behavior. Browsing behavior characteristics include the distribution of product categories viewed, browsing duration, and browsing time periods; adding to cart behavior characteristics involve the conversion rate between added and viewed products, and the price range distribution of added products; ordering behavior characteristics include order frequency, the combination of product categories ordered, and the distribution of device information at the time of ordering; payment behavior characteristics include payment method preferences, payment time distribution, and the distribution of payment failures. These user behavior characteristics constitute a multi-dimensional feature set, with each dimension reflecting the user's performance in a specific behavior.

[0014] The extraction of risk characteristics revolves around common risk types on e-commerce platforms, such as fraudulent transaction risk, fake return risk, and account theft risk. For fraudulent transaction risk, characteristics include the distribution of fraudulent transaction amounts, the distribution of time periods in which fraudulent transactions occur, and the distribution of product categories involved in fraudulent transactions. Characteristics of fake return risk include the percentage of orders with fake returns and the distribution of price ranges for products with fake returns. Characteristics of account theft risk include the percentage of transactions after abnormal account login and the distribution of differences between the actions of the stolen account and the original user's actions.

[0015] Extracting data distribution features involves describing the distribution of the aforementioned data types, including the distribution of thresholds in business rules, the probability distribution of data across various dimensions of user behavior characteristics, and the statistical distribution of indicators in risk performance characteristics. For example, the distribution of transaction amount thresholds in business rules exhibits frequency distribution across different intervals; the distribution of browsing time in user behavior characteristics conforms to a certain probability distribution model; and the distribution of fraudulent transaction amounts in risk performance characteristics shows specific statistical patterns.

[0016] Step S112: Extract the domain attribute feature set of the target domain risk control scenario. This domain attribute feature set includes business rule features, user behavior features, risk performance features and data distribution features.

[0017] In the risk control scenario of financial business on social platforms, business rule features include social relationship chain verification rules, financial product purchase eligibility review rules, and fund transfer risk control rules. Social relationship chain verification rules involve rules linking the number of a user's friends to their eligibility to participate in financial business, and rules linking friend interaction frequency to credit assessment. Financial product purchase eligibility review rules include rules corresponding to the user's social account level and the types of products they can purchase, and rules linking the compliance of the user's social behavior to purchase limits. Fund transfer risk control rules involve rules linking transfer amount to the closeness of the user's social relationship, and rules linking transfer frequency to account anomaly risk. These business rule features are also extracted in a structured form and contain rule attributes similar to those in the source domain.

[0018] In terms of user behavior characteristics, social media platform user behavior includes social interaction behavior, content publishing behavior, and financial business participation behavior. Social interaction behavior characteristics include the distribution of friend interaction frequency, message type distribution, and interaction time distribution; content publishing behavior characteristics involve the distribution of published content type, publishing frequency, and interaction volume; financial business participation behavior characteristics include the frequency of inquiries about financial products, the duration of viewing financial product information, and the frequency of initiating financial transactions. These user behavior characteristics collectively constitute the user behavior characteristic set of the target domain. While differing in dimensions from the user behavior characteristics of the source domain, they also share the characteristic of being multi-dimensional.

[0019] Risk characteristics are extracted based on the risk types of financial businesses on social media platforms, such as social media fraud inducement risk, false advertising risk of financial products, and illegal transfer of funds risk. Social media fraud inducement risk characteristics include the distribution of the frequency of fraudulent inducement messages and the distribution of the types of financial businesses involved in the fraud; false advertising risk characteristics include the proportion of products falsely advertised and the distribution of the scope of dissemination of false advertising content; illegal transfer of funds risk characteristics involve the distribution of the amount of illegally transferred funds and the distribution of the time periods in which the illegal transfers occurred.

[0020] Similar to the source domain, the extraction of data distribution features describes the distribution of business rule data, user behavior data, and risk performance data in the target domain, forming a set of data distribution features for the target domain.

[0021] Step S113: Construct a domain difference assessment dimension system, which includes rule adaptation dimension, behavior similarity dimension, risk mapping dimension, and distribution offset dimension.

[0022] The rule fit dimension is used to evaluate the degree of fit between business rules in the source domain and the target domain. This dimension focuses on the similarity between the business rules of the two domains in terms of rule structure, rule logic, and applicable scenarios. By comparing various attributes of the rules, the applicability of the business rules in the source domain to the target domain is determined. The higher the fit, the smaller the difference between the business rules of the two domains.

[0023] The behavioral similarity dimension aims to measure the degree of similarity between user behavior in the source and target domains, evaluating aspects such as user behavior patterns, frequency, and preferences. Higher behavioral similarity indicates that users in the two domains are more alike in their behavioral habits, making knowledge transfer relatively easier.

[0024] The risk mapping degree dimension is used to assess the mapping relationship between risk types in the source domain and the target domain, that is, whether there are corresponding risk types in the target domain for risk types in the source domain, and the strength of the correlation between the two. The higher the risk mapping degree, the more similar the risk characteristics of the two domains are, and the easier the transfer of risk knowledge is.

[0025] The distribution offset dimension focuses on the degree of deviation between the data distribution of the source domain and the target domain. Data distribution includes the distribution of business rule data, user behavior data, risk performance data, etc. The smaller the distribution offset, the closer the statistical distribution of the data in the two domains is, and the easier it is to achieve data-level migration and adaptation.

[0026] Step S114: For the rule adaptability dimension, the business rule features of the source domain risk control scenario and the business rule features of the target domain risk control scenario are compared in a structured manner, the difference points of rule elements are extracted, and the rule adaptability difference coefficient is calculated based on the proportion of the number of difference points.

[0027] In assessing the rule compatibility between e-commerce platform transaction risk control scenarios and social platform financial business risk control scenarios, it is necessary to use structured comparison to uncover the differences in the characteristics of the two business rules. The structured comparison must be conducted according to a pre-defined rule element framework to ensure that all key elements are covered, thereby accurately extracting differences and calculating the difference coefficients. Step S1141: Analyze the business rule features of the risk control scenario in the source domain and generate a set of rule elements in the source domain. This set of rule elements in the source domain includes rule condition elements, rule threshold elements, rule execution logic elements, and rule output result elements. When analyzing the business rule characteristics of e-commerce platform transaction risk control scenarios, rule condition elements cover transaction amount range, transaction location attributes, user account status, etc. For example, in the rule "transaction amount greater than a specific value and transaction location is in a different location," both "transaction amount greater than a specific value" and "transaction location is in a different location" are rule condition elements. Rule threshold elements include transaction amount thresholds, transaction frequency thresholds, etc., such as "specific number of transactions" in "daily transaction count exceeds a specific number to trigger review." Rule execution logic elements reflect the relationship between rule conditions and execution actions, such as the logical structure "if condition A and condition B are met, then execute action C." Rule output result elements are the specific results generated after rule execution, such as "freeze transaction" and "send verification SMS." These elements are systematically organized to form a set of source domain rule elements, each element having a clear attribute description and classification identifier. Step S1142: Analyze the business rule features of the risk control scenario in the target domain and generate a set of rule elements in the target domain. This set of rule elements includes rule condition elements, rule threshold elements, rule execution logic elements, and rule output result elements. When analyzing the business rule characteristics of risk control scenarios in social media platform financial transactions, rule condition elements include social relationship intimacy, duration of financial transaction participation, and account activity, such as the conditions in "social friend intimacy is below a certain value and it is the first time participating in financial transactions"; rule threshold elements include transfer amount thresholds and social interaction frequency thresholds; rule execution logic elements include "if condition X or condition Y is true, then execute action Z"; and rule output result elements include "limit transfer amount" and "require supplementary identity information". Similarly, these elements are organized into a target domain rule element set, with element classification consistent with the source domain, for subsequent comparison. Step S1143: Construct a rule element comparison template, which includes element type identifier, element attribute description and element association description. The rule element comparison template identifies the element type corresponding to rule conditions, thresholds, execution logic, and output results, ensuring a one-to-one correspondence between element types during comparison. The element attribute description clearly defines the specific content and characteristics of each element; for example, the attribute description of rule condition elements must explain the judgment criteria for the condition. The element relationship description reflects the logical connection between different elements, such as the relationship between rule condition elements and execution logic elements. This template standardizes the comparison process and reduces omissions and errors. Step S1144: Based on the rule element comparison template, perform an element-by-element comparison operation between the source domain rule element set and the target domain rule element set, recording the element entries with differences as rule element difference points. Based on the rule element comparison template, rule elements in the source and target domains are compared one by one. In the rule condition element comparison, the "transaction location is different" condition for e-commerce platforms and the "social friend intimacy is below a certain value" condition for social platforms are different and are recorded as differences. In the rule threshold element comparison, if the transaction amount threshold for e-commerce platforms differs from the transfer amount threshold for social platforms, this is also recorded as a difference. In the rule execution logic element comparison, the difference between "AND" and "OR" logic is recorded. The difference between "freeze transaction" and "limit transfer amount" in the rule output result elements is also recorded. All differences are clearly labeled with the element type, attribute, and relationship details. Step S1145: Count the total number of difference points of rule elements, calculate the ratio of the total number to the total number of elements in the source domain rule element set, and determine the ratio as the rule fit difference coefficient.

[0028] After counting the total number of rule element differences across all records, this number is compared to the total number of elements in the source domain's rule element set. The resulting ratio is the rule fit difference coefficient. This rule fit difference coefficient objectively reflects the degree of difference between the business rule elements of the two domains.

[0029] Step S115: For the behavioral similarity dimension, perform time-series pattern matching between the user behavior features of the source domain risk control scenario and the user behavior features of the target domain risk control scenario to generate a behavioral sequence similarity score. Subtract the behavioral sequence similarity score from 1 to obtain the behavioral similarity difference coefficient.

[0030] First, the behavioral characteristics of users on the source domain e-commerce platform and the target domain social platform are converted into time-series behavioral sequences. The time-series behavioral sequences in the source domain are based on time, recording the order and intervals of user behaviors such as browsing, adding to cart, placing orders, and making payments. The time-series behavioral sequences in the target domain record the temporal relationships of user behaviors such as social interactions, content publishing, financial business inquiries, and fund operations.

[0031] Then, a temporal pattern matching algorithm is used to match the two temporal behavior sequences. First, the behavior sequences are segmented, dividing behaviors within similar time periods into behavior segments. Then, the similarity between corresponding behavior segments in the source and target domains is calculated. For example, the similarity between a user's browsing-adding to cart-placing an order segment in the source domain on a weekday evening and a user's social interaction-financial business consultation-fund transfer segment in the target domain on a weekday evening is compared, calculated based on factors such as the frequency of behavior occurrence, the probability of transition between behaviors, and the duration of each behavior.

[0032] After calculating the similarity of each behavioral segment, a weighted average is calculated for all segments. The weights are determined based on the importance of the behavioral segment in the entire time series; for example, behavioral segments involving financial operations have higher weights. The resulting weighted average is the behavioral sequence similarity score. The higher the score, the more similar the time series patterns of user behavior in the two domains are.

[0033] Finally, subtracting the behavioral sequence similarity score from 1 yields the behavioral similarity difference coefficient, which reflects the degree of difference in user behavior between the two domains.

[0034] Step S116: For the risk mapping degree dimension, construct a risk type mapping dictionary, statistically analyze the overlap ratio and semantic association strength of risk types in the source domain risk control scenario and the target domain risk control scenario, and calculate the risk mapping degree difference coefficient based on the weighted average.

[0035] When constructing a risk type mapping dictionary, we first sort out all risk types in the transaction risk control scenarios of e-commerce platforms in the source domain, such as fraudulent transaction risk, fake return risk, and account theft risk. Then, we sort out the risk types in the financial business risk control scenarios of social platforms in the target domain, such as social fraud inducement risk, false advertising risk of financial products, and illegal transfer of funds risk.

[0036] Then, the relationships between these risk types are analyzed, mapping risk types with similar nature or causal connections. For example, fraudulent transaction risk in the source domain is related to social fraud inducement risk in the target domain because both involve fraudulent behavior; account theft risk in the source domain is associated with the risk of illegal fund transfer due to account theft in the target domain. These mapping relationships are recorded in a risk type mapping dictionary, where each mapping entry includes a risk type in the source domain, a risk type in the target domain, and a correlation strength score. The correlation strength score is determined based on the similarity between the risk types and the tightness of the causal relationship.

[0037] When calculating the overlap ratio of risk types, the ratio is obtained by dividing the number of completely identical or essentially identical risk types in the two domains by the total number of risk types in the source domain. For example, if the source domain has 5 main risk types, and 2 of them completely overlap with the risk types in the target domain, then the overlap ratio is 2 / 5.

[0038] The semantic association strength is represented by the average association strength score of all mapping entries in the risk type mapping dictionary. The higher the average score, the stronger the semantic association between the two domain risk types.

[0039] Finally, a weighted average of the overlap ratio and semantic association strength is calculated, with the weights set according to their importance in the risk mapping assessment, to obtain the risk mapping score. Subtracting this risk mapping score from 1 yields the risk mapping difference coefficient.

[0040] Step S117: For the distribution offset dimension, the maximum mean difference algorithm is used to calculate the distribution distance between the data distribution characteristics of the source domain risk control scenario and the data distribution characteristics of the target domain risk control scenario. The distribution distance is then standardized and used as the distribution offset difference coefficient.

[0041] First, determine the data distribution characteristics that need to be compared in the source and target domains, such as the frequency distribution of various behaviors in user behavior characteristics and the probability distribution of risk occurrence in risk performance characteristics.

[0042] Then, the maximum mean difference algorithm is applied to calculate the distance between these data distribution features. This algorithm maps the data to a reproducing kernel Hilbert space and calculates the mean difference between the two domain data in that space to measure the degree of shift between the distributions. For example, for the source domain user browsing time distribution and the target domain user social interaction time distribution, the distribution distance between the two is calculated. The larger the distribution distance, the greater the shift between the two distributions.

[0043] After obtaining the distribution distance, it is standardized to convert it to the range of 0 to 1. The standardization method involves dividing the distribution distance by the maximum possible distribution distance, ensuring that the processed result can be compared with the difference coefficients of other dimensions on the same scale. The standardized result is the distribution offset difference coefficient.

[0044] Step S118: Arrange the rule fit difference coefficient, behavior similarity difference coefficient, risk mapping difference coefficient, and distribution offset difference coefficient in row vector form to generate a domain difference measurement matrix.

[0045] The four difference coefficients calculated earlier are arranged into a row vector in the order of rule fit difference coefficient, behavior similarity difference coefficient, risk mapping difference coefficient, and distribution offset difference coefficient. For example, if the rule fit difference coefficient is A, the behavior similarity difference coefficient is B, the risk mapping difference coefficient is C, and the distribution offset difference coefficient is D, then the vector is represented as [A, B, C, D]. This row vector is the domain difference measurement matrix, which integrates the differences between the source domain and the target domain across the four evaluation dimensions.

[0046] Step S120: Construct a hierarchical knowledge transfer channel based on the domain difference metric matrix, perform selective transfer processing on the multi-level network parameters of the source domain risk prediction model, and generate the target domain initial risk prediction model.

[0047] The purpose of this step is to utilize the differences between the two domains reflected by the domain difference metric matrix to construct a targeted knowledge transfer channel, select network parameters suitable for transfer to the target domain from the source domain risk prediction model, and perform corresponding transformation processing to initialize the target domain risk prediction model.

[0048] Step S121: Analyze the network architecture of the source domain risk prediction model and determine the multi-level network structure that includes an input layer, a hidden layer, and an output layer. The hidden layer includes a feature extraction sub-layer, a feature fusion sub-layer, and a risk assessment sub-layer. Each network structure corresponds to a set of network parameters.

[0049] This paper analyzes the network architecture of a risk prediction model for a transaction risk control scenario on an e-commerce platform in the source domain. The model adopts a deep learning network architecture. The input layer is responsible for receiving raw user feature data and transaction feature data. Its network parameters include the weight matrix and bias vector of the input features, which determine the initial importance of different input features in the model.

[0050] The feature extraction sublayer in the hidden layer consists of multiple convolutional layers and pooling layers. The network parameters of the convolutional layers include the kernel weights and biases, which are used to extract local features from the input data. The parameters of the pooling layers determine the method and range of the pooling operation, such as max pooling or average pooling, as well as the size of the pooling window. The feature extraction sublayer processes the input data using these parameters to obtain a preliminary feature representation.

[0051] The feature fusion sublayer employs a fully connected layer structure. Its network parameters include connection weight matrices and bias vectors, which are used to fuse multiple local features obtained from the feature extraction sublayer to form a more representative comprehensive feature. The parameters of this feature extraction sublayer determine the degree of contribution of different local features in the fusion process.

[0052] The risk assessment sublayer is also a fully connected layer. Its network parameters include a weight matrix and a bias vector, which are used to process the fused features and output intermediate risk assessment results. These parameters directly affect the assessment of the correlation between different features and risks.

[0053] The output layer is responsible for converting the intermediate results of the risk assessment sublayer into the final risk prediction probability distribution. Its network parameters include the output weight matrix and bias vector, which determine the mapping relationship between the intermediate results and the final prediction results.

[0054] Step S122: Calculate the domain migration weights of network parameters at each level based on the domain difference metric matrix. The domain migration weights are negatively correlated with the difference coefficients of the corresponding levels.

[0055] Step S1221: Extract the difference coefficients of the rule fit dimension in the domain difference metric matrix, and determine the rule fit difference coefficients as the difference evaluation benchmarks for the input layer network parameters.

[0056] The rule fit difference coefficient in the domain difference measurement matrix reflects the differences in business rules between the source domain and the target domain. Since the input layer network parameters are mainly related to the input features involved in the business rules, this coefficient is used as the difference evaluation benchmark for the input layer network parameters to measure the degree of difference faced by the input layer parameters when migrating across domains.

[0057] Step S1222: Extract the difference coefficients of the behavior similarity dimension and the risk mapping dimension from the domain difference measurement matrix, and calculate the arithmetic mean of the two as the benchmark for evaluating the difference of hidden layer network parameters.

[0058] The feature extraction, fusion, and risk assessment processes in the hidden layer are related to both user behavior characteristics and risk performance characteristics. Therefore, both the behavior similarity difference coefficient and the risk mapping difference coefficient are considered. By calculating the arithmetic mean of these two coefficients, an evaluation benchmark that comprehensively reflects the differences in the feature domains processed by the hidden layer is obtained, which is used to evaluate the transfer adaptability of the hidden layer network parameters.

[0059] Step S1223: Extract the difference coefficients of the distribution offset dimension in the domain difference metric matrix, and determine the distribution offset difference coefficients as the difference evaluation benchmarks for the output layer network parameters.

[0060] The network parameters of the output layer determine the distribution of the risk prediction results of the final output of the model, and the distribution offset difference coefficient reflects the offset of the data distribution between the two domains. Therefore, it is used as the benchmark for evaluating the difference of the output layer network parameters to measure the degree of fit between the output layer parameters and the target domain data distribution when transferring across domains.

[0061] Step S1224: Construct a difference evaluation benchmark normalization function. Input the difference evaluation benchmarks of the input layer, hidden layer and output layer into the difference evaluation benchmark normalization function to generate the relative difference values ​​of each layer. The relative difference values ​​range from 0 to 1.

[0062] The difference assessment benchmark normalization function employs the min-max normalization method. This method first determines the minimum and maximum values ​​among all difference assessment benchmarks. Then, it subtracts the minimum value from the difference assessment benchmark for each level and divides by the difference between the maximum and minimum values ​​to obtain the relative difference value for that level. For example, if the difference assessment benchmark for the input layer is E, for the hidden layer it is F, and for the output layer it is G, the calculated relative difference values ​​are E_norm for the input layer, F_norm for the hidden layer, and G_norm for the output layer. These values ​​are all within the range of 0 to 1. For instance, when the difference assessment benchmark E of the input layer is at a moderate level among all benchmarks, after normalization, E_norm will be in the middle position between 0 and 1, intuitively reflecting the relative degree of the input layer's cross-domain difference.

[0063] Step S1225: Subtract the relative difference value of each level from 1 to obtain the domain transfer weight of each level of network parameters. The value range of the domain transfer weight is 0 to 1.

[0064] After obtaining the relative difference values ​​E_norm for the input layer, F_norm for the hidden layer, and G_norm for the output layer, subtracting these values ​​from 1 yields the corresponding domain transfer weights. Specifically, the domain transfer weight for the input layer is 1 minus E_norm, for the hidden layer it's 1 minus F_norm, and for the output layer it's 1 minus G_norm. Since the relative difference values ​​range from 0 to 1, the domain transfer weights must also range from 0 to 1. A higher domain transfer weight indicates that the network parameters at that layer are more suitable for migration from the source domain to the target domain, because a higher weight means that the domain difference at that layer is smaller, and the migrated parameters are better suited to the needs of the target domain. For example, if the relative difference value E_norm for the input layer is small, it means that the domain difference of the input layer is small, so the domain transfer weight obtained by subtracting E_norm from 1 will be larger, indicating that the network parameters of the input layer have higher migration value.

[0065] Step S123: Construct a hierarchical filtering mechanism, which includes a migration weight threshold judgment module and a parameter retention decision module. The migration weight threshold judgment module is used to compare the domain migration weight of each level of network parameters with a preset threshold. The parameter retention decision module decides whether to retain the corresponding level of network parameters based on the comparison results.

[0066] The construction of a hierarchical screening mechanism first requires determining a preset threshold. This preset threshold is set based on experience in domain transfer and the characteristics of the target domain, and is used to determine whether the parameters of the hierarchical network have transfer value. The core function of the transfer weight threshold judgment module is to compare the domain transfer weights calculated for each layer with the preset threshold one by one. For example, the domain transfer weights of the input layer, the hidden layer, and the output layer are compared with the preset thresholds respectively to determine whether the weight of each layer is higher or lower than the preset threshold.

[0067] The parameter retention decision module makes its decision based on the comparison results of the transfer weight threshold judgment module. When the domain transfer weight of a certain level is higher than the preset threshold, the parameter retention decision module determines that the network parameters at that level have high transfer value and decides to retain these parameters; when the domain transfer weight of a certain level is lower than or equal to the preset threshold, the module determines that the network parameters at that level have low transfer value and decides not to retain these parameters. This mechanism ensures that only those network parameters with small differences from the target domain and high adaptability are selected for transfer, improving the efficiency and accuracy of knowledge transfer.

[0068] Step S124: Perform selective filtering on multi-level network parameters through a hierarchical filtering mechanism, and retain the hierarchical network parameters with transfer weights higher than a preset threshold as transferable knowledge parameters.

[0069] The network parameters of each layer of the source domain risk prediction model are input into a layer-level screening mechanism. The migration weight threshold judgment module first compares the domain migration weights of the input layer network parameters with a preset threshold. If the domain migration weights of the input layer are higher than the preset threshold, the network parameters of that layer, including the weight matrix and bias vector of the input features, are initially determined to be parameters that can be retained; if they are lower than or equal to the preset threshold, they are initially determined to be parameters that cannot be retained.

[0070] Next, the network parameters of each sub-layer of the hidden layer are screened. The convolutional kernel weights, biases, and pooling layer parameters of the feature extraction sub-layer, the connection weight matrix and bias vector of the feature fusion sub-layer, and the weight matrix and bias vector of the risk assessment sub-layer are judged based on the comparison results of the domain transfer weights of their respective hidden layers with preset thresholds. If the domain transfer weights of the hidden layers are higher than the preset threshold, the network parameters of these sub-layers are included in the range of retainable parameters; otherwise, they are excluded.

[0071] Finally, the network parameters of the output layer are filtered, and the output weight matrix and bias vector of the output layer are determined whether to be retained based on the comparison results between the domain transfer weight of the output layer and the preset threshold.

[0072] After selective filtering by a hierarchical screening mechanism, all hierarchical network parameters deemed retainable are aggregated to form a set of transferable knowledge parameters. These transferable knowledge parameters represent the most suitable parts of the source domain risk prediction model for transfer to the target domain.

[0073] Step S125: Construct a hierarchical knowledge transfer channel, which includes an input layer transfer sub-channel, a hidden layer transfer sub-channel, and an output layer transfer sub-channel. Each transfer sub-channel corresponds to a different parameter transformation rule.

[0074] The parameter transformation rules for the input layer transfer sub-channel are primarily designed for the network parameters of the input layer. Considering the differences in input features between the source and target domains, these rules include two parts: feature dimension mapping and weight adjustment. Feature dimension mapping maps the feature dimensions corresponding to the input layer parameters in the source domain to the input feature dimensions in the target domain. For example, a feature dimension related to e-commerce transaction amounts in the source domain input layer might correspond to a feature dimension related to the amount of funds in financial transactions in the target domain. This mapping ensures that the parameters can adapt to the input structure of the target domain. Weight adjustment, on the other hand, fine-tunes the weight matrix and bias vector of the source domain input layer based on the importance distribution of the input features in the target domain, making the transferred parameters more consistent with the characteristics of the input data in the target domain.

[0075] The parameter transformation rules for the hidden layer transfer sub-channels are more complex because the hidden layer contains multiple sub-layers. For the feature extraction sub-layer, the parameter transformation rules involve adjusting the size of the convolutional kernels and redistributing weights. When the feature extraction requirements of the target domain differ from those of the source domain, the size of the convolutional kernels needs to be adjusted to adapt to the new feature scale, and the weights of the convolutional kernels need to be redistributed according to the importance of the features in the target domain. The parameter transformation rules for the feature fusion sub-layer mainly adjust the connection weight matrix to adapt to the correlation between different features in the target domain. For example, the correlation strength between social interaction features and financial behavior features in the target domain may differ from the correlation strength between product browsing features and transaction features in the source domain. The weight matrix is ​​adjusted to reflect this difference. The parameter transformation rules for the risk assessment sub-layer focus on adjusting the weight matrix and bias vector to accurately capture the mapping relationship between features and risks in the target domain.

[0076] The parameter transformation rules for the output layer migration sub-channel primarily target the network parameters of the output layer. The aim is to ensure that the migrated parameters produce results that conform to the risk assessment standards of the target domain. These parameter transformation rules include output dimension adaptation and probability distribution adjustment. Output dimension adaptation ensures that the migrated parameters correspond to the number of risk types in the target domain. Probability distribution adjustment adjusts the weight matrix and bias vector of the output layer based on the overall probability distribution of risk occurrence in the target domain, making the output risk probability distribution more closely match the actual situation in the target domain.

[0077] Step S126: Input the transferable knowledge parameters into the corresponding transfer sub-channel, perform cross-domain parameter mapping and transformation processing, and generate target domain hierarchical parameters that are adapted to the target domain network structure.

[0078] The input layer network parameters from the transferable knowledge parameters are input into the input layer transfer sub-channel and processed according to the parameter transformation rules of this sub-channel. First, feature dimension mapping is performed, mapping the feature dimensions of the source domain input layer parameters to the input feature dimensions of the target domain, generating preliminary dimension mapping parameters. Then, based on the importance distribution of the target domain input features, the preliminary weight matrix and bias vector are adjusted to obtain target domain input layer parameters adapted to the target domain input layer structure.

[0079] The hidden layer network parameters in the transferable knowledge parameters are input into the hidden layer transfer sub-channels and processed according to the transformation rules corresponding to each sub-layer. The parameters of the feature extraction sub-layer are first adjusted in the sub-channel by adjusting the convolution kernel size to adapt to the feature extraction requirements of the target domain. Then, the convolution kernel weights and biases are redistributed according to the distribution of target domain features to generate the target domain parameters for the feature extraction sub-layer. The connection weight matrix of the feature fusion sub-layer is adjusted in the sub-channel according to the correlation between target domain features to obtain a weight matrix and bias vector suitable for target domain feature fusion. The parameters of the risk assessment sub-layer are generated by adjusting the weight matrix and bias vector to accurately assess the correlation between target domain features and risk. These target domain parameters of the sub-layers collectively constitute the target domain hidden layer parameters.

[0080] The output layer network parameters in the transferable knowledge parameters are input to the output layer transfer sub-channel. After output dimension adaptation processing, it is ensured that the parameters can correspond to the number of risk types in the target domain. Then, through probability distribution adjustment, the weight matrix and bias vector of the output layer generate results that conform to the risk probability distribution of the target domain, and finally the output layer parameters of the target domain are obtained.

[0081] Step S127: Load the target domain hierarchical parameters into the target domain network framework in hierarchical order to generate the target domain initial risk prediction model.

[0082] The target domain network framework is designed based on the needs of risk control scenarios in the financial business of social platforms in the target domain. Its structure is similar to the network architecture of the source domain risk prediction model, including an input layer, a hidden layer, and an output layer. The hidden layer also includes a feature extraction sub-layer, a feature fusion sub-layer, and a risk assessment sub-layer, but the specific configuration of each layer has been optimized for the target domain.

[0083] When loading parameters in hierarchical order, the target domain input layer parameters are first loaded into the input layer of the target domain network framework to ensure that the input layer can correctly receive and process the target domain input data. Next, the target domain hidden layer parameters are loaded into the corresponding positions of the hidden layers in the order of feature extraction sublayer, feature fusion sublayer, and risk assessment sublayer, enabling the hidden layers to perform feature extraction, fusion, and risk assessment operations that meet the requirements of the target domain. Finally, the target domain output layer parameters are loaded into the output layer to ensure that the output layer can output the risk prediction results required by the target domain.

[0084] Once the parameters are loaded, the target domain network framework becomes the target domain initial risk prediction model. It inherits the transferable knowledge from the source domain risk prediction model and adapts to the characteristics of the target domain.

[0085] Step S130: Use the target domain risk sample set to perform dynamic adaptation training on the target domain initial risk prediction model to generate a cross-domain risk prediction model.

[0086] The core of this step is to train the initialized target domain risk prediction model using risk sample data from the target domain. The model parameters are continuously adjusted to better adapt the model to the risk characteristics and data distribution of the target domain, ultimately generating a cross-domain risk prediction model capable of accurately predicting risks in the target domain. During dynamic adaptation training, the training strategy can be dynamically adjusted based on the training results to improve the model's performance and generalization ability.

[0087] Step S131: Divide the target domain risk sample set into an adaptation training sample subset, a validation sample subset, and a test sample subset. The proportion of the adaptation training sample subset to the total number of samples is greater than that of the validation sample subset and the test sample subset. The adaptation training sample subset is used for model parameter adjustment, the validation sample subset is used for training process monitoring, and the test sample subset is used for final performance evaluation.

[0088] The target domain risk sample set contains a large amount of risk case data from social platform financial business. Each sample includes the user's social characteristics, financial behavior characteristics and corresponding risk tags, such as whether there is social fraud inducement risk, whether there is illegal transfer of funds risk, etc.

[0089] When dividing the sample into subsets, the proportion of each subset is first determined. The proportion of the training sample subset to the total number of samples is usually set to a large proportion, such as more than half, to ensure that there are enough samples for adjusting and optimizing the model parameters. The proportions of the validation sample subset and the test sample subset are relatively small, and the proportions of the two can be the same or slightly different.

[0090] The partitioning process employs random sampling to ensure that each subset represents the overall distribution of risk samples in the target domain. During random sampling, all samples are first shuffled, and then samples are selected sequentially according to preset proportions to form the training, validation, and test subsets. For example, from the shuffled samples, a larger proportion is first selected as the training subset, then a certain proportion is selected as the validation subset, and the remaining samples are used as the test subset.

[0091] The adaptation training sample subset is used to iteratively update model parameters during model training. By calculating the difference between the predicted results and the true risk labels, the model's weights and biases are continuously adjusted. The validation sample subset is input into the model after each or multiple training rounds. Metrics such as validation accuracy are used to monitor the model's training progress and overfitting. The test sample subset is used to evaluate the model's final performance after training is complete, determining whether the model has achieved the expected predictive effect.

[0092] Step S132: Construct a dynamic loss function based on the domain difference metric matrix. This dynamic loss function includes a basic loss term and a domain adaptation loss term. The basic loss term uses the cross-entropy loss function to calculate the difference between the predicted risk type and the actual risk type. The domain adaptation loss term adjusts the loss value of different risk types based on the domain difference coefficient.

[0093] Step S1321: Extract the distribution offset dimension difference coefficient from the domain difference metric matrix, and determine the distribution offset dimension difference coefficient as the basic adjustment factor for the domain adaptation loss term.

[0094] The distribution offset dimension difference coefficient reflects the overall degree of offset between the data distributions of the source and target domains. The purpose of the domain adaptation loss term is to reduce the loss bias caused by data distribution differences during model training in the target domain. Therefore, using this coefficient as a basic adjustment factor is reasonable. The larger the value of this basic adjustment factor, the greater the difference in data distribution between the two domains, and the stronger the role of the domain adaptation loss term in the dynamic loss function needs to be to better guide the model to adapt to the data distribution of the target domain.

[0095] Step S1322: Construct an iterative decay coefficient generation function. The input of this iterative decay coefficient generation function is the number of training iterations, and the output is the iterative decay coefficient that gradually decays from 1 to 0 as the number of iterations increases.

[0096] The design of the iterative decay coefficient generation function is based on the model training pattern. In the early stages of training, the model has limited knowledge of the target domain and requires strong domain adaptation guidance, resulting in a larger iterative decay coefficient. As the number of training iterations increases, the model gradually adapts to the characteristics of the target domain, reducing the need for domain adaptation guidance and causing the iterative decay coefficient to gradually decrease. This iterative decay coefficient generation function can adopt an exponential decay form. For example, the number of iterations can be used as the independent variable, and the decay coefficient can be calculated using an exponential function, so that the coefficient decreases proportionally with each additional iteration, eventually approaching 0. For instance, when the number of iterations is 0, the decay coefficient is 1; as the number of iterations increases, the decay coefficient gradually decreases, and when the number of iterations reaches the preset maximum number of iterations, the decay coefficient approaches 0.

[0097] Step S1323: Perform a product operation between the basic adjustment factor and the iterative decay coefficient to generate the domain adaptation loss weight coefficient.

[0098] The base adjustment factor reflects the static degree of domain distribution differences, while the iteration decay coefficient reflects the dynamic changes in domain adaptation requirements during training. Multiplying these two factors yields the domain adaptation loss weight coefficient, which considers both the inherent characteristics of domain differences and the changing domain adaptation requirements during training. In the early stages of training, the iteration decay coefficient is 1, and the domain adaptation loss weight coefficient equals the base adjustment factor. At this point, the domain adaptation loss term has a relatively large weight, effectively guiding the model to adapt to the data distribution of the target domain. As the number of iterations increases, the iteration decay coefficient gradually decreases, and the domain adaptation loss weight coefficient also decreases accordingly. This causes the proportion of the base loss term in the dynamic loss function to gradually increase, and the model focuses more on the prediction accuracy of the target domain samples themselves.

[0099] Step S1324: Construct a risk type weight vector. Each element of the risk type weight vector corresponds to a risk type, and the element value is inversely proportional to the frequency of occurrence of that risk type in the target domain.

[0100] First, all risk types in the risk control scenarios of financial business on social media platforms in the target domain are identified, such as social fraud inducement risk, false advertising risk of financial products, and illegal transfer of funds risk. A corresponding vector element position is assigned to each risk type. Then, the frequency of occurrence of each risk type in the risk sample set of the target domain is calculated. The frequency is the ratio of the number of samples of that risk type to the total number of samples. For risk types with low occurrence frequency, higher weights are assigned to improve the model's ability to identify them; therefore, the element value is inversely proportional to the frequency. For example, if the frequency of social fraud inducement risk is low, the corresponding element value is larger; if the frequency of false advertising risk of financial products is high, the corresponding element value is smaller. The element values ​​of the risk type weight vector are normalized so that their sum is 1 to ensure the rationality of the weights.

[0101] Step S1325: Introduce the risk type weight vector into the basic loss term, and perform a weighted summation of the cross-entropy loss values ​​for different risk types to generate a weighted basic loss term.

[0102] The basic loss term is calculated using the cross-entropy loss function. For each sample, the cross-entropy loss value is calculated based on the probability distribution of the risk type predicted by the model and the true risk label. Then, the cross-entropy loss value for each risk type is multiplied by the corresponding element in the risk type weight vector to obtain the weighted loss value for that risk type. Finally, the weighted loss values ​​for all risk types are summed to obtain the weighted basic loss term. This process allows the model to focus more on low-frequency but important risk types during training, improving the model's overall predictive ability for various risks.

[0103] Step S1326: Perform a product operation between the domain adaptation loss weight coefficient and the weighted basic loss term to generate the domain adaptation loss term.

[0104] The domain adaptation loss weight coefficient determines the degree of influence of the domain adaptation loss term in the dynamic loss function. Multiplying it by the weighted basic loss term yields a domain adaptation loss term that incorporates information from both the basic loss and the adjustment factors related to domain adaptation. When the domain adaptation loss weight coefficient is large, the domain adaptation loss term contributes significantly to the dynamic loss function, and model training focuses more on reducing the loss caused by domain differences. Conversely, when the coefficient is small, the contribution of the domain adaptation loss term decreases, and the model focuses more on reducing the basic loss.

[0105] Step S1327: Construct a dynamic loss function expression, which is the arithmetic sum of the weighted basic loss term and the domain adaptation loss term, and determine this dynamic loss function expression as the dynamic loss function.

[0106] The final form of the dynamic loss function is the sum of the weighted base loss term and the domain adaptation loss term, thus combining the accuracy loss of the model's prediction of samples in the target domain with the adaptation loss caused by domain differences. During training, the model adjusts its parameters by minimizing this dynamic loss function, enabling the model to accurately predict the risk type in the target domain while also adapting well to the differences between the source and target domains, thereby improving the model's cross-domain generalization ability.

[0107] Step S133: Initialize the weight coefficients of the dynamic loss function. These weight coefficients include the base loss weight and the domain adaptation loss weight. In the initial state, the base loss weight is greater than the domain adaptation loss weight.

[0108] The base loss weights and domain adaptation loss weights are used to adjust the relative importance of the weighted base loss term and the domain adaptation loss term in the dynamic loss function. Initially, the model needs to learn the basic risk prediction patterns of the target domain first; therefore, the base loss weight is set to a larger value, and the domain adaptation loss weight is set to a smaller value. For example, the base loss weight can be set to 0.7, and the domain adaptation loss weight to 0.3. This allows the weighted base loss term to dominate the dynamic loss function, ensuring that the model prioritizes accurate predictions of samples from the target domain.

[0109] Step S134: Divide the subset of adaptive training samples into multiple training batches according to the preset batch size, and input each training batch into the target domain to initialize the risk prediction model in sequence.

[0110] The preset batch size is determined based on the hardware resources and sample size for model training. The choice of batch size affects training efficiency and model convergence. Setting the batch size too small increases the number of training iterations, prolonging training time, and may lead to overly frequent model parameter updates, resulting in training instability. Setting the batch size too large consumes more memory resources and may cause the model to converge to a local optimum. In this embodiment, considering both the total amount of risk samples in the target domain and the memory capacity of the training device, the preset batch size is determined to be an appropriate value, ensuring that each batch of samples can be efficiently processed by the training device.

[0111] When dividing the training batches, the subset of suitable training samples is first randomly shuffled to avoid the order of samples affecting model training. Then, samples are sequentially extracted from the shuffled subset according to a preset batch size to form multiple training batches. For example, if the number of samples in the suitable training sample subset is the total number of suitable training samples, and the preset batch size is a set value, then the number of training batches is the result obtained by dividing the total number of suitable training samples by the preset batch size and rounding up.

[0112] When each training batch is sequentially input into the target domain to initialize the risk prediction model, the model's input layer receives sample feature data from the training batch. After processing by the input layer, this data is passed to the feature extraction sublayer of the hidden layer. The feature extraction sublayer performs feature extraction on the input feature data and passes the extracted local features to the feature fusion sublayer. The feature fusion sublayer fuses the local features to obtain a comprehensive feature, which is then passed to the risk assessment sublayer. The risk assessment sublayer performs risk assessment on the comprehensive feature, generating an intermediate risk assessment result. Finally, the output layer converts the intermediate result into the corresponding risk prediction value.

[0113] Step S135: Calculate the predicted risk value for each training batch through forward propagation, and calculate the batch loss value by combining the sample's true risk label with the dynamic loss function.

[0114] After initializing the risk prediction model with input from the target domain in each training batch, the model undergoes a forward propagation process. The input layer performs preliminary processing on the feature data of the training samples, assigning different initial weights, and then passes it to the feature extraction sublayer of the hidden layer. The feature extraction sublayer extracts representative local features from the input features through convolution and pooling operations. These local features reflect local information related to risk in the samples.

[0115] The local features processed by the feature extraction sublayer are passed to the feature fusion sublayer. The feature fusion sublayer fuses multiple local features through fully connected operations, integrating the information contained in different local features to form a comprehensive feature that fully reflects the risk characteristics of the sample. The comprehensive feature is then passed to the risk assessment sublayer, which further processes and analyzes the comprehensive feature. Through multi-layer fully connected operations, the comprehensive feature is mapped onto the risk assessment space, generating intermediate risk assessment results.

[0116] The intermediate results are passed to the output layer, which processes them using an activation function to obtain the predicted risk value for each sample. The predicted risk value is a multi-dimensional vector, with each dimension corresponding to the predicted probability of a risk type.

[0117] For each sample in the training batch, there is a corresponding true risk label. The true risk label is also a multi-dimensional vector. Each dimension indicates whether the sample belongs to the corresponding risk type. If it does, the value of that dimension is 1, otherwise it is 0.

[0118] The predicted risk value and the true risk label of each sample are input into a dynamic loss function. The dynamic loss function calculates the loss value for each sample based on the difference between the two. Then, the loss values ​​of all samples in the batch are averaged to obtain the batch loss value for that training batch. The batch loss value reflects the gap between the model's prediction results and the actual situation in the current training batch. The larger the loss value, the worse the model's prediction performance.

[0119] Step S136: Based on the batch loss value, the backpropagation algorithm is used to update the network parameters of the target domain initial risk prediction model. During the update process, the momentum optimization strategy is used to accelerate convergence.

[0120] The core of the backpropagation algorithm is to calculate the gradients of the network parameters at each layer of the model based on the batch loss value, and then update the network parameters in the direction of gradient descent. Specifically, it first calculates the gradient of the network parameters of the output layer with respect to the batch loss value, which reflects the degree of influence of changes in the output layer parameters on the loss value; then, it calculates the gradient of the network parameters of the risk assessment sublayer based on the gradient of the output layer, and so on, progressively calculating the gradients of the network parameters of the feature fusion sublayer, feature extraction sublayer, and input layer.

[0121] After calculating the gradients of the network parameters at each layer, a momentum optimization strategy is used to update the network parameters. This momentum optimization strategy simulates the concept of momentum in physics, considering not only the current gradient but also the direction and magnitude of the previous parameter update during parameter updates. Specifically, the update amount for each network parameter consists of two parts: the product of the current gradient and the learning rate, and the product of the previous parameter update amount and the momentum coefficient. These two parts are then added together to obtain the current parameter update amount.

[0122] The learning rate is a pre-defined parameter that determines the step size for parameter updates. An excessively large learning rate may cause the parameters to oscillate around the optimal solution, while an excessively small learning rate will make the training process too slow. The momentum coefficient is also a pre-defined parameter, ranging from 0 to 1. A larger momentum coefficient indicates a greater influence of the direction of the previous parameter update on the current update, helping to accelerate the model's convergence in the correct direction and reduce oscillations.

[0123] By using the backpropagation algorithm and momentum optimization strategy, the network parameters of the input layer, hidden layer (including feature extraction sub-layer, feature fusion sub-layer, and risk assessment sub-layer), and output layer of the initial risk prediction model in the target domain are updated, so that the prediction results of the model gradually approach the true risk label of the sample and reduce the batch loss value.

[0124] Step S137: After completing a preset number of iterations of training, input the validation sample subset into the current model, calculate the validation set prediction accuracy, and when the improvement of the validation set prediction accuracy after K consecutive iterations is lower than a preset threshold, dynamically adjust the weight coefficients of the dynamic loss function and increase the proportion of the domain adaptation loss weight.

[0125] Preset-number of iterative training refers to the process of training the model once for all training batches. After completing the preset number of iterative training iterations, the model's parameters have been updated multiple times. At this point, a subset of validation samples is input into the model that has undergone multiple parameter updates. The model then performs risk prediction on the validation samples, obtaining a predicted risk label for each validation sample.

[0126] To calculate the validation set prediction accuracy, the predicted risk label of each validation sample is compared with the true risk label of the sample. The number of correctly predicted samples is counted, and then the number of correctly predicted samples is divided by the total number of samples in the validation sample subset to obtain the validation set prediction accuracy. The validation set prediction accuracy reflects the model's generalization ability on validation samples that were not used in training. The higher the accuracy, the stronger the model's generalization ability.

[0127] K consecutive iterations refer to K consecutive complete training iterations. After each iteration, the validation set prediction accuracy is calculated. The difference between the validation set prediction accuracy of each iteration and the previous iteration is calculated to obtain the accuracy improvement. When the accuracy improvement of K consecutive iterations is lower than a preset threshold, it indicates that the model, under the current loss function settings, is unlikely to further improve its generalization ability, and the weight coefficients of the dynamic loss function need to be adjusted.

[0128] When dynamically adjusting the weight coefficients, the sum of the base loss weights and the domain adaptation loss weights remains constant. The proportion of the domain adaptation loss weights is increased, while the proportion of the base loss weights is decreased. For example, if the original ratio of base loss weights to domain adaptation loss weights was the base ratio, the adjusted ratio will be a new one, with the proportion of domain adaptation loss weights increasing. This adjustment makes the model pay more attention to the differences between domains during training, improving the model's adaptability to the target domain.

[0129] Step S138: When the prediction accuracy of the validation set remains stable for a preset number of consecutive times or reaches a preset target value, stop training, input a subset of test samples into the current model to perform performance evaluation, generate a model performance evaluation report, and determine the model that passes the performance evaluation as a cross-domain risk prediction model based on the model performance evaluation report.

[0130] Maintaining stable validation set prediction accuracy for a preset number of consecutive training iterations means that the variation in validation set prediction accuracy remains within a very small range, with almost no significant increase or decrease, during a preset number of training iterations. The preset target value is a threshold for validation set prediction accuracy pre-set based on the requirements of the risk control scenario in the target domain. When the validation set prediction accuracy reaches this threshold, it indicates that the model's performance has met the requirements.

[0131] When the above-mentioned conditions for stopping training are met, training of the initial risk prediction model for the target domain is stopped. At this point, a subset of test samples is input into the currently trained model, and the model performs risk prediction on the test samples to obtain the predicted risk value for each test sample.

[0132] When performing performance evaluation, in addition to calculating the prediction accuracy on the test set, other performance metrics are also calculated, such as precision, recall, and F1 score. Precision refers to the proportion of samples predicted as positive that were actually positive, reflecting the accuracy of the model's prediction results; recall refers to the proportion of samples that were actually positive that were predicted as positive, reflecting the model's ability to identify positive samples; the F1 score is the harmonic mean of precision and recall, comprehensively reflecting the overall performance of the model.

[0133] Based on these performance metrics, a model performance evaluation report is generated. The report details the values ​​of each metric and the model's performance on a subset of test samples. If all performance metrics of the model meet the preset evaluation criteria, the model is identified as a cross-domain risk prediction model through performance evaluation. If the evaluation criteria are not met, it is necessary to re-examine the parameter settings and other issues in the model training process, make adjustments, and retrain.

[0134] Step S140: Input the target domain sample to be evaluated into the cross-domain risk prediction model, perform risk feature transfer reasoning operation, and generate preliminary risk prediction scores.

[0135] Step S141: Perform risk feature extraction on the sample to be evaluated in the target domain to extract the basic risk feature set of the sample. The basic risk feature set includes user attribute features, behavioral transaction features, historical risk features and related entity features.

[0136] When extracting risk features from samples to be evaluated in the target domain, we first collect various types of data contained in the samples to be evaluated, and then extract risk-related features from these data.

[0137] User attribute characteristics include user registration information characteristics and identity information characteristics. Registration information characteristics include the user's registration duration, the type of device used during registration, and the type of region where the registration took place; identity information characteristics involve the user's age range, gender, and occupation. These characteristics reflect some inherent features of the user from the perspective of basic attributes and may be related to the user's risky behavior.

[0138] Behavioral transaction characteristics encompass users' social and financial transaction behaviors on social platforms. Social behavior characteristics include the number of friends a user has, the frequency of friend interactions, and the type and frequency of content posted; financial transaction behavior characteristics include the types of financial transactions a user participates in, the range of transaction amounts, the frequency of transactions, and the time of day for transactions. These characteristics reflect users' behavioral habits in social interactions and financial transactions and are important bases for risk assessment.

[0139] Historical risk features are extracted based on a user's past risk performance, including the type, frequency, and time interval of risk events that occurred in the user's history, as well as records of penalties imposed on the user for these risk events. These features reflect a user's risk history and are valuable for predicting the risk of the current sample.

[0140] Related entity characteristics involve the features of other entities associated with the user being evaluated, such as the proportion of the user's friends with risk records or the frequency of risk events occurring in the groups the user participates in. The risk status of related entities may affect the risk status of the user being evaluated; therefore, related entity characteristics are also an important component of the basic risk characteristic set.

[0141] Through the above extraction operations, a multi-dimensional set of basic risk features is obtained. Each dimension corresponds to a specific risk feature, which together constitute the initial feature representation of the sample to be evaluated.

[0142] Step S142: Input the basic risk feature set into the feature preprocessing layer of the cross-domain risk prediction model, perform feature standardization processing, and generate a standardized risk feature set.

[0143] The feature preprocessing layer of the cross-domain risk prediction model is responsible for standardizing the input set of basic risk features to eliminate the influence of different features due to differences in units and numerical ranges, enabling the model to more effectively handle and learn the relationships between features. After performing the above standardization process on all features in the basic risk feature set, a standardized risk feature set is obtained.

[0144] Step S143: Perform cross-domain feature mapping processing on the standardized risk feature set through the feature transfer layer of the cross-domain risk prediction model. The feature transfer layer includes a domain difference perception module, which dynamically adjusts the feature mapping rules based on the domain difference metric matrix.

[0145] The feature transfer layer of the cross-domain risk prediction model is the key part to realize the cross-domain transfer of risk features. Its core function is to map the standardized risk feature set from the feature space of the target domain to the feature space of the source domain, or to a feature space shared by the two domains, so as to use the risk knowledge of the source domain for risk assessment.

[0146] The domain difference awareness module of the feature transfer layer stores the previously generated domain difference metric matrix. This module can dynamically adjust the feature mapping rules based on the difference coefficients of each dimension in the domain difference metric matrix. For example, when the rule fit difference coefficient in the domain difference metric matrix is ​​large, it indicates that the business rules of the two domains are significantly different. The domain difference awareness module will adjust the feature mapping rules to make the features related to the business rules more adaptively adjusted during the mapping process. When the behavior similarity difference coefficient is small, it indicates that the user behavior of the two domains is highly similar. The feature mapping rules will reduce the adjustment range of user behavior features.

[0147] Adjusting the feature mapping rules involves changing the parameters of the mapping function. These parameters are dynamically updated based on the specific values ​​of the domain difference metric matrix to ensure that the feature mapping process can minimize the impact of domain differences and retain valuable feature information for risk assessment. Through the processing of the feature transfer layer, the standardized risk feature set is transformed into a feature representation that can adapt to the risk knowledge of the source domain.

[0148] Step S144: Perform dimension alignment processing on the standardized risk feature set through the feature alignment submodule of the feature transfer layer to generate a dimension-aligned risk feature set.

[0149] The feature alignment submodule of the feature transfer layer is designed to address the potential inconsistencies in the dimensions of risk features between the source and target domains. Through dimension alignment, it ensures that the dimensions of the standardized risk feature set match the expected feature dimensions of the source domain risk prediction model, or that both are processed within a unified dimensional space.

[0150] Dimension alignment processing includes two scenarios: dimension increase and dimension decrease. When the dimensions of the standardized risk feature set in the target domain are less than the feature dimensions required by the risk prediction model in the source domain, the feature alignment submodule generates corresponding supplementary features based on the missing dimensions in the source domain features. The values ​​of the supplementary features can be estimated based on the feature distribution of the target domain and the domain difference metric matrix. When the dimensions of the standardized risk feature set in the target domain are more than the feature dimensions required by the risk prediction model in the source domain, the feature alignment submodule filters out dimensions with low correlation to the source domain features for deletion, or combines multiple dimensions into one dimension through feature combination to reduce the feature dimensions. For example, the feature space of the risk prediction model in the source domain includes user consumption frequency features, while the standardized risk feature set in the target domain does not have this feature but does have user financial transaction frequency features. The feature alignment submodule transforms the user financial transaction frequency features based on the behavioral similarity difference coefficient between the two domains to generate supplementary features that align with the dimension of the user consumption frequency features. After dimension alignment processing, the dimensions of the standardized risk feature set match the dimensions of the risk features in the source domain, generating a dimension-aligned risk feature set.

[0151] Step S145: Perform distribution adaptation processing on the dimension-aligned risk feature set through the feature distribution transformation submodule of the feature transfer layer, adjust the feature value distribution based on the source domain feature distribution parameters, and generate a distribution-adapted risk feature set.

[0152] Step S145: Perform distribution adaptation processing on the dimension-aligned risk feature set through the feature distribution transformation submodule of the feature transfer layer, adjust the feature value distribution based on the source domain feature distribution parameters, and generate a distribution-adapted risk feature set. The feature distribution transformation submodule aims to make the risk feature set after dimension alignment more suitable for the target domain in terms of distribution. By referencing the feature distribution parameters of the source domain, it adjusts the feature value distribution and improves the effectiveness of the features. For example, step S1451: Extract the feature distribution parameters of the source domain risk feature space, which include the mean vector and covariance matrix of each feature dimension. Feature distribution parameters are extracted from the source domain risk feature space of e-commerce platform transaction risk control scenarios. The mean vector reflects the average level of each feature dimension, such as the mean of the user transaction amount feature dimension and the mean of the transaction frequency feature dimension. The covariance matrix reflects the correlation between different feature dimensions, such as the covariance between transaction amount and transaction frequency, thus fully describing the distribution characteristics of source domain risk features. Step S1452: Calculate the current mean and current covariance of each feature dimension in the dimension-aligned risk feature set. For the set of risk features of financial business on social platforms after dimension alignment, calculate the current mean of each feature dimension, such as the mean of the user transfer amount feature dimension; at the same time, calculate the current covariance between each feature dimension, such as the covariance between transfer amount and social interaction frequency, so as to reflect the current distribution status of features in the target domain. Step S1453: Construct a distribution transformation function that takes the feature value, the mean of the source domain feature dimension, the covariance of the source domain feature dimension, the current mean of the target domain feature dimension, and the current covariance of the target domain feature dimension as input. The purpose of a distribution transformation function is to transform the feature values ​​of the target domain from their current distribution to a state that more closely resembles the distribution characteristics of the source domain. The function comprehensively considers various input parameters and achieves distribution adaptation by adjusting the position and dispersion of the feature values. Step S1454: For each feature value in the dimension-aligned risk feature set, perform a standardization transformation process using a distribution transformation function to transform the feature value from the current distribution of the target domain to a transformed feature value that conforms to the distribution characteristics of the source domain. For each feature value in the dimension-aligned risk feature set, a distribution transformation function is used for processing. For example, the feature value of a social platform user's transfer amount is transformed based on the mean and covariance of transaction amounts in the source domain, as well as its own current mean and covariance, to obtain a new feature value that more closely approximates the distribution of the corresponding feature in the source domain. Step S1455: Arrange all transformed feature values ​​in the order of their original feature dimensions to generate a distribution-adapted risk feature set. Each transformed feature value is organized and arranged according to the original feature dimension order to form a distribution-adapted risk feature set, which not only retains the dimensional structure of the features, but also makes the distribution more similar to the features in the source domain.

[0153] Step S146: The feature enhancement submodule of the feature transfer layer performs importance weighting on the distribution adaptation risk feature set, determines the weight value of each feature based on the risk mapping degree difference coefficient in the domain difference metric matrix, and generates the transfer adaptation risk feature set.

[0154] The feature enhancement submodule of the feature transfer layer is used to highlight features that are more important to risk assessment in the distribution-adapted risk feature set, while suppressing less important features. Through importance-weighted processing, it improves the expressive power of the features. In other words, its core is to determine the weight value of each feature. The weight value is determined based on the risk mapping degree difference coefficient in the domain difference metric matrix. The risk mapping degree difference coefficient reflects the mapping relationship and degree of difference between the risk types in the source and target domains. Features with high relevance to risk assessment should have larger weight values, and vice versa.

[0155] Specifically, the feature enhancement submodule first analyzes the correlation between each feature in the distribution-adapted risk feature set and the risk type. Combining this with the risk mapping difference coefficient, it calculates the importance score for each feature. The importance score considers both the feature's importance in the source domain risk assessment and the strength of its correlation with the risk type in the target domain. A smaller risk mapping difference coefficient indicates a closer correlation between the risk types in the two domains, and the weight of important features in the source domain will correspondingly increase in the target domain.

[0156] After obtaining the importance score for each feature, the feature enhancement submodule normalizes these scores so that the sum of the weights of all features is one. The normalization process involves dividing the importance score of each feature by the sum of the importance scores of all features. This ensures the reasonableness of the weight allocation and avoids excessive suppression of other features due to excessively high scores for individual features.

[0157] After determining the weight values, the feature enhancement submodule multiplies each feature in the distribution-adapted risk feature set with its corresponding weight value. For example, if the distribution-adapted risk feature set contains multiple features such as feature A, feature B, and feature C, each of which is a multi-dimensional vector, with feature A having a weight of W_A, feature B having a weight of W_B, feature C having a weight of W_C, and so on, then after weighting, feature A becomes the product of feature A and W_A, feature B becomes the product of feature B and W_B, feature C becomes the product of feature C and W_C, and so on.

[0158] All weighted features are arranged and combined according to their original order to form a transfer-adaptive risk feature set. This transfer-adaptive risk feature set retains the distributional characteristics of the features in the distribution-adaptive risk feature set, while highlighting key features through importance weighting, making it more adaptable to the risk assessment needs of the target domain.

[0159] Step S147: Input the set of transfer adaptation risk features into the risk assessment layer of the cross-domain risk prediction model. The risk assessment layer contains a multilayer perceptron network. Through multilayer nonlinear transformation, risk level mapping processing is performed on the set of transfer adaptation risk features to generate a high-level risk feature vector.

[0160] The risk assessment layer of the cross-domain risk prediction model is the core component for mapping risk levels. Its multilayer perceptron network consists of an input layer, multiple hidden layers, and an output layer. The transfer-adapted risk feature set serves as the input to the multilayer perceptron network, first entering the input layer. The input layer transforms the feature set into a vector form suitable for network processing and then passes it to the first hidden layer.

[0161] The first hidden layer performs a non-linear transformation on the input vector, processing each element of the vector through an activation function (such as the ReLU function) to enhance the non-linear expressive power of the features. The processed result is passed to the next hidden layer, and each hidden layer repeats a similar non-linear transformation process. The number of neurons and the connection method of different hidden layers are set according to the needs of risk assessment. Through multiple layers of non-linear transformation, deep information related to the risk level in the features is gradually extracted.

[0162] After processing through all hidden layers, the result is passed to the output layer, which transforms the result into a high-level risk feature vector. This high-level risk feature vector is a multi-dimensional vector, where each dimension represents an abstract feature related to a specific risk level. These features integrate key information from the transfer-adapted risk feature set, more accurately reflecting the risk status of the sample to be evaluated in the target domain.

[0163] Step S148: Call the output layer of the cross-domain risk prediction model to perform probability transformation processing on the high-level risk feature vector, calculate the probability distribution of different risk levels through the softmax activation function, and determine the risk level score corresponding to the maximum probability value in the probability distribution as the preliminary risk prediction score.

[0164] The output layer of the cross-domain risk prediction model receives high-level risk feature vectors and processes them using the softmax activation function. The softmax activation function converts each element in the high-level risk feature vector into a probability value between 0 and 1, and the sum of all probability values ​​is 1, thus obtaining the probability distribution of different risk levels.

[0165] For example, a high-level risk feature vector contains elements related to three levels: low risk, medium risk, and high risk. After processing with the softmax activation function, the probabilities of low risk, medium risk, and high risk are obtained. These three probability values ​​together constitute the probability distribution.

[0166] After obtaining the probability distribution, the maximum probability value is identified. The score corresponding to the risk level of this maximum probability value is the preliminary risk prediction score. The preliminary risk prediction score reflects the risk assessment result of the sample to be evaluated in the target area under the current model.

[0167] Step S150: Perform cross-domain calibration processing on the preliminary risk prediction score based on the domain difference metric matrix to generate the final risk prediction result.

[0168] The preliminary risk prediction score is calculated based on the model obtained from transfer learning. Since there are certain differences between the source domain and the target domain, the preliminary risk prediction score may not be able to fully and accurately reflect the actual risk situation in the target domain. Therefore, it is necessary to perform cross-domain calibration based on the domain difference metric matrix to improve the accuracy of risk prediction and finally generate prediction results that can truly reflect the risk of the sample to be evaluated in the target domain.

[0169] Step S151: Extract the rule fit difference coefficient, behavior similarity difference coefficient, risk mapping difference coefficient, and distribution offset difference coefficient from the domain difference measurement matrix.

[0170] The rule fit difference coefficient, behavior similarity difference coefficient, risk mapping difference coefficient, and distribution offset difference coefficient are extracted from the domain difference measurement matrix. These four coefficients comprehensively reflect the differences between the source domain and the target domain in different dimensions and are the key basis for cross-domain calibration processing.

[0171] Step S152: Construct a calibration coefficient calculation model. The calibration coefficient calculation model includes a difference coefficient weighting module and a calibration value generation module. The difference coefficient weighting module performs weighted summation processing on the rule fit difference coefficient, behavior similarity difference coefficient, risk mapping difference coefficient and distribution offset difference coefficient based on the importance weight of each difference coefficient to generate a comprehensive difference index.

[0172] The difference coefficient weighting module of the calibration coefficient calculation model first assigns an importance weight to each difference coefficient. The importance weight is determined based on the degree of influence of each difference coefficient in cross-domain risk prediction. The difference coefficient with the greater influence on the risk prediction result has a greater importance weight.

[0173] Then, the difference coefficient weighting module multiplies each difference coefficient by its corresponding importance weight to obtain a weighted difference coefficient. Finally, all weighted difference coefficients are summed to obtain the comprehensive difference index. The comprehensive difference index integrates the information from the four difference coefficients, reflecting the overall degree of difference between the source and target domains.

[0174] Step S153: Input the comprehensive difference index into the calibration value generation module, and convert the comprehensive difference index into calibration coefficients through the sigmoid function. The calibration coefficients range from 0 to 1.

[0175] The sigmoid function in the calibration value generation module can convert the overall difference index (which can be any real number) into a calibration coefficient between 0 and 1. When the overall difference index is large, the output of the sigmoid function is close to 1, indicating that the overall difference between the source and target domains is large, requiring a greater degree of calibration; when the overall difference index is small, the output is close to 0, indicating that the difference between the two domains is small, and the degree of calibration is also small.

[0176] Step S154: Construct a cross-domain calibration function. The input of the cross-domain calibration function is the preliminary risk prediction score and the calibration coefficient, and the output is the calibrated risk score.

[0177] The construction of cross-domain calibration functions is based on the relationship between preliminary risk prediction scores and calibration coefficients. The core idea is to adjust the preliminary risk prediction scores according to the calibration coefficients to eliminate prediction biases caused by domain differences. The form of the calibration function can be set according to the actual situation; for example, it can be the product of the preliminary risk prediction score and (1 plus the calibration coefficient), or other function forms that can reasonably reflect the calibration relationship.

[0178] Step S155: Input the preliminary risk prediction score and calibration coefficient into the cross-domain calibration function, perform calibration calculation processing, and generate the calibrated risk score.

[0179] The preliminary risk prediction score obtained in step S148 and the calibration coefficient obtained in step S153 are substituted into the cross-domain calibration function, and the calculation is performed according to the calculation method set by the cross-domain calibration function to obtain the calibrated risk score. Compared with the preliminary risk prediction score, the calibrated risk score is more adapted to the risk assessment standards of the target domain, reducing the prediction error caused by domain differences.

[0180] Step S156: Determine the risk level based on the calibrated risk score and generate the final risk prediction result. The final risk prediction result includes the calibrated risk score, risk level, probability distribution of risk type, and explanation of the impact of domain differences.

[0181] Based on the pre-defined risk level classification standards for the target domain, the calibrated risk scores are mapped to the corresponding risk levels (such as low risk, medium risk, and high risk). Simultaneously, by combining the probability distributions of different risk types output by the cross-domain risk prediction model (extracted from the probability distribution in step S148) and the impact analysis of each difference coefficient in the domain difference measurement matrix on the prediction results, a domain difference impact explanation is formed.

[0182] The final risk prediction result integrates the calibrated risk score, the corresponding risk level, the probability distribution of risk type, and the impact of domain differences on the prediction result. Those skilled in the art can take corresponding risk control measures based on the result.

[0183] Figure 2The risk control cross-domain risk prediction system 100 combining transfer learning, as shown, includes a processor 1001 and a memory 1003. The processor 1001 and memory 1003 are connected, for example, via a bus 1002. Optionally, the risk control cross-domain risk prediction system 100 combining transfer learning may further include a transceiver 1004. The transceiver 1004 can be used for data interaction between this risk control cross-domain risk prediction system combining transfer learning and other risk control cross-domain risk prediction systems combining transfer learning, such as sending and / or receiving data. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of this risk control cross-domain risk prediction system 100 combining transfer learning does not constitute a limitation on the embodiments of this application.

[0184] The memory 1003 is used to store program code for executing the embodiments of this application, and its execution is controlled by the processor 1001. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0185] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application, without departing from the technical concept of this application, also fall within the protection scope of the embodiments of this application.

Claims

1. A cross-domain risk prediction method for risk control combining transfer learning, characterized in that, The method includes: Perform domain difference measurement operations on the source domain risk control scenario and the target domain risk control scenario to generate a domain difference measurement matrix; A hierarchical knowledge transfer channel is constructed based on the domain difference metric matrix to perform selective transfer processing on the multi-level network parameters of the source domain risk prediction model, generating a target domain initial risk prediction model. The multi-level network includes an input layer, a hidden layer, and an output layer. The hierarchical knowledge transfer channel includes an input layer transfer sub-channel, a hidden layer transfer sub-channel, and an output layer transfer sub-channel, each corresponding to different parameter transformation rules. The domain difference metric matrix is ​​used to determine the domain transfer weights of the network parameters at each level. A knowledge transfer channel is generated based on these domain transfer weights, and selective transfer is performed on the multi-level network parameters. A cross-domain risk prediction model is generated by performing dynamic adaptation training on the target domain initialization risk prediction model using a target domain risk sample set. Input the target domain sample to be evaluated into the cross-domain risk prediction model, perform risk feature transfer reasoning operation, and generate a preliminary risk prediction score; Based on the domain difference metric matrix, cross-domain calibration is performed on the preliminary risk prediction score to generate the final risk prediction result. The step of performing a domain difference measurement operation on the source domain risk control scenario and the target domain risk control scenario to generate a domain difference measurement matrix includes: Extract the domain attribute feature set of the source domain risk control scenario. The domain attribute feature set includes business rule features, user behavior features, risk performance features, and data distribution features. The business rule features include transaction verification rules, refund risk control rules, and fraudulent transaction identification rules. The user behavior features include user browsing behavior features, add-to-cart behavior features, order placement behavior features, and payment behavior features on the e-commerce platform. The browsing behavior features include the category distribution of browsed products, the browsing duration distribution, and the browsing time period distribution. The add-to-cart behavior features involve the conversion rate between added-to-cart products and browsed products, and the price range distribution of added-to-cart products. The order placement behavior features include the order frequency, the category combination of ordered products, and the device information distribution when placing an order. The payment behavior features include payment method preferences, payment time period distribution, and payment failure number distribution. The data distribution features include the distribution of each threshold in the business rules, the probability distribution of each dimension of user behavior feature data, and the statistical distribution of each indicator of risk performance feature. Extract the domain attribute feature set of the target domain risk control scenario. The domain attribute feature set includes business rule features, user behavior features, risk performance features and data distribution features. The business rule features include social relationship chain verification rules, financial product purchase qualification review rules and fund transfer risk control rules. The user behavior features include social interaction behavior on social platforms, content publishing behavior and financial business participation behavior. Social interaction behavior characteristics include the distribution of friend interaction frequency, interaction message type, and interaction time period; content publishing behavior characteristics include the distribution of published content type, publishing frequency, and the distribution of published content interaction volume; financial business participation behavior characteristics include the frequency of participating in financial product consultation, the duration of viewing financial product information, and the frequency of initiating financial transactions. Based on the set of domain attribute features of the source domain risk control scenario and the target domain risk control scenario, the rule fit difference coefficient, behavior similarity difference coefficient, risk mapping difference coefficient and distribution offset difference coefficient are determined. The rule fit difference coefficient, behavior similarity difference coefficient, risk mapping difference coefficient and distribution offset difference coefficient are arranged in row vector form to generate the domain difference measurement matrix. The step of inputting the target domain sample to be evaluated into the cross-domain risk prediction model, performing risk feature transfer inference operations, and generating a preliminary risk prediction score includes: A risk feature extraction operation is performed on the target domain samples to be evaluated to extract a basic risk feature set. The basic risk feature set includes user attribute features, behavioral transaction features, historical risk features, and related entity features. The user attribute features include user registration information features and identity information features. The historical risk features are features extracted based on the user's past risk performance, including the type, frequency, and time interval of risk events that occurred in the user's history, as well as the user's penalty records for risk events. The related entity features involve the features of other entities associated with the user of the sample to be evaluated, including the proportion of risk records among the user's friends and the frequency of risk events occurring in the groups the user participates in. The behavioral transaction features cover the user's social behavior features and financial transaction behavior features on social platforms. The social behavior features include the number of the user's friends, the frequency of friend interactions, and the type and frequency of content posted. The financial transaction behavior features include the type of financial business the user participates in, the range of transaction amounts, the transaction frequency, and the transaction time period. The basic risk feature set is input into the feature preprocessing layer of the cross-domain risk prediction model, and feature standardization processing is performed to generate a standardized risk feature set. The cross-domain risk prediction model performs cross-domain feature mapping on the standardized risk feature set through its feature transfer layer. The feature transfer layer includes a domain difference perception module, which dynamically adjusts the feature mapping rules based on the domain difference metric matrix. The feature alignment submodule of the feature transfer layer performs dimension alignment processing on the standardized risk feature set to generate a dimension-aligned risk feature set. The feature distribution transformation submodule of the feature transfer layer performs distribution adaptation processing on the dimension-aligned risk feature set, adjusts the feature value distribution based on the source domain feature distribution parameters, and generates a distribution-adapted risk feature set. The feature enhancement submodule of the feature transfer layer performs importance weighting processing on the distribution adaptation risk feature set, and determines the weight value of each feature based on the risk mapping degree difference coefficient in the domain difference metric matrix to generate the transfer adaptation risk feature set. The migration adaptation risk feature set is input into the risk assessment layer of the cross-domain risk prediction model. The risk assessment layer contains a multilayer perceptron network and performs risk level mapping processing on the migration adaptation risk feature set through multilayer nonlinear transformation to generate a high-level risk feature vector. The output layer of the cross-domain risk prediction model is invoked to perform probability transformation processing on the high-level risk feature vector. The probability distribution of different risk levels is calculated through the softmax activation function, and the risk level score corresponding to the maximum probability value in the probability distribution is determined as the preliminary risk prediction score.

2. The risk control cross-domain risk prediction method combining transfer learning according to claim 1, characterized in that, The determination of rule fit difference coefficient, behavior similarity difference coefficient, risk mapping difference coefficient, and distribution offset difference coefficient based on the domain attribute feature set of the source domain risk control scenario and the target domain risk control scenario includes: A domain difference assessment dimension system is constructed, which includes rule adaptability dimension, behavior similarity dimension, risk mapping dimension, and distribution offset dimension. Regarding the rule adaptability dimension, the business rule features of the source domain risk control scenario and the business rule features of the target domain risk control scenario are structurally compared, the rule element difference points are extracted, and the rule adaptability difference coefficient is calculated based on the proportion of the number of difference points. For the behavioral similarity dimension, time-series pattern matching is performed on the user behavioral features of the source domain risk control scenario and the user behavioral features of the target domain risk control scenario to generate a behavioral sequence similarity score. The behavioral similarity difference coefficient is obtained by subtracting the behavioral sequence similarity score from 1. For the risk mapping degree dimension, a risk type mapping dictionary is constructed, and the overlap ratio and semantic association strength of risk types in the source domain risk control scenario and the target domain risk control scenario are statistically analyzed. The risk mapping degree difference coefficient is calculated based on the weighted average. For the distribution offset dimension, the maximum mean difference algorithm is used to calculate the distribution distance between the data distribution characteristics of the source domain risk control scenario and the data distribution characteristics of the target domain risk control scenario. The distribution distance is then standardized and used as the distribution offset difference coefficient. The rule fit difference coefficient, the behavior similarity difference coefficient, the risk mapping difference coefficient, and the distribution offset difference coefficient are arranged in row vector form to generate the domain difference measurement matrix.

3. The risk control cross-domain risk prediction method combining transfer learning according to claim 2, characterized in that, Regarding the rule adaptability dimension, a structured comparison is performed between the business rule features of the source domain risk control scenario and the business rule features of the target domain risk control scenario. Differences in rule elements are extracted, and a rule adaptability difference coefficient is calculated based on the proportion of these differences. This includes: The business rule features of the source domain risk control scenario are analyzed to generate a set of source domain rule elements, which includes rule condition elements, rule threshold elements, rule execution logic elements, and rule output result elements. The business rule features of the risk control scenario in the target domain are analyzed to generate a set of rule elements in the target domain. The set of rule elements in the target domain includes rule condition elements, rule threshold elements, rule execution logic elements, and rule output result elements. Construct a rule element comparison template, which includes element type identifier, element attribute description, and element association relationship description; Based on the rule element comparison template, perform an element-by-element comparison operation on the source domain rule element set and the target domain rule element set, and record the element entries with differences as rule element difference points; The total number of difference points of the rule elements is counted, and the ratio of the total number of difference points of the rule elements to the total number of elements in the source domain rule element set is calculated. The ratio is determined as the rule fit difference coefficient.

4. The risk control cross-domain risk prediction method combining transfer learning according to claim 1, characterized in that, The process of constructing a hierarchical knowledge transfer channel based on the domain difference metric matrix, performing selective transfer processing on the multi-level network parameters of the source domain risk prediction model, and generating a target domain initial risk prediction model includes: The network architecture of the source domain risk prediction model is analyzed to determine a multi-level network structure including an input layer, a hidden layer and an output layer. The hidden layer includes a feature extraction sub-layer, a feature fusion sub-layer and a risk assessment sub-layer. Each network structure corresponds to a set of network parameters. The domain migration weights of the network parameters at each level are calculated based on the domain difference metric matrix. The domain migration weights are negatively correlated with the difference coefficients of the corresponding levels. A hierarchical filtering mechanism is constructed, which includes a migration weight threshold judgment module and a parameter retention decision module. The migration weight threshold judgment module is used to compare the domain migration weight of each level of network parameters with a preset threshold. The parameter retention decision module decides whether to retain the corresponding level of network parameters based on the comparison result. The hierarchical filtering mechanism performs selective filtering on the multi-level network parameters, retaining the hierarchical network parameters with transfer weights higher than a preset threshold as transferable knowledge parameters. The transferable knowledge parameters are input into the corresponding transfer sub-channel, and cross-domain parameter mapping and transformation processing is performed to generate target domain hierarchical parameters that are adapted to the target domain network structure. The target domain hierarchical parameters are loaded into the target domain network framework in hierarchical order to generate the target domain initial risk prediction model.

5. The risk control cross-domain risk prediction method combining transfer learning according to claim 4, characterized in that, The calculation of the domain transfer weights for each layer of network parameters based on the domain difference metric matrix includes: Extract the difference coefficients of the rule fit dimension from the domain difference metric matrix, and determine the rule fit difference coefficients as the difference evaluation benchmarks for the input layer network parameters; Extract the difference coefficients of the behavior similarity dimension and the risk mapping dimension from the domain difference measurement matrix, and calculate the arithmetic mean of the two as the difference evaluation benchmark for the hidden layer network parameters; Extract the difference coefficients of the distribution offset dimension from the domain difference metric matrix, and determine the distribution offset difference coefficients as the difference evaluation benchmarks for the output layer network parameters; A difference evaluation benchmark normalization function is constructed. The difference evaluation benchmarks of the input layer, hidden layer and output layer are input into the difference evaluation benchmark normalization function to generate the relative difference value of each layer. The value of the relative difference value is in the range of 0 to 1. Subtracting the relative difference value of each level from 1 yields the domain migration weight of the network parameters at each level, where the domain migration weight ranges from 0 to 1.

6. The risk control cross-domain risk prediction method combining transfer learning according to claim 1, characterized in that, The step of using a risk sample set from the target domain to perform dynamic adaptation training on the initial risk prediction model for the target domain, generating a cross-domain risk prediction model, includes: The target domain risk sample set is divided into an adaptation training sample subset, a validation sample subset, and a test sample subset. The proportion of the adaptation training sample subset to the total number of samples is greater than that of the validation sample subset and the test sample subset. The adaptation training sample subset is used for model parameter adjustment, the validation sample subset is used for training process monitoring, and the test sample subset is used for final performance evaluation. A dynamic loss function is constructed based on the domain difference metric matrix. The dynamic loss function includes a basic loss term and a domain adaptation loss term. The basic loss term uses the cross-entropy loss function to calculate the difference between the predicted risk type and the actual risk type. The domain adaptation loss term is based on the domain difference coefficient to weight and adjust the loss value of different risk types. Initialize the weight coefficients of the dynamic loss function, wherein the weight coefficients include the base loss weight and the domain adaptation loss weight, and in the initial state the base loss weight is greater than the domain adaptation loss weight; The subset of adaptive training samples is divided into multiple training batches according to a preset batch size, and each training batch is sequentially input into the target domain to initialize the risk prediction model. The predicted risk value for each training batch is calculated through forward propagation, and the batch loss value is calculated using the dynamic loss function in combination with the true risk label of the sample. Based on the batch loss value, the network parameters of the target domain initial risk prediction model are updated using the backpropagation algorithm, and a momentum optimization strategy is used to accelerate convergence during the update process. After completing a preset number of iterations of training, the subset of validation samples is input into the current model, and the prediction accuracy of the validation set is calculated. When the improvement of the prediction accuracy of the validation set after K consecutive iterations is less than a preset threshold, the weight coefficient of the dynamic loss function is dynamically adjusted to increase the proportion of the domain adaptation loss weight. Training stops when the prediction accuracy of the validation set remains stable for a preset number of consecutive times or reaches a preset target value. The test sample subset is then input into the current model to perform a performance evaluation, a model performance evaluation report is generated, and the model that passes the performance evaluation is determined as the cross-domain risk prediction model based on the model performance evaluation report.

7. The risk control cross-domain risk prediction method combining transfer learning according to claim 6, characterized in that, The construction of the dynamic loss function based on the domain difference metric matrix includes: Extract the distribution offset dimension difference coefficient from the domain difference metric matrix, and determine the distribution offset dimension difference coefficient as the basic adjustment factor for the domain adaptation loss term; Construct an iterative decay coefficient generation function, wherein the input of the iterative decay coefficient generation function is the number of training iterations, and the output is an iterative decay coefficient that gradually decays from 1 to 0 as the number of iterations increases; The basic adjustment factor and the iterative decay coefficient are multiplied to generate the domain adaptation loss weight coefficient. Construct a risk type weight vector, where each element of the risk type weight vector corresponds to a risk type, and the element value is inversely proportional to the frequency of occurrence of that risk type in the target domain. The risk type weight vector is introduced into the basic loss term, and the cross-entropy loss values ​​of different risk types are weighted and summed to generate a weighted basic loss term. The domain adaptation loss weight coefficient is multiplied by the weighted basic loss term to generate the domain adaptation loss term; Construct a dynamic loss function expression, which is the arithmetic sum of the weighted basic loss term and the domain adaptation loss term, and determine the dynamic loss function expression as the dynamic loss function.

8. The risk control cross-domain risk prediction method combining transfer learning according to claim 1, characterized in that, The process of performing cross-domain calibration on the preliminary risk prediction score based on the domain difference metric matrix to generate the final risk prediction result includes: Extract the rule fit difference coefficient, behavior similarity difference coefficient, risk mapping difference coefficient, and distribution offset difference coefficient from the domain difference measurement matrix; A calibration coefficient calculation model is constructed, which includes a difference coefficient weighting module and a calibration value generation module. The difference coefficient weighting module performs weighted summation processing on the rule fit difference coefficient, the behavior similarity difference coefficient, the risk mapping difference coefficient, and the distribution offset difference coefficient based on the importance weight of each difference coefficient to generate a comprehensive difference index. The comprehensive difference index is input into the calibration value generation module, and the comprehensive difference index is converted into a calibration coefficient by the sigmoid function. The calibration coefficient has a value range of 0 to 1. Construct a cross-domain calibration function, the input of which is the preliminary risk prediction score and the calibration coefficient, and the output is the calibrated risk score; The preliminary risk prediction score and the calibration coefficient are input into the cross-domain calibration function to perform calibration calculation and generate the calibrated risk score. Based on the calibrated risk score, the risk level is determined, and a final risk prediction result is generated. The final risk prediction result includes the calibrated risk score, risk level, probability distribution of risk type, and explanation of the impact of domain differences.

9. A cross-domain risk prediction system for risk control combining transfer learning, characterized in that, The method includes a processor and a computer-readable storage medium storing machine-executable instructions that, when executed by the processor, implement the cross-domain risk prediction method for risk control combining transfer learning as described in any one of claims 1-8.

Images