Financial platform risk prediction method and system based on multi-source user behavior data

Abstract

The present application provides a financial platform risk prediction method and system based on multi-source user behavior data, which obtains multi-source user behavior data indicated by the financial platform risk prediction task, and uses graph autoencoding technology to process the data to generate graph autoencoding structure data, thereby performing a preliminary assessment of user behavior in the preliminary risk prediction stage to generate first risk prediction data, and then determining the influence weights of different features on risk prediction based on the influence feature domain and probability score of the graph autoencoding structure data, and then using the influence weights output by the target graph autoencoder to converge the encoded data to generate feature interpretation data, and finally generating an attention weight distribution based on the feature interpretation data, and blending it with the original multi-source user behavior data to generate blended data, and performing advanced risk prediction based on the blended data to generate more specific and accurate second risk prediction data. In this way, user behavior risks can be comprehensively and accurately assessed.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method and system for predicting financial platform risks based on multi-source user behavior data. Background Art

[0002] With the rapid development of financial technology, risk management of financial platforms has become an important part of ensuring financial security and maintaining market order. Traditional risk management methods often rely on a single data source and simple statistical analysis, which makes it difficult to comprehensively and accurately assess user behavior risks. Therefore, it is of great significance to develop a financial platform risk prediction method based on multi-source user behavior data.

[0003] Most existing risk prediction methods focus on the analysis of a single data source, ignoring the correlation and complementarity between multi-source data. At the same time, traditional risk assessment models often lack in-depth mining of user behavior patterns, making it difficult to accurately capture potential risks. In addition, existing risk prediction methods often face problems such as large computational complexity and low efficiency when processing complex user behavior data. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a financial platform risk prediction method based on multi-source user behavior data, the method comprising:

[0005] Acquire multi-source user behavior data indicated by a financial platform risk prediction task, and perform graph autoencoding processing on the multi-source user behavior data to generate graph autoencoding structure data corresponding to the multi-source user behavior data; the graph autoencoding structure data includes graph autoencoding structure data of multiple encoding dimensions corresponding to each graph autoencoder generated by performing graph autoencoding processing on the multi-source user behavior data according to multiple graph autoencoders respectively;

[0006] Performing preliminary risk prediction on the multi-source user behavior data according to the graph autoencoder structure data to generate first risk prediction data of the multi-source user behavior data, wherein the first risk prediction data includes financial platform risk knowledge points and probability scores corresponding to the multi-source user behavior data;

[0007] Based on the influence feature domain of the graph autoencoder structure data and the probability score, determining the influence weight of the graph autoencoder structure data on the risk prediction of the multi-source user behavior data; the influence feature domain is the feature domain of the perception node in the graph autoencoder structure data;

[0008] Extracting a graph autoencoder that meets the target feature requirements from the multiple graph autoencoders to generate a target graph autoencoder, aggregating the graph autoencoder structure data of multiple encoding dimensions of the target graph autoencoder based on the influence weights corresponding to the graph autoencoder structure data output by the target graph autoencoder, and generating feature interpretation data representing the description of the multi-source user behavior data as the corresponding financial platform risk knowledge point;

[0009] An attention weight distribution reflecting the feature influence state weights of the multi-source user behavior data is generated based on the feature interpretation data, and the multi-source user behavior data and the attention weight distribution are blended to generate blended data. Advanced risk prediction is performed on the multi-source user behavior data based on the blended data to generate second risk prediction data for the multi-source user behavior data, wherein the second risk prediction data represents the target financial platform risk knowledge points corresponding to the multi-source user behavior data.

[0010] In a possible implementation of the first aspect, performing preliminary risk prediction on the multi-source user behavior data according to the graph autoencoder structure data to generate first risk prediction data for the multi-source user behavior data includes:

[0011] Using a risk assessment unit to perform preliminary risk prediction on the multi-source user behavior data according to the graph autoencoding structure data, to generate first risk prediction data for the multi-source user behavior data;

[0012] The step of performing advanced risk prediction on the multi-source user behavior data based on the blended data to generate second risk prediction data for the multi-source user behavior data includes:

[0013] An optimization evaluation unit is used to perform advanced risk prediction on the multi-source user behavior data based on the blended data to generate second risk prediction data for the multi-source user behavior data.

[0014] In a possible implementation of the first aspect, before the step of using a risk assessment unit to perform preliminary risk prediction on the multi-source user behavior data according to the graph autoencoder structure data to generate first risk prediction data for the multi-source user behavior data, the method further includes:

[0015] Acquire sample learning data, wherein the sample learning data includes template multi-source user behavior data and its corresponding annotated risk knowledge point data and sample annotation probability;

[0016] Using a risk assessment unit to perform graph autoencoding processing on the template multi-source user behavior data, generate template graph autoencoding structure data corresponding to the template multi-source user behavior data, and perform preliminary risk prediction on the template multi-source user behavior data based on the template graph autoencoding structure data to generate predicted risk knowledge points and predicted probability scores;

[0017] Based on the influence feature domain of the template graph auto-encoding structure data and the prediction probability score, calculating the template influence weight of the template graph auto-encoding structure data for risk prediction of the template multi-source user behavior data;

[0018] Aggregating the template graph auto-encoding structure data based on the template influence weight to generate template feature interpretation data representing the matching of the template multi-source user behavior data with the predicted risk knowledge point;

[0019] Using the optimization evaluation unit to perform advanced risk prediction on the template multi-source user behavior data according to the template feature interpretation data to generate target prediction risk knowledge points;

[0020] Calculating the characteristic distance between the predicted risk knowledge point and the annotated risk knowledge point data to generate a first training cost value, and calculating the characteristic distance between the target predicted risk knowledge point and the annotated risk knowledge point data to generate a second training cost value;

[0021] The weight and bias information of the risk assessment unit are updated based on the first training cost value, and the weight and bias information of the optimization assessment unit are updated based on the second training cost value until the training convergence requirement is met.

[0022] In a possible implementation of the first aspect, the template graph autoencoding structure data includes the template graph autoencoding structure data of multiple encoding dimensions corresponding to each graph autoencoder generated by performing graph autoencoding processing on the template multi-source user behavior data according to multiple graph autoencoders of the risk assessment unit, and the template graph autoencoding structure data is aggregated based on the template influence weight to generate template feature interpretation data representing the description of the template multi-source user behavior data matching the predicted risk knowledge point, including:

[0023] Extract a graph autoencoder that meets the target feature requirements from multiple graph autoencoders to generate a target graph autoencoder;

[0024] Based on the template influence weight corresponding to the template graph autoencoder structure data output by the target graph autoencoder, the template graph autoencoder structure data of multiple encoding dimensions of the target graph autoencoder are aggregated to generate template feature interpretation data representing the matching of the template multi-source user behavior data with the predicted risk knowledge point.

[0025] In a possible implementation of the first aspect, the sample learning data further includes risk knowledge point feature data corresponding to the template multi-source user behavior data, and the optimization evaluation unit performs advanced risk prediction on the template multi-source user behavior data according to the template feature interpretation data to generate target predicted risk knowledge points, including:

[0026] Generate a feature explanation data sequence according to the risk knowledge point feature data and the template feature explanation data;

[0027] Selecting target feature explanation data from the feature explanation data sequence;

[0028] The optimization evaluation unit is used to interpret the data based on the target features, perform advanced risk prediction on the template multi-source user behavior data, and generate target prediction risk knowledge points.

[0029] In a possible implementation of the first aspect, the utilizing the optimization evaluation unit to perform advanced risk prediction on the template multi-source user behavior data according to the target feature interpretation data to generate target prediction risk knowledge points includes:

[0030] Generating a template attention weight distribution reflecting the feature influence state weights of the template multi-source user behavior data based on the target feature interpretation data;

[0031] Blending the template multi-source user behavior data and the template attention weight distribution to generate blended multi-source user behavior data;

[0032] An optimization evaluation unit is used to perform advanced risk prediction on the template multi-source user behavior data based on the blended multi-source user behavior data to generate target predicted risk knowledge points.

[0033] In a possible implementation of the first aspect, the step of generating, based on the target feature interpretation data, a template attention weight distribution reflecting the feature influence state weight of the template multi-source user behavior data includes:

[0034] Initialize an attention mechanism model, the attention mechanism model is used to receive target feature explanation data as input and output attention weight distribution, the attention mechanism model is constructed based on a deep learning framework;

[0035] In the attention mechanism model, an attention function is defined, the attention function is used to calculate the importance of each feature in the target feature explanation data to risk prediction, and the attention function uses one of the mechanisms such as dot product attention, additive attention or self-attention;

[0036] Inputting the target feature explanation data into the attention mechanism model, calculating the attention weight of each feature of the target feature explanation data for risk prediction through the attention function in the attention mechanism model, wherein the attention weight reflects the importance of different features in risk prediction, and the attention function uses one of the mechanisms such as dot product attention, additive attention or self-attention;

[0037] The calculated attention weights are normalized to generate a template attention weight distribution that reflects the feature influence state weights of the template's multi-source user behavior data.

[0038] In a possible implementation of the first aspect, the step of blending the template multi-source user behavior data and the template attention weight distribution to generate blended multi-source user behavior data includes:

[0039] Applying the weight values ​​in the template attention weight distribution to the corresponding feature values ​​of the template multi-source user behavior data to weight each feature value with its corresponding weight value to generate a weighted feature value;

[0040] The weighted feature values ​​are linearly combined or concatenated to generate a blended feature vector, and the blended feature vector is integrated with other unweighted feature values ​​in the template multi-source user behavior data to generate the blended multi-source user behavior data.

[0041] In a possible implementation of the first aspect, after performing advanced risk prediction on the multi-source user behavior data based on the blended data to generate second risk prediction data for the multi-source user behavior data, the method further includes:

[0042] Based on the second risk prediction data, target multi-source user behavior data matching the target financial platform risk knowledge point corresponding to the multi-source user behavior data are retrieved, and sample aggregation and collection of the target multi-source user behavior data are performed to facilitate neural network model learning; or,

[0043] Based on the second risk prediction data, adding the multi-source user behavior data to a knowledge point database that matches the target financial platform risk knowledge point corresponding to the multi-source user behavior data; or,

[0044] Based on the second risk prediction data, risk warning push information of the target financial platform risk knowledge point corresponding to the multi-source user behavior data is generated.

[0045] On the other hand, an embodiment of the present application also provides a financial platform risk prediction system based on multi-source user behavior data, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0046] Based on the above aspects, the embodiments of the present application realize the accurate prediction and interpretation of the user behavior risks of the financial platform. First, the multi-source user behavior data is processed using graph autoencoding technology to generate graph autoencoding structure data containing rich user behavior patterns, providing a comprehensive and accurate data basis for risk prediction. Secondly, the first risk prediction data is generated through preliminary risk prediction, including risk knowledge points and probability scores, providing immediate risk warnings for the financial platform. Furthermore, based on the influencing feature domain and probability scores of the graph autoencoding structure data, the influence weights of different features on risk prediction are determined, making the risk prediction more accurate. At the same time, by extracting graph autoencoders that meet the target feature requirements from multiple graph autoencoders, a target graph autoencoder is generated, and the encoded data is aggregated based on the influence weights of its output to generate feature interpretation data, providing a detailed explanation for the risk prediction results. Finally, the attention weight distribution is generated based on the feature interpretation data, and it is blended with the original multi-source user behavior data to generate blended data. Advanced risk prediction based on fusion data generates more accurate and specific second risk prediction data, characterizing the risk knowledge points of the target financial platform, which not only improves the accuracy of risk prediction, but also provides strong support for risk management and decision-making of the financial platform. Therefore, by combining graph autoencoding technology and attention mechanism, in-depth analysis and accurate prediction of multi-source user behavior data are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the execution flow of the financial platform risk prediction method based on multi-source user behavior data provided in an embodiment of the present application.

[0048] Figure 2 This is a schematic diagram of the hardware architecture of a financial platform risk prediction system based on multi-source user behavior data provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a financial platform risk prediction method based on multi-source user behavior data provided by an embodiment of the present application. The financial platform risk prediction method based on multi-source user behavior data is introduced in detail below.

[0050] Step S110, obtaining multi-source user behavior data indicated by the financial platform risk prediction task, and performing graph autoencoding processing on the multi-source user behavior data to generate graph autoencoding structure data corresponding to the multi-source user behavior data. The graph autoencoding structure data includes graph autoencoding structure data of multiple encoding dimensions corresponding to each graph autoencoder generated by performing graph autoencoding processing on the multi-source user behavior data according to multiple graph autoencoders.

[0051] In detail, the multi-source user behavior data refers to user behavior records from different data sources, reflecting the various behaviors of users on the financial platform, such as transaction records, browsing history, search queries, etc. These data come from diverse sources and contain rich user behavior information, which is of great significance for analyzing user behavior patterns and predicting potential risks.

[0052] For example, the multi-source user behavior data includes:

[0053] Transaction records: Details of users’ transactions on financial platforms, including transaction time, transaction amount, transaction object, etc.

[0054] Browsing history: the user’s browsing history on the financial platform, such as the financial product pages viewed, browsing time, etc.

[0055] Search query: The search keywords that users conduct on financial platforms reflect their interests and needs.

[0056] Graph autoencoder technology is an unsupervised learning method based on graph structure, which is used to learn low-dimensional representation of graph data. Through graph autoencoders, complex graph structure data can be converted into low-dimensional vectors, which is convenient for subsequent analysis and processing. Graph autoencoders are implemented in two parts: encoder and decoder. The encoder maps graph data to low-dimensional space, and the decoder attempts to reconstruct the original graph data from the low-dimensional representation. For example, in the financial field, graph autoencoder technology can be used to learn low-dimensional representation of user behavior graph data. For example, the transaction relationship and browsing relationship between users are constructed into graph structure data, and the low-dimensional behavior characteristics of users are learned through graph autoencoders to facilitate subsequent risk prediction analysis.

[0057] The graph autoencoding structure data is a low-dimensional vector representation obtained by processing multi-source user behavior data through graph autoencoding technology, which contains complex patterns and potential relationships of user behavior, and is convenient for subsequent risk prediction models to analyze and process. For example, suppose that multi-source user behavior data describes the transaction and browsing behavior of users on a financial platform. After processing these data through graph autoencoding technology, the graph autoencoding structure data of each user can be obtained. These data represent the user's behavioral characteristics in the form of vectors, such as transaction frequency, browsing preferences, etc.

[0058] The graph autoencoder is the core component of the graph autoencoder technology, which is used to map graph structure data to a low-dimensional space. Each graph autoencoder is responsible for encoding user behavior data from different dimensions and generating corresponding graph autoencoder structure data. For example, in the financial field, multiple graph autoencoders can be constructed to process multi-source user behavior data. For example, one graph autoencoder focuses on transaction behavior features, and another graph autoencoder focuses on browsing behavior features. Through the combination of multiple graph autoencoders, the multi-dimensional features of user behavior can be fully captured.

[0059] Thus, in this embodiment, the server receives a financial platform risk prediction task, which requires analyzing the user's behavior data on the financial platform to predict potential risks. In detail, the server first obtains multi-source user behavior data from multiple data sources (such as user transaction records, browsing history, search queries, etc.). Then, the server uses graph autoencoding technology to process these multi-source user behavior data. Specifically, the server constructs a model containing multiple graph autoencoders, each of which is responsible for encoding user behavior data from different dimensions. Through this step, the server generates graph autoencoding structure data corresponding to the multi-source user behavior data, which reflects the complex patterns and potential relationships of user behavior.

[0060] Step S120, performing preliminary risk prediction on the multi-source user behavior data based on the graph auto-encoding structure data, generating first risk prediction data of the multi-source user behavior data, wherein the first risk prediction data includes financial platform risk knowledge points and probability scores corresponding to the multi-source user behavior data.

[0061] In detail, the preliminary risk prediction is a risk assessment of user behavior based on the graph autoencoder structure data. Through the preliminary risk prediction, possible risk points can be identified and corresponding probability scores can be given. For example, after obtaining the graph autoencoder structure data of the user, the risk assessment unit can be used to perform preliminary risk prediction on these data. For example, by analyzing the user's transaction frequency, transaction amount and other characteristics, the risk assessment unit may predict that the user has an abnormal transaction risk and give a corresponding probability score.

[0062] That is, in this embodiment, after obtaining the graph auto-encoding structure data, the server further uses these data to make a preliminary risk prediction of the user behavior. A risk assessment unit in the server receives the graph auto-encoding structure data as input and runs a prediction algorithm to identify possible risk points. During this process, the risk assessment unit analyzes the patterns in the graph auto-encoding structure data and compares them with known risk patterns. Finally, the risk assessment unit outputs the first risk prediction data, which includes the financial platform risk knowledge points that may be triggered by user behavior and the probability score of each risk point.

[0063] Step S130, based on the influence feature domain of the graph autoencoder structure data and the probability score, determine the influence weight of the graph autoencoder structure data on the risk prediction of the multi-source user behavior data. The influence feature domain is the feature domain of the perception node in the graph autoencoder structure data.

[0064] In detail, the influencing feature domain refers to the set of features in the graph autoencoder structure data that have a significant impact on the risk prediction results, reflecting the key patterns or anomalies in user behavior. For example, in the graph autoencoder structure data, certain features may have a significant impact on the risk prediction results. For example, features such as the user's abnormal transaction frequency and the degree of association with high-risk users may be considered as part of the influencing feature domain, and these features will be focused on and analyzed during the risk prediction process.

[0065] The impact weight is used to quantify the importance of different features in the graph autoencoder structure data to the risk prediction results. By calculating the impact weight, it can be clarified which features play a key role in risk prediction, providing more targeted information for subsequent risk assessment. For example, in the risk prediction process, the impact weight can be calculated by analyzing the impact feature domain and probability score of the graph autoencoder structure data. For example, if a feature has been proven to be highly correlated with the risk point in multiple risk predictions, then the impact weight of the feature will be assigned a higher value, indicating its importance in risk prediction.

[0066] That is, in this embodiment, in order to more accurately assess the risk, the server needs to determine which parts of the graph autoencoder structure data have the greatest impact on risk prediction. Therefore, the server calculates the contribution of each perception node to risk prediction, that is, the influence weight, based on the influence feature domain of the graph autoencoder structure data (i.e., the feature domain of the perception node) and the previously calculated probability score. This process helps the server understand which user behavior features are most critical in risk prediction.

[0067] Step S140, extracting a graph autoencoder that meets the target feature requirements from the multiple graph autoencoders to generate a target graph autoencoder, and based on the influence weights corresponding to the graph autoencoder structure data output by the target graph autoencoder, aggregating the graph autoencoder structure data of multiple encoding dimensions of the target graph autoencoder to generate feature interpretation data that describes the multi-source user behavior data as the corresponding financial platform risk knowledge points.

[0068] In detail, the target graph autoencoder is a graph autoencoder selected from multiple graph autoencoders and meets the specific target feature requirements. These target graph autoencoders can capture and represent the user behavior features that are most important for risk prediction tasks. For example, suppose that in the risk prediction task of a financial platform, five graph autoencoders are used to process multi-source user behavior data respectively, and these five graph autoencoders each focus on different user behavior feature dimensions. By evaluating the performance of each graph autoencoder in preliminary risk prediction, it is found that two of the graph autoencoders perform particularly well in identifying fraud risks. Therefore, these two graph autoencoders are used as target graph autoencoders for subsequent advanced risk prediction.

[0069] The feature explanation data is generated by aggregating the graph autoencoder structure data of multiple encoding dimensions of the target graph autoencoder, and is used to describe the correlation between multi-source user behavior data and risk knowledge points of the financial platform, providing a detailed explanation of the risk prediction results, and helping to understand which user behavior characteristics have an important impact on the risk prediction results. For example, assume that the target graph autoencoder generates graph autoencoder structure data of five encoding dimensions, each dimension representing a specific aspect of user behavior. By calculating the impact weight of each dimension and multiplying it with the corresponding encoding data and then aggregating it, the feature explanation data is obtained. The feature explanation data may indicate that the user's abnormal transaction frequency and correlation with high-risk users are two key features that trigger fraud risks.

[0070] That is, after determining the influence weights, the server begins to screen out those graph autoencoders that meet specific feature requirements from multiple graph autoencoders. These specific feature requirements can be set based on previous experience or domain knowledge, aiming to identify the most indicative user behavior features for risk prediction. Once the target graph autoencoder is selected, the server aggregates the outputs of these target graph autoencoders using the influence weights corresponding to the graph autoencoder structure data output by them. This aggregation process integrates information from multiple encoding dimensions to generate a comprehensive feature explanation data that can more comprehensively describe the relationship between user behavior and risk knowledge points of the financial platform.

[0071] Step S150, generating an attention weight distribution reflecting the feature influence state weight of the multi-source user behavior data based on the feature interpretation data, and blending the multi-source user behavior data and the attention weight distribution to generate blended data, performing advanced risk prediction on the multi-source user behavior data based on the blended data, and generating second risk prediction data for the multi-source user behavior data, wherein the second risk prediction data represents the target financial platform risk knowledge points corresponding to the multi-source user behavior data.

[0072] In detail, the attention weight distribution is generated based on the feature explanation data, which reflects the importance or degree of influence of different features in multi-source user behavior data on the risk prediction results, helping the model to pay more attention to those features that have a significant impact on the risk prediction results in advanced risk prediction. For example, after obtaining the feature explanation data, the importance score of each feature in risk prediction can be calculated and normalized to obtain the attention weight distribution. For example, if the frequency of abnormal transactions and the association with high-risk users are given higher weights in the feature explanation data, then they will also have larger weight values ​​in the attention weight distribution, indicating that these two features should be given priority attention in advanced risk prediction.

[0073] The blended data is generated by blending the original multi-source user behavior data with the attention weight distribution. This process is equivalent to weighting the original data, so that in advanced risk prediction, more attention is paid to those features that have a significant impact on the risk prediction results. For example, suppose there is a user behavior data vector containing multiple features, and a corresponding attention weight distribution. By multiplying each feature in the user behavior data vector with the corresponding attention weight, a weighted user behavior data vector is obtained, namely the blended data. The blended data not only retains all the feature information of the original data, but also emphasizes those features that are more important to the risk prediction results through the attention weight.

[0074] Advanced risk prediction is a more in-depth risk assessment based on the preliminary risk prediction. It uses fusion data and more complex models or algorithms to generate more accurate risk prediction results. For example, in the preliminary risk prediction stage, some possible risk points have been identified and probability scores have been given. However, these prediction results may not be accurate enough. Therefore, advanced risk prediction is performed using fusion data and the optimized assessment unit. By deeply analyzing the key features in the fusion data and the correlation between them, the optimized assessment unit can generate more accurate and specific risk prediction results, such as specific risk types, risk levels, and possible losses.

[0075] That is, in this embodiment, at the end, the server uses the feature interpretation data to generate an attention weight distribution, which reflects the importance of different features in the multi-source user behavior data to risk prediction. Then, the server blends the original multi-source user behavior data with this attention weight distribution to generate a new blended data, which not only contains the original user behavior information, but also incorporates the weight information of key features in risk prediction. Based on this blended data, the server performs advanced risk prediction and generates second risk prediction data. These second risk prediction data more accurately point out the specific financial platform risk knowledge points that may be triggered by user behavior, providing strong support for the risk management of the financial platform.

[0076] Based on the above steps, the embodiment of the present application realizes the accurate prediction and interpretation of the user behavior risk of the financial platform. First, the multi-source user behavior data is processed using graph autoencoding technology to generate graph autoencoding structure data containing rich user behavior patterns, providing a comprehensive and accurate data basis for risk prediction. Secondly, the first risk prediction data is generated through preliminary risk prediction, including risk knowledge points and probability scores, providing immediate risk warnings for the financial platform. Furthermore, based on the influencing feature domain and probability scores of the graph autoencoding structure data, the influence weights of different features on risk prediction are determined, making the risk prediction more accurate. At the same time, by extracting graph autoencoders that meet the target feature requirements from multiple graph autoencoders, a target graph autoencoder is generated, and the encoded data is aggregated based on the influence weights of its output to generate feature interpretation data, providing a detailed explanation for the risk prediction results. Finally, the attention weight distribution is generated based on the feature interpretation data, and it is blended with the original multi-source user behavior data to generate blended data. Advanced risk prediction based on fusion data generates more accurate and specific second risk prediction data, characterizing the risk knowledge points of the target financial platform, which not only improves the accuracy of risk prediction, but also provides strong support for risk management and decision-making of the financial platform. Therefore, by combining graph autoencoding technology and attention mechanism, in-depth analysis and accurate prediction of multi-source user behavior data are achieved.

[0077] In a possible implementation, step S120 includes:

[0078] A risk assessment unit is used to perform preliminary risk prediction on the multi-source user behavior data according to the graph auto-encoding structure data to generate first risk prediction data of the multi-source user behavior data.

[0079] In this embodiment, the server uses the risk assessment unit to perform preliminary risk prediction on the multi-source user behavior data according to the graph self-encoding structure data, and generates first risk prediction data of these graph self-encoding structure data.

[0080] In detail, the server has collected and integrated multi-source user behavior data from multiple data sources (such as transaction systems, user browsing records, search logs, etc.). Through graph autoencoding technology, the server has converted these multi-source data into graph autoencoding structure data, which reflects the complex patterns and potential correlations of user behaviors.

[0081] Next, the risk assessment unit of the server receives these graph autoencoding structure data as input. The risk assessment unit runs a pre-trained risk assessment model inside, which can identify possible risk patterns in user behavior. The risk assessment model begins to analyze the graph autoencoding structure data to identify potential risks. After a series of calculations and analyses, the first risk prediction data is output. These first risk prediction data include financial platform risk knowledge points that may be triggered by user behavior and the probability score of each risk point. As a result, the server stores the first risk prediction data in the database and prepares for the next step of advanced risk prediction. At the same time, these preliminary prediction results can also serve as immediate risk warnings to notify the operators of the financial platform to pay attention to certain users or transaction behaviors.

[0082] Step S150 includes:

[0083] An optimization evaluation unit is used to perform advanced risk prediction on the multi-source user behavior data based on the blended data to generate second risk prediction data for the multi-source user behavior data.

[0084] In this embodiment, the server uses the optimization evaluation unit to perform advanced risk prediction on the multi-source user behavior data based on the fusion data to generate second risk prediction data of the multi-source user behavior data.

[0085] In detail, after completing the preliminary risk prediction, the server has determined which parts of the graph autoencoder structure data have the greatest impact on risk prediction, and calculated the impact weights accordingly. The server selects encoders that meet the target feature requirements from multiple graph autoencoders, and uses the impact weights of their outputs to aggregate the encoded data to generate feature interpretation data. The server uses the feature interpretation data to generate an attention weight distribution, which reflects the importance of different features in multi-source user behavior data to risk prediction.

[0086] Next, the server blends the original multi-source user behavior data with the attention weight distribution to obtain blended data, which not only contains the original information but also highlights key risk features. The server's optimization and evaluation unit receives the blended data as input. A more sophisticated risk assessment model runs inside the optimization and evaluation unit, aiming to conduct a deeper risk analysis. The optimization and evaluation model begins to process the blended data, identifying risk points more accurately by analyzing the weighted user behavior features. After the calculation and analysis process, the second risk prediction data is output, which more accurately points out the specific financial platform risk knowledge points that may be triggered by user behavior. As a result, the server stores the second risk prediction data in the database, and can generate risk reports or warning information based on these data.

[0087] Financial platforms can use these second risk prediction data to make risk management decisions, such as adjusting user credit limits, strengthening transaction monitoring, etc. In addition, these second risk prediction data can also be used to further optimize risk assessment models and improve the accuracy of future risk predictions.

[0088] In a possible implementation manner, before the step of using the risk assessment unit to perform preliminary risk prediction on the multi-source user behavior data according to the graph autoencoding structure data to generate first risk prediction data for the multi-source user behavior data, the method further includes:

[0089] Step S101, obtaining sample learning data, wherein the sample learning data includes template multi-source user behavior data and its corresponding annotated risk knowledge point data and sample annotation probability.

[0090] In this embodiment, the server obtains a set of sample learning data from the database, which includes template multi-source user behavior data, such as user transaction records, browsing history, search records, etc., as well as the labeled risk knowledge point data and sample labeling probability corresponding to these template multi-source user behavior data. The labeled risk knowledge point data is the possible risk points marked by experts based on user behavior, such as fraudulent behavior, abnormal transactions, etc., and the sample labeling probability is the possibility of these risk points occurring.

[0091] Step S102, using the risk assessment unit to perform graph auto-encoding processing on the template multi-source user behavior data, generate template graph auto-encoding structure data corresponding to the template multi-source user behavior data, and perform preliminary risk prediction on the template multi-source user behavior data based on the template graph auto-encoding structure data to generate predicted risk knowledge points and predicted probability scores.

[0092] In this embodiment, the server inputs the acquired template multi-source user behavior data into the risk assessment unit. The risk assessment unit first performs graph self-encoding processing on these template multi-source user behavior data to convert complex user behavior data into structured graph self-encoding data. This processing makes the data easier to perform risk prediction analysis.

[0093] Based on the converted template graph autoencoding structure data, the risk assessment unit performs preliminary risk prediction, identifies possible risk knowledge points, such as abnormal transaction patterns or suspicious user behaviors, and assigns a prediction probability score to each predicted risk knowledge point, indicating the possibility of the risk point occurring.

[0094] Step S103, based on the influence feature domain of the template graph auto-encoding structure data and the prediction probability score, calculate the template influence weight of the template graph auto-encoding structure data for risk prediction of the template multi-source user behavior data.

[0095] The server further analyzes the influencing feature domains and predicted probability scores of the template graph self-encoding structure data, and calculates the importance or weight of these template graph self-encoding structure data for risk prediction. This process helps to understand which user behavior characteristics have the greatest impact on risk prediction results.

[0096] Step S104, aggregating the template graph auto-encoding structure data based on the template influence weight, and generating template feature interpretation data representing the matching of the template multi-source user behavior data with the predicted risk knowledge point.

[0097] According to the calculated template influence weight, the server performs weighted aggregation on the template graph self-encoding structure data to generate a new set of data, namely, template feature explanation data. This set of template feature explanation data more intuitively describes the matching relationship between user behavior and predicted risk knowledge points.

[0098] Step S105 , using the optimization evaluation unit to perform advanced risk prediction on the template multi-source user behavior data according to the template feature interpretation data, and generate target predicted risk knowledge points.

[0099] The server inputs the template feature interpretation data into the optimization evaluation unit for more refined risk prediction analysis, which aims to identify more specific and accurate risk knowledge points, namely, target prediction risk knowledge points.

[0100] Step S106, calculating the characteristic distance between the predicted risk knowledge point and the annotated risk knowledge point data to generate a first training cost value, and calculating the characteristic distance between the target predicted risk knowledge point and the annotated risk knowledge point data to generate a second training cost value.

[0101] The server calculates the feature distance between the predicted risk knowledge point and the annotated risk knowledge point to obtain the first training cost value. At the same time, it also calculates the feature distance between the target predicted risk knowledge point and the annotated risk knowledge point to obtain the second training cost value. These two cost values ​​reflect the prediction accuracy of the risk assessment unit and the optimization assessment unit respectively.

[0102] Step S107, updating the weight and bias information of the risk assessment unit based on the first training cost value, and updating the weight and bias information of the optimization assessment unit based on the second training cost value, until the training convergence requirement is met.

[0103] Based on the first training epoch value, the server updates the weights and bias information of the risk assessment unit to improve its prediction accuracy. Similarly, based on the second training epoch value, the server updates the weights and bias information of the optimization assessment unit. This process is repeated until the training results of the two assessment units meet the preset training convergence requirements, which means that the gap between the prediction results and the actual risk points has been narrowed to an acceptable range, and the model has achieved a high prediction accuracy.

[0104] In a possible implementation, the template graph autoencoding structure data includes the template graph autoencoding structure data of multiple encoding dimensions corresponding to each graph autoencoder generated by performing graph autoencoding processing on the template multi-source user behavior data according to multiple graph autoencoders of the risk assessment unit. Step S104 includes:

[0105] Step S1041, extracting a graph autoencoder that meets the target feature requirements from multiple graph autoencoders to generate a target graph autoencoder.

[0106] Step S1042, based on the template influence weight corresponding to the template graph autoencoder structure data output by the target graph autoencoder, the template graph autoencoder structure data of multiple encoding dimensions of the target graph autoencoder are aggregated to generate template feature interpretation data describing the matching of the template multi-source user behavior data with the predicted risk knowledge point.

[0107] In this embodiment, the graph autoencoder is a neural network structure used to encode graph structure data into low-dimensional vectors to facilitate subsequent analysis and processing; the template graph autoencoder structure data is structured data obtained after multiple graph autoencoders process the template multi-source user behavior data; the template influence weight reflects the importance of different graph autoencoder structure data for risk prediction.

[0108] The server first analyzes the template graph autoencoder structure data output by multiple graph autoencoders. These template graph autoencoder structure data are obtained by graph autoencoding the template multi-source user behavior data. Each graph autoencoder generates data of different encoding dimensions. The server evaluates these graph autoencoders according to the preset target feature requirements, such as encoding efficiency, feature expression ability, etc.

[0109] After evaluation, the server selected several graph autoencoders with excellent performance and meeting the target feature requirements, and marked them as target graph autoencoders. These target graph autoencoders will play an important role in subsequent risk prediction and feature interpretation.

[0110] After determining the target graph autoencoders, the server begins to pay attention to the template graph autoencoder structure data output by these encoders and their corresponding template influence weights. The template influence weights are calculated in the preliminary risk prediction stage and reflect the importance of different graph autoencoder structure data to the risk prediction results.

[0111] Based on these weight information, the server performs weighted aggregation on the template graph autoencoder structure data of multiple encoding dimensions of the target graph autoencoder. Specifically, according to the weight corresponding to the data of each encoding dimension, they are combined in a certain proportion to form a new data representation that integrates all important features.

[0112] This process is equivalent to a feature fusion and refinement of the original multi-source user behavior data, so that the generated data is more focused on the features related to the predicted risk knowledge points. In this way, the server can more accurately describe the matching relationship between the template multi-source user behavior data and the predicted risk knowledge points.

[0113] After the above-mentioned weighted aggregation processing, the server obtains a new set of data representations, namely, template feature interpretation data. This set of data not only integrates the output information of multiple graph autoencoders, but also highlights the contribution of key features to risk prediction through weight adjustment.

[0114] Therefore, this set of template feature explanation data can more intuitively reflect the correlation between user behavior and predicted risk knowledge points. For example, in anti-fraud scenarios, it may reveal the close connection between certain specific user transaction behaviors and fraud risks; in credit risk assessment, it may point out which user profile information is highly correlated with default risks, etc.

[0115] In short, through this series of processing and analysis steps, the server is able to generate more explanatory and predictive data representations, providing strong support for subsequent risk management work.

[0116] In a possible implementation, the sample learning data further includes risk knowledge point feature data corresponding to the template multi-source user behavior data, and step S105 includes:

[0117] Step S1051, generating a feature interpretation data sequence according to the risk knowledge point feature data and the template feature interpretation data.

[0118] Step S1052: Select target feature interpretation data from the feature interpretation data sequence.

[0119] Step S1053, using the optimization evaluation unit to interpret the data based on the target features, perform advanced risk prediction on the template multi-source user behavior data, and generate target prediction risk knowledge points.

[0120] In this embodiment, risk knowledge point feature data refers to feature information related to risk points, template feature interpretation data is obtained through previous steps and is used to describe the relationship between user behavior and predicted risk points, and advanced risk prediction is a more in-depth analysis and prediction based on preliminary risk prediction.

[0121] The server first obtains the risk knowledge point feature data corresponding to the template multi-source user behavior data, which may include risk points that have appeared in history, user behavior patterns related to risk points, etc. Then, the server combines these risk knowledge point feature data with the previously generated template feature interpretation data.

[0122] Specifically, the server may integrate these two types of data through some algorithm (such as feature splicing, weighted fusion, etc.) to form a new data sequence, namely the feature interpretation data sequence, which not only contains the characteristics of user behavior, but also incorporates feature information directly related to risk points, thereby providing a richer data foundation for subsequent advanced risk prediction.

[0123] After generating feature explanation data sequences, the server needs to screen these feature explanation data sequences to select the most representative data for advanced risk prediction. This process may involve evaluation of data quality, relevance, importance, etc.

[0124] For example, the server may use some statistical methods or machine learning algorithms to evaluate the contribution of each feature to risk prediction, and then select those features with higher contribution as target feature interpretation data, which are believed to be able to more effectively reveal the relationship between user behavior and potential risks.

[0125] Finally, the server inputs the selected target feature interpretation data into the optimization evaluation unit, which can be a more complex machine learning model or deep learning network, which can perform more in-depth analysis and processing on the input data. Through the processing of the optimization evaluation unit, the server can generate more accurate risk prediction results, namely, target prediction risk knowledge points, which can include specific risk types, probability of risk occurrence, possible losses, etc. This information has important guiding significance for subsequent risk management and decision-making.

[0126] For example, in a credit risk assessment scenario, the server may first integrate the user's credit history, income status, debt status and other risk knowledge point feature data with the user's borrowing behavior, repayment record and other template feature interpretation data. Then, by screening and analyzing these data, the most representative target feature interpretation data is selected. Finally, the optimized evaluation unit is used to perform advanced risk prediction on these data, generating specific prediction results about the user's credit risk, such as the probability of default, possible losses, etc. This information will help financial institutions make more accurate credit decisions.

[0127] In a possible implementation, step S1053 includes:

[0128] Step S1053-1, based on the target feature interpretation data, generate a template attention weight distribution that reflects the feature influence state weight of the template multi-source user behavior data.

[0129] In this embodiment, after obtaining the target feature interpretation data, the server begins to analyze the impact of these target feature interpretation data on the template multi-source user behavior data. In order to quantify this impact, the server uses an attention mechanism to calculate the importance or weight of each feature.

[0130] Specifically, the server can process the target feature interpretation data through a specific algorithm (such as the attention layer in a neural network) to generate a template attention weight distribution that reflects the degree of influence of different features in the user behavior data. For example, certain user behavior features (such as high-frequency trading, abnormal login time, etc.) may be more important in risk prediction and therefore will be given a higher attention weight.

[0131] In this way, the server can clearly identify which features play a key role in risk prediction and provide more targeted information for subsequent risk assessment.

[0132] Step S1053-2, blending the template multi-source user behavior data and the template attention weight distribution to generate blended multi-source user behavior data.

[0133] In this embodiment, after obtaining the template attention weight distribution, the server needs to blend the template attention weight distribution with the original template multi-source user behavior data. The purpose of blending is to strengthen those important features and relatively weaken those features that have little impact on risk prediction.

[0134] The specific fusion process may include multiplying each user behavior feature with its corresponding attention weight to obtain a weighted user behavior data representation, so that important features will play a greater role in subsequent risk prediction.

[0135] Through the fusion operation, the server generates fused multi-source user behavior data, which not only retains the characteristic information of the original user behavior, but also incorporates the importance weight of the feature.

[0136] Step S1053-3: Using the optimization evaluation unit to perform advanced risk prediction on the template multi-source user behavior data based on the blended multi-source user behavior data, to generate target predicted risk knowledge points.

[0137] Finally, the server inputs the blended multi-source user behavior data into the optimization and evaluation unit for advanced risk prediction. The optimization and evaluation unit can be a complex machine learning model, such as a deep learning network, which can process the weighted user behavior data and output more accurate risk prediction results.

[0138] In this process, the optimization and evaluation unit will comprehensively consider the various characteristics and weights of the integrated data, and generate target prediction risk knowledge points through complex calculations and analysis. These target prediction risk knowledge points can include specific risk categories, risk levels, potential losses and other information, providing strong data support for risk management and decision-making.

[0139] For example, in a fraud detection scenario on an e-commerce platform, the server may analyze the user's shopping behavior, payment behavior and other characteristics, and combine them with the attention weight distribution to predict which user behaviors may have fraud risks. The target prediction risk knowledge points generated in the end can help the platform take preventive measures in a timely manner to reduce potential economic losses.

[0140] In a possible implementation, step S1053-1 includes:

[0141] Step S1053-11, initialize an attention mechanism model, which is used to receive target feature explanation data as input and output attention weight distribution, and the attention mechanism model is constructed based on a deep learning framework.

[0142] In this embodiment, the server first initializes an attention mechanism model built on a deep learning framework, which is designed to receive target feature interpretation data as input and output an attention weight distribution. The deep learning framework provides tools for building and training complex neural network models, enabling the server to efficiently implement the attention mechanism.

[0143] For example, in deep learning frameworks such as TensorFlow or PyTorch, the server can define key components such as the model's hierarchy, activation functions, and optimizers to build an attention mechanism model suitable for processing target feature explanation data.

[0144] Step S1053-12, in the attention mechanism model, an attention function is defined, which is used to calculate the importance of each feature in the target feature explanation data to risk prediction, and the attention function uses one of the mechanisms such as dot product attention, additive attention or self-attention.

[0145] In the attention mechanism model, the server defines an attention function, which calculates the importance of each feature in the target feature explanation data for risk prediction. To achieve this, the server can choose one of the mechanisms such as dot product attention, additive attention or self-attention to construct this attention function.

[0146] Taking dot product attention as an example, the server can calculate the correlation score between each feature and the target risk prediction, which reflects the importance of the feature in risk prediction. Additive attention and self-attention mechanisms use different calculation methods to evaluate the importance of features.

[0147] Step S1053-13, input the target feature explanation data into the attention mechanism model, and calculate the attention weight of each feature of the target feature explanation data for risk prediction through the attention function in the attention mechanism model. The attention weight reflects the importance of different features in risk prediction. The attention function uses one of the mechanisms such as dot product attention, additive attention or self-attention.

[0148] When the server inputs the target feature explanation data into the attention mechanism model, the attention function in the model begins to calculate the attention weights of each feature for risk prediction. These weights reflect the importance of different features in risk prediction.

[0149] Specifically, for each feature, the attention function will give a score based on its relevance to the target risk prediction. The higher the score, the greater the impact of the feature on risk prediction, and therefore the greater its attention weight.

[0150] Step S1053-14, normalize the calculated attention weights to generate a template attention weight distribution that reflects the feature influence state weights of the template multi-source user behavior data.

[0151] After calculating the attention weights of all features, the server can normalize these weights. The purpose of normalization is to ensure that the sum of all weights is 1, which facilitates subsequent data processing and interpretation.

[0152] Normalization can be achieved through the softmax function. The softmax function converts the original attention weight into a probability distribution, so that the weight of each feature is between 0 and 1, and the sum of all weights is 1. After normalization, the server obtains a template attention weight distribution that reflects the feature influence state weight of the template multi-source user behavior data. The template attention weight distribution can intuitively show which features play a key role in the risk prediction process and which features have relatively small effects, which has important guiding significance for subsequent risk prediction and feature selection.

[0153] In a possible implementation, step S1053-2 includes:

[0154] Step S1053-21, applying the weight values ​​in the template attention weight distribution to the corresponding feature values ​​of the template multi-source user behavior data, so as to weight each feature value with its corresponding weight value to generate a weighted feature value.

[0155] Step S1053-22, linearly combining or concatenating the weighted feature values ​​to generate a blended feature vector, integrating the blended feature vector with other unweighted feature values ​​in the template multi-source user behavior data to generate the blended multi-source user behavior data.

[0156] In this embodiment, after obtaining the template attention weight distribution, the server begins to apply the template attention weight distribution to the template multi-source user behavior data. Specifically, the server can traverse each feature value in the template multi-source user behavior data and multiply these feature values ​​with the corresponding weight values ​​in the template attention weight distribution.

[0157] For example, suppose there is a feature in the template multi-source user behavior data called "user login frequency", and the template attention weight distribution assigns this feature a weight value of 0.8 (indicating that this feature is more important in risk prediction). The server can multiply the actual value of the feature "user login frequency" by 0.8 to obtain the weighted feature value.

[0158] This process is performed for each feature in the template's multi-source user behavior data, ensuring that each feature is appropriately weighted according to its importance.

[0159] Through the above weighting process, the server obtains a set of weighted feature values. These weighted feature values ​​not only retain the information of the original features, but also incorporate the importance weights of the features, so that important features can play a greater role in subsequent risk predictions.

[0160] The server then processes the weighted eigenvalues ​​to generate a unified fused eigenvector, which can be achieved through linear combination (i.e., calculating the weighted sum of the weighted eigenvalues) or concatenation (connecting the weighted eigenvalues ​​in a certain order).

[0161] If you choose linear combination, the server can calculate the sum of all weighted feature values ​​to obtain a single value that combines all feature information. If you choose concatenation, the server will connect all weighted feature values ​​in the order of the original features into a longer feature vector.

[0162] Whether through linear combination or concatenation, the server will eventually obtain a fused feature vector, which integrates the information of each feature in the template's multi-source user behavior data and weights the features according to the attention weight.

[0163] Finally, the server can integrate this fused feature vector with other unweighted feature values ​​in the template multi-source user behavior data, which usually means merging the fused feature vector with other feature values ​​into a more complete user behavior data representation so that subsequent risk prediction models can comprehensively consider all relevant information to make decisions.

[0164] For example, in the user behavior analysis of an e-commerce platform, the server may weight features such as the user's browsing history and purchase history, and integrate them with unweighted features such as the user's geographic location and device information to obtain a more comprehensive and accurate user behavior portrait for subsequent tasks such as risk prediction and personalized recommendations.

[0165] In a possible implementation, after step S150, the method further includes:

[0166] Step S160, based on the second risk prediction data, retrieve the target multi-source user behavior data that matches the target financial platform risk knowledge point corresponding to the multi-source user behavior data, and perform sample aggregation and collection on the target multi-source user behavior data to facilitate neural network model learning.

[0167] In this embodiment, after obtaining the second risk prediction data, the server will use the second risk prediction data to retrieve other multi-source user behavior data that matches the risk knowledge points of the specific financial platform. For example, if the second risk prediction data indicates that a certain user behavior has a fraud risk, the server can search the database for other user behavior data with similar fraud risk characteristics.

[0168] For example, on an online payment platform, if the server detects that a user's transaction behavior is abnormal and there may be risks, based on this risk prediction data, the server can automatically retrieve other user data with similar transaction patterns in the historical database to further analyze and prevent potential risks.

[0169] After retrieving the target multi-source user behavior data that matches the risk knowledge points, the server can aggregate and organize these data to form a structured dataset, which will then be used to train or optimize the neural network model to improve the model's ability to predict future risks.

[0170] For example, in an insurance claims system, the server finds a series of claims suspected of insurance fraud based on risk prediction data. The server aggregates the data of these applications to form a data set specifically for training anti-fraud models. Through this data set, the neural network model can learn more about the characteristics of fraudulent behavior, thereby improving the ability to identify potential fraudulent behavior.

[0171] Alternatively, in step S170, based on the second risk prediction data, the multi-source user behavior data is added to a knowledge point database that matches the target financial platform risk knowledge point corresponding to the multi-source user behavior data.

[0172] When the server determines that a certain multi-source user behavior data matches a specific financial platform risk knowledge point, it will add this data to the corresponding knowledge point database. This can enrich the content of the database and provide more comprehensive data support for subsequent risk analysis and early warning.

[0173] For example, on a public trading platform, the server predicts that a user may be engaging in insider trading by analyzing the user's trading behavior. The server can add the user's trading data and related risk prediction data to the database of insider trading risk knowledge points, so that similar risky behaviors can be identified and prevented more quickly in the future.

[0174] Alternatively, step S180, based on the second risk prediction data, generates risk warning push information of the target financial platform risk knowledge point corresponding to the multi-source user behavior data.

[0175] Based on the second risk prediction data, the server can generate risk warning push information for risk knowledge points of specific financial platforms, which will be sent to relevant risk managers or systems in a timely manner so that they can take prompt measures to deal with potential risks.

[0176] For example, on an online lending platform, the server discovers through risk prediction that a borrower's repayment ability is abnormal and there may be a risk of default. The server can immediately generate a risk warning push message to notify the platform's risk management department. After receiving the information, the risk management department can quickly take measures, such as contacting the borrower to confirm the situation, adjusting the loan amount, or taking other necessary risk control measures.

[0177] Figure 2 The hardware structure of the financial platform risk prediction system 100 based on multi-source user behavior data for implementing the above-mentioned financial platform risk prediction method based on multi-source user behavior data provided in the embodiment of the present application is shown as follows: Figure 2 As shown, the financial platform risk prediction system 100 based on multi-source user behavior data may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0178] In one possible design, the financial platform risk prediction system 100 based on multi-source user behavior data can be a single server or a server group. The server group can be centralized or distributed (for example, the financial platform risk prediction system 100 based on multi-source user behavior data can be a distributed system). In some embodiments, the financial platform risk prediction system 100 based on multi-source user behavior data can be local or remote. For example, the financial platform risk prediction system 100 based on multi-source user behavior data can access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the financial platform risk prediction system 100 based on multi-source user behavior data can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the financial platform risk prediction system 100 based on multi-source user behavior data can be implemented on a financial platform risk prediction system based on multi-source user behavior data. As an example only, the financial platform risk prediction system based on multi-source user behavior data can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any combination thereof.

[0179] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the financial platform risk prediction system 100 based on multi-source user behavior data uses to execute or use to complete the exemplary method described in this application.

[0180] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the financial platform risk prediction method based on multi-source user behavior data in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0181] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned financial platform risk prediction system 100 based on multi-source user behavior data. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0182] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned financial platform risk prediction method based on multi-source user behavior data is implemented.

[0183] It should be noted that in order to simplify the description disclosed in the present application and thereby facilitate the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, various features are sometimes grouped into one embodiment, drawing or description thereof.

Claims

1. A financial platform risk prediction method based on multi-source user behavior data, characterized in that: The method comprises: Acquire multi-source user behavior data indicated by a financial platform risk prediction task, and perform graph autoencoding processing on the multi-source user behavior data to generate graph autoencoding structure data corresponding to the multi-source user behavior data; the graph autoencoding structure data includes graph autoencoding structure data of multiple encoding dimensions corresponding to each graph autoencoder generated by performing graph autoencoding processing on the multi-source user behavior data according to multiple graph autoencoders respectively; Performing preliminary risk prediction on the multi-source user behavior data according to the graph autoencoder structure data to generate first risk prediction data of the multi-source user behavior data, wherein the first risk prediction data includes financial platform risk knowledge points and probability scores corresponding to the multi-source user behavior data; Based on the influence feature domain of the graph autoencoder structure data and the probability score, determining the influence weight of the graph autoencoder structure data on the risk prediction of the multi-source user behavior data; the influence feature domain is the feature domain of the perception node in the graph autoencoder structure data; Extracting a graph autoencoder that meets the target feature requirements from the multiple graph autoencoders to generate a target graph autoencoder, aggregating the graph autoencoder structure data of multiple encoding dimensions of the target graph autoencoder based on the influence weights corresponding to the graph autoencoder structure data output by the target graph autoencoder, and generating feature interpretation data representing the description of the multi-source user behavior data as the corresponding financial platform risk knowledge point; An attention weight distribution reflecting the feature influence state weights of the multi-source user behavior data is generated based on the feature interpretation data, and the multi-source user behavior data and the attention weight distribution are blended to generate blended data. Advanced risk prediction is performed on the multi-source user behavior data based on the blended data to generate second risk prediction data for the multi-source user behavior data, wherein the second risk prediction data represents the target financial platform risk knowledge points corresponding to the multi-source user behavior data.

2. The financial platform risk prediction method based on multi-source user behavior data according to claim 1 is characterized in that: The performing preliminary risk prediction on the multi-source user behavior data according to the graph auto-encoding structure data to generate first risk prediction data of the multi-source user behavior data includes: Using a risk assessment unit to perform preliminary risk prediction on the multi-source user behavior data according to the graph autoencoding structure data, to generate first risk prediction data for the multi-source user behavior data; The step of performing advanced risk prediction on the multi-source user behavior data based on the blended data to generate second risk prediction data for the multi-source user behavior data includes: An optimization evaluation unit is used to perform advanced risk prediction on the multi-source user behavior data based on the blended data to generate second risk prediction data for the multi-source user behavior data.

3. The financial platform risk prediction method based on multi-source user behavior data according to claim 2 is characterized in that: Before the step of using a risk assessment unit to perform preliminary risk prediction on the multi-source user behavior data according to the graph autoencoding structure data to generate first risk prediction data for the multi-source user behavior data, the method further includes: Acquire sample learning data, wherein the sample learning data includes template multi-source user behavior data and its corresponding annotated risk knowledge point data and sample annotation probability; Using a risk assessment unit to perform graph autoencoding processing on the template multi-source user behavior data, generate template graph autoencoding structure data corresponding to the template multi-source user behavior data, and perform preliminary risk prediction on the template multi-source user behavior data based on the template graph autoencoding structure data to generate predicted risk knowledge points and predicted probability scores; Based on the influence feature domain of the template graph auto-encoding structure data and the prediction probability score, calculating the template influence weight of the template graph auto-encoding structure data for risk prediction of the template multi-source user behavior data; Aggregating the template graph auto-encoding structure data based on the template influence weight to generate template feature interpretation data representing the matching of the template multi-source user behavior data with the predicted risk knowledge point; Using the optimization evaluation unit to perform advanced risk prediction on the template multi-source user behavior data according to the template feature interpretation data to generate target prediction risk knowledge points; Calculating the characteristic distance between the predicted risk knowledge point and the annotated risk knowledge point data to generate a first training cost value, and calculating the characteristic distance between the target predicted risk knowledge point and the annotated risk knowledge point data to generate a second training cost value; The weight and bias information of the risk assessment unit are updated based on the first training cost value, and the weight and bias information of the optimization assessment unit are updated based on the second training cost value until the training convergence requirement is met.

4. The financial platform risk prediction method based on multi-source user behavior data according to claim 3 is characterized in that: The template graph autoencoding structure data includes the template graph autoencoding structure data of multiple encoding dimensions corresponding to each graph autoencoder generated by performing graph autoencoding processing on the template multi-source user behavior data according to the multiple graph autoencoders of the risk assessment unit, and the template graph autoencoding structure data is aggregated based on the template influence weight to generate template feature interpretation data representing the description of the template multi-source user behavior data matching the predicted risk knowledge point, including: Extract a graph autoencoder that meets the target feature requirements from multiple graph autoencoders to generate a target graph autoencoder; Based on the template influence weight corresponding to the template graph autoencoder structure data output by the target graph autoencoder, the template graph autoencoder structure data of multiple encoding dimensions of the target graph autoencoder are aggregated to generate template feature interpretation data representing the matching of the template multi-source user behavior data with the predicted risk knowledge point.

5. The financial platform risk prediction method based on multi-source user behavior data according to claim 3 is characterized in that: The sample learning data also includes risk knowledge point feature data corresponding to the template multi-source user behavior data, and the optimization evaluation unit performs advanced risk prediction on the template multi-source user behavior data according to the template feature interpretation data to generate target predicted risk knowledge points, including: Generate a feature explanation data sequence according to the risk knowledge point feature data and the template feature explanation data; Selecting target feature explanation data from the feature explanation data sequence; The optimization evaluation unit is used to interpret the data based on the target features, perform advanced risk prediction on the template multi-source user behavior data, and generate target prediction risk knowledge points.

6. The financial platform risk prediction method based on multi-source user behavior data according to claim 5 is characterized in that: The optimization evaluation unit performs advanced risk prediction on the template multi-source user behavior data according to the target feature interpretation data to generate target prediction risk knowledge points, including: Generating a template attention weight distribution reflecting the feature influence state weights of the template multi-source user behavior data based on the target feature interpretation data; Blending the template multi-source user behavior data and the template attention weight distribution to generate blended multi-source user behavior data; An optimization evaluation unit is used to perform advanced risk prediction on the template multi-source user behavior data based on the blended multi-source user behavior data to generate target predicted risk knowledge points.

7. The financial platform risk prediction method based on multi-source user behavior data according to claim 6 is characterized in that: The step of generating a template attention weight distribution reflecting the feature influence state weight of the template multi-source user behavior data based on the target feature interpretation data comprises: Initialize an attention mechanism model, the attention mechanism model is used to receive target feature explanation data as input and output attention weight distribution, the attention mechanism model is constructed based on a deep learning framework; In the attention mechanism model, an attention function is defined, the attention function is used to calculate the importance of each feature in the target feature explanation data to risk prediction, and the attention function uses one of the mechanisms such as dot product attention, additive attention or self-attention; Inputting the target feature explanation data into the attention mechanism model, calculating the attention weight of each feature of the target feature explanation data for risk prediction through the attention function in the attention mechanism model, wherein the attention weight reflects the importance of different features in risk prediction, and the attention function uses one of the mechanisms such as dot product attention, additive attention or self-attention; The calculated attention weights are normalized to generate a template attention weight distribution that reflects the feature influence state weights of the template's multi-source user behavior data.

8. The financial platform risk prediction method based on multi-source user behavior data according to claim 6 is characterized in that: The step of blending the template multi-source user behavior data and the template attention weight distribution to generate blended multi-source user behavior data includes: Applying the weight values ​​in the template attention weight distribution to the corresponding feature values ​​of the template multi-source user behavior data to weight each feature value with its corresponding weight value to generate a weighted feature value; The weighted feature values ​​are linearly combined or concatenated to generate a blended feature vector, and the blended feature vector is integrated with other unweighted feature values ​​in the template multi-source user behavior data to generate the blended multi-source user behavior data.

9. The financial platform risk prediction method based on multi-source user behavior data according to any one of claims 1 to 8, characterized in that: After performing advanced risk prediction on the multi-source user behavior data based on the blended data to generate second risk prediction data for the multi-source user behavior data, the method further includes: Based on the second risk prediction data, target multi-source user behavior data matching the target financial platform risk knowledge point corresponding to the multi-source user behavior data are retrieved, and sample aggregation and collection of the target multi-source user behavior data are performed to facilitate neural network model learning; or, Based on the second risk prediction data, adding the multi-source user behavior data to a knowledge point database that matches the target financial platform risk knowledge point corresponding to the multi-source user behavior data; or, Based on the second risk prediction data, risk warning push information of the target financial platform risk knowledge point corresponding to the multi-source user behavior data is generated.

10. A financial platform risk prediction system based on multi-source user behavior data, characterized in that: The financial platform risk prediction system based on multi-source user behavior data includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the financial platform risk prediction method based on multi-source user behavior data as described in any one of claims 1 to 9 above.

Images