Multi-source data scoring model construction method and device based on dynamic feature weight

Abstract

The invention discloses a multi-source data scoring model construction method and device based on dynamic feature weights, and the method comprises the steps: obtaining input data of a target user, the input data comprising financial data, communication behavior data and consumption habit data; determining a first weight of the financial data, a second weight of the communication behavior data and a third weight of the consumption habit data according to an available financial data dimension number of the financial data and a total dimension number of the input data; and processing the input data, the first weight, the second weight and the third weight by using an interpretable credit scoring model, and obtaining and outputting a credit score and a scoring basis of the target user.

Description

Technical Field

[0001] This application relates to the field of scoring model technology, and in particular to a method and apparatus for constructing a multi-source data scoring model based on dynamic feature weights. Background Technology

[0002] Currently, banks and financial institutions generally use credit scoring models to assess users' credit scores in order to provide corresponding services based on those scores.

[0003] Existing credit scoring models suffer from the following main problems. Firstly, feature fusion is inefficient. Traditional credit scoring models, such as logistic regression and the FICO model, primarily rely on fixed weight allocations to calculate credit scores. This makes them unable to dynamically adapt to the differences in data characteristics among different customer groups. For example, for customers without credit records, the weight of communication behavior data cannot be effectively increased. Furthermore, the correlation between non-financial data such as communication behavior and consumption habits and credit risk is non-linear. Therefore, manually preset fixed weights, such as setting operator data at 20%, can easily lead to a decrease in model prediction accuracy, for example, causing the Area Under the ROC Curve (AUC) to drop by 15% to 25%. The ROC curve is a graph used to show the model's performance at different "judgment thresholds."

[0004] On the other hand, there is a lack of model interpretability. Black-box models, such as credit scores output by deep neural networks, lack interpretability, making it difficult for financial institutions to quickly identify risk triggers. For example, deep neural network models cannot explain the impact of "increased nighttime call frequency" on the score. Furthermore, the lack of model interpretability makes it difficult for these black-box models to generate compliant explanatory reports.

[0005] On the other hand, there is insufficient coverage of long-tail customers. For customers with blank credit records, single financial data models are ineffective, and existing technologies have not achieved effective alternative assessments of non-financial data, such as failing to map "package fee stability" into an income stability indicator. Summary of the Invention

[0006] To improve the accuracy of credit scoring models, this application discloses the following technical solutions:

[0007] The first aspect of this application provides a method for constructing a multi-source data scoring model based on dynamic feature weights, including:

[0008] Obtain input data from the target user, including financial data, communication behavior data, and consumption habit data;

[0009] determine a first weight of the financial data, a second weight of the communication behavior data, and a third weight of the consumption habit data according to a number of available financial data dimensions of the financial data and a total number of dimensions of the input data;

[0010] process the input data, the first weight, the second weight, and the third weight by using the interpretable credit scoring model, and obtain and output a credit score and a scoring basis of the target user.

[0011] The second aspect of the present application provides a multi-source data scoring model construction device based on dynamic feature weights, comprising:

[0012] an input unit configured to obtain input data of a target user, the input data comprising financial data, communication behavior data, and consumption habit data;

[0013] a determination unit configured to determine a first weight of the financial data, a second weight of the communication behavior data, and a third weight of the consumption habit data according to a number of available financial data dimensions of the financial data and a total number of dimensions of the input data;

[0014] a processing unit configured to process the input data, the first weight, the second weight, and the third weight by using the interpretable credit scoring model, and obtain and output a credit score and a scoring basis of the target user.

[0015] The beneficial effects of the present application are that, based on the number of available financial data dimensions of the financial data of the target user and the total number of dimensions of the input data, the first weight of the financial data, the second weight of the communication behavior data, and the third weight of the consumption habit data are dynamically determined, so that the determined weights are more in line with the personalized features of the target user, thereby improving the accuracy of the credit score determined based on the above weights. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on the provided drawings.

[0017] Figure 1 is a flowchart of a multi-source data scoring model construction method based on dynamic feature weights provided by an embodiment of the present application;

[0018] Figure 2 is a method flowchart for determining the data weight proportion of the communication behavior data and the data weight proportion of the consumption habit data provided by an embodiment of the present application;

[0019] Figure 3 is a method flowchart provided by an embodiment of the present application for obtaining a credit score and a score basis according to an explainable credit scoring model;

[0020] Figure 4 is a schematic diagram of a visual contribution graph provided by an embodiment of the present application;

[0021] Figure 5 is a flowchart of another method for constructing a multi-source data scoring model based on dynamic feature weights provided by an embodiment of the present application;

[0022] Figure 6 is a structural schematic diagram of another device for constructing a multi-source data scoring model based on dynamic feature weights provided by an embodiment of the present application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0024] The present embodiment provides a method for constructing a multi-source data scoring model based on dynamic feature weights, as shown in Figure 1 The method can include the following steps.

[0025] 1. Obtain a plurality of different types of input data of a target user.

[0026] The target user refers to any user who needs to be evaluated for a credit score. Specifically, the target user can be a user who has not been evaluated for a credit score before, or a user who has been evaluated for a credit score before but needs to be reevaluated currently.

[0027] The input data can include communication behavior data, consumption habit data, and financial data. The communication behavior data can include any one or more of the following: night call frequency, call time period distribution, base station location trajectory, and international call frequency. The base station location trajectory refers to a set of address locations of base stations accessed by the target user.

[0028] The consumption habit data can include any one or more of the following: package consumption data, package fee fluctuation rate, traffic usage periodicity, and terminal device replacement frequency. The financial data can include any one or more of the following: credit record, account flow, historical overdue times, and historical lending behavior.

[0029] 2. Dynamically fuse the input data.

[0030] Step 2 specifically includes performing feature standardization on the input data. The method of feature standardization can be different for different types of input data, and is not specifically limited.

[0031] In some embodiments, the method of feature standardization can be to cluster unstructured data in the input data to obtain corresponding clustering features. For example, the base station location trajectory is a typical unstructured data, and in step 2, the base station location trajectory can be spatially clustered to convert into a daily activity radius dispersion value feature, which is equivalent to a kind of clustering feature.

[0032] In some embodiments, the method of feature standardization can be to calculate the coefficient of variation (CV value) within a sliding window for time series data in the input data, and the calculated coefficient of variation is taken as a time series feature.

[0033] For example, the package consumption data is a typical time series data, and the coefficient of variation (CV value) of the package consumption data within each sliding window can be calculated as the corresponding time series feature of the package consumption data.

[0034] The sliding window is a time window with a certain length, and the specific length can be set as needed, for example, one week, 10 days, etc. The length of the sliding window used when calculating the coefficient of variation for different time series data can be the same or different.

[0035] Optionally, the length of the sliding window can be determined according to the change period of the time series data, for example, it can be equal to N times of the change period, and N is a preset integer greater than 1. As an example, when calculating the coefficient of variation of the package consumption data, according to the billing rules of the operator, the package consumption data is generally counted and changed monthly, so the change period of the package consumption data is one month, and the sliding window used to calculate the coefficient of variation of the package consumption data can be N months.

[0036] For any time series data, the sliding window can be slid forward from the current time, and a coefficient of variation can be calculated according to the time series data within the current sliding window each time the sliding window is slid. The set of multiple coefficients of variation thus calculated can be taken as the corresponding time series feature of the time series data. For example, for the package consumption data, assuming that the sliding window is 3 months, a coefficient of variation can be calculated for the package consumption data in the last 3 months, another coefficient of variation can be calculated for the package consumption data from the previous 4 months to the previous 1 month, and another coefficient of variation can be calculated for the package consumption data from the previous 5 months to the previous 2 months, and so on. The set of multiple coefficients of variation thus obtained is taken as the corresponding time series feature of the package consumption data.

[0037] The method of calculating the coefficient of variation of the time series data can refer to related prior art, which is not described in detail.

[0038] It should be noted that the process of determining clustering features and temporal features described above is an optional process in the method of this embodiment. It is possible to obtain clustering features and temporal features or not.

[0039] Step 2 may also include dynamic weight allocation.

[0040] Specifically, the weighting of non-financial data can be dynamically adjusted based on the lack of financial data for the target user. Non-financial data can include communication behavior data and consumption habit data.

[0041] The specific adjustment method can be represented by the following formula (1).

[0042]

[0043] Where W_{nonfinancial} represents the weighting of non-financial data determined based on the lack of financial data; A represents a preset adjustment coefficient, the value of which can be set as needed without limitation, for example, it can be set to 0.7; k is a preset curve steepness factor, which can generally be set to k=3, but can also be set to other values ​​as needed in other embodiments; D_{financial} represents the number of available financial data dimensions; D_{total} represents the total number of dimensions of the input data; and e is the base of the natural logarithm.

[0044] According to formula (1), the smaller the number of available financial data dimensions, the more severe the lack of financial data in the input data, and the greater the corresponding weight of non-financial data. In other words, the weight of non-financial data is negatively correlated with the number of available financial data dimensions that represent the lack of financial data. As an example, suppose the target user has no relevant financial data at all, that is, no credit records, account statements, and historical lending behavior data. Then the number of available financial data dimensions D_{financial} for the target user is equal to 0. Based on formula (1), the weight of non-financial data for the target user can be calculated to be 85%.

[0045] Optionally, given that non-financial data includes two categories: communication behavior data and consumption habit data, after obtaining the data weight ratio of non-financial data, the data weight ratio of non-financial data can be allocated to communication behavior data and consumption habit data in a certain proportion, thereby determining the data weight ratio of communication behavior data and the data weight ratio of consumption habit data.

[0046] Please see Figure 2 The methods for determining the data weight ratio of communication behavior data and the data weight ratio of consumption habit data may include the following steps.

[0047] S201, determine the first data volume of communication behavior data and the second data volume of consumption habit data.

[0048] S202, based on the first data volume and the second data volume, the data weight ratio of non-financial data is divided into the data weight ratio of communication behavior data and the data weight ratio of consumption habit data, wherein the data weight ratio is positively correlated with the corresponding data volume.

[0049] The advantage of determining the data weight ratio of communication behavior data and consumption habit data using the above method is that it allocates data weight ratios according to different types of data volume, thereby highlighting the input data with a larger data volume and improving the accuracy of credit scoring.

[0050] In other words, for communication behavior data and consumption habit data, the larger the volume of the latter type of data, the greater the weight of that type of data.

[0051] The first data volume of communication behavior data can be represented by the number of communication behavior data entries of the target user collected after removing duplicate data; the second data volume of consumption habit data can be represented by the number of consumption habit data entries of the target user collected after removing duplicate data.

[0052] In some optional embodiments, the ratio of the first data volume to the second data volume (denoted as the data volume ratio) can be determined, and the data weight of non-financial data can be proportionally divided into the data weight of communication behavior data and the data weight of consumption habit data according to the data volume ratio. For example, assuming the ratio of the first data volume to the second data volume is 10:7, and the data weight of non-financial data is 85%, then 85% can be divided according to the 10:7 ratio. Thus, the data weight of communication behavior data is determined to be 50% (10 / 17 of 85%), and the data weight of consumption habit data is determined to be 35% (7 / 17 of 85%).

[0053] After obtaining the data weight ratio of non-financial data according to formula (1), the data weight ratio of financial data (denoted as W_{financial}) can be obtained based on the data weight ratio of non-financial data. Specifically, the data weight ratio of financial data can be obtained by subtracting the data weight ratio of non-financial data from 1, i.e., W_{financial}=1-W_{nonfinancial}.

[0054] Optionally, the weighting of financial data can be dynamically adjusted based on changes in the financial data. Specifically, when it is determined that a negative financial event has occurred to a target user based on the target user's financial data, the weighting of the financial data can be increased by a preset amount, such as by 20%.

[0055] It should be noted that the weighting of each data point above is determined based on the target user's personal financial data, communication behavior data, and consumption habit data. Therefore, the corresponding data weighting may vary for different users to suit their user characteristics.

[0056] 3. Process the input data according to the interpretability credit scoring model to obtain the credit score and the scoring basis.

[0057] The interpretable credit scoring model used in step 3 can be a deep neural network model that employs a hybrid architecture of optimized distributed gradient boosting library (Extreme Gradient Boosting, XGBoost) and game theory-based Shapley Additive exPlanations (SHAP). For specific implementation methods of XGBoost and SHAP, please refer to relevant existing technologies, which will not be elaborated here.

[0058] In the interpretable credit scoring model used in step 3, XGBoost mainly generates the initial credit score, while the SHAP algorithm can be used to analyze the feature contribution.

[0059] In some optional embodiments, processing the input data according to an interpretable credit scoring model to obtain a credit score and scoring criteria may further include, for example: Figure 3 The steps are shown.

[0060] S301, obtain fused data of communication behavior data and financial data.

[0061] S302, Process the fused data and input data according to the interpretable credit scoring model to obtain the credit score and the scoring basis.

[0062] For example, in step S301, the frequency of nighttime calls in the communication behavior data and the number of historical overdue payments in the financial data can be multiplied together, and the resulting product can be used as the fused data and incorporated into the model input.

[0063] In step S302, the fused data, the target user's input data, the data weight ratio of communication behavior data, the data weight ratio of consumption habit data, and the data weight ratio of financial data can be input into the interpretable credit scoring model for processing to obtain a credit score and scoring basis.

[0064] The beneficial effect of this embodiment is that by fusing communication behavior data and financial data to obtain fused data, and by determining credit scores based on fused data, the interpretable credit scoring model can comprehensively consider the inherent correlation between different types of input data when determining credit scores, which is conducive to improving the accuracy of credit scoring.

[0065] The output credit score has no limit on its range, for example, it can be between 0 and 1000 points. The higher the score, the better the credit status of the target user.

[0066] The scoring criteria can include any of the following: a visual contribution graph, positive factors, and negative factors. The visual contribution graph can include the type of input data and the contribution of each type of input data to the credit score. This contribution can be calculated based on the aforementioned SHAP algorithm. As an example, the obtained visual contribution graph can be as follows: Figure 4 As shown.

[0067] Positive factors can include those that improve credit scores and the magnitude of that improvement, while negative factors can include those that decrease credit scores and the magnitude of that decrease. For example, a positive factor could be "package stability (+120 points)"; a negative factor could be "high base station location dispersion (-50 points)". Here, +120 points represents the magnitude of the credit score improvement, and -50 points represents the magnitude of the credit score decrease.

[0068] If clustering features and time series features are not obtained from the input data when performing step 3, the data weight ratios of the target user's input data, communication behavior data, consumption habit data, and financial data can be input into the interpretable credit scoring model for processing to obtain credit scores and scoring basis.

[0069] If clustering features and time-series features have already been obtained from the input data, then the clustering features, time-series features, the aforementioned input data, and the data weight ratios can also be input into the interpretable credit scoring model for processing to obtain credit scores and scoring criteria.

[0070] The following examples illustrate the application scenarios of the method in this embodiment.

[0071] Example 1: Credit score for customers with no credit history. The input data obtained did not include financial data such as bank loan records. Only communication behavior data and consumption habit data were obtained from telecommunications operators. Communication behavior data included: 8% of calls were made at night (23:00-5:00) in the past 3 months, and the base station locations covered 2 cities. Consumption habit data included: a standard deviation of 4% in package fees, and the terminal device was a mid-range model that had been used for more than 2 years.

[0072] Given the lack of financial data, the number of available financial data dimensions D_{financial} is equal to 0, indicating that the total number of dimensions of the input data D_{total} is equal to 8. Substituting this into formula (1), the data weight ratio of non-financial data is calculated to be 85%. Figure 2After the data was divided using this method, the weighting of communication behavior data was determined to be 50%, consumption habit data to be 35%, and financial data to be 15%.

[0073] The input data and the weighting of each data point are fed into the interpretable credit scoring model. The credit score output by the interpretable credit scoring model is 650 points, which is greater than the credit threshold of 600 points. Therefore, a loan can be granted to the target user. The interpretable credit scoring model also outputs the scoring criteria: positive factor: package stability (+120 points); negative factor: cross-city activities (-70 points).

[0074] Example 2: Dynamic Risk Warning. User B's original credit score was 780. Recently, communication behavior data of this target user was obtained: the frequency of international calls increased by 300% week-on-week, and financial data showed that the credit card bill payment was delayed by 3 days. Therefore, User B's credit score was reassessed.

[0075] First, new negative financial events are identified based on financial data. Therefore, the weighting of financial data is increased, for example, from 20% to 40%, while the weighting of communication behavior data and consumption habit data is each reduced by 10%.

[0076] The interpretable credit scoring model recalculates the credit score based on the adjusted weighting of the above data, resulting in a new credit score of 620, which triggers a risk warning.

[0077] Furthermore, the interpretable credit scoring model outputs the following scoring criteria:

[0078] Warning reason: Abnormal international call (contributes 90 points) + repayment delay (contributes 70 points).

[0079] The system suggests freezing the credit limit increase application and initiating a manual review process.

[0080] The beneficial effects of the multi-source data scoring model construction method based on dynamic feature weights in this embodiment are as follows:

[0081] First, the model accuracy has been improved. On the test set of customers with no credit history, the AUC value has increased from 0.62 to 0.83; and the coverage of long-tail customers has increased to 78% (compared to 32% for the traditional model).

[0082] Second, decision-making efficiency is optimized. Dynamic weight adjustment shortens the model iteration cycle from 30 days to real-time updates; automated interpretation report generation reduces manual review time by 65%.

[0083] Third, compliance is guaranteed. It meets the requirements of relevant personal information protection regulations, and all data usage is recorded on a blockchain with customer authorization.

[0084] Furthermore, the method in this embodiment employs dynamic weight allocation technology, breaking through the limitations of traditional fixed weights. By adaptively adjusting the data missing rate, it addresses the complex correlation between non-financial data and credit risk. Additionally, it utilizes an interpretability enhancement design to provide quantitative risk factor analysis, meeting regulatory compliance requirements and reducing legal risks for financial institutions.

[0085] Based on the above embodiments, please refer to Figure 5 This application also provides a method for constructing a multi-source data scoring model based on dynamic feature weights, which may include the following steps.

[0086] S501, obtain input data from the target user, including financial data, communication behavior data, and consumption habit data.

[0087] S502, based on the number of available financial data dimensions and the total number of input data dimensions, determine the first weight of financial data, the second weight of communication behavior data, and the third weight of consumption habit data.

[0088] S503 uses an interpretable credit scoring model to process input data, first weight, second weight, and third weight, and obtains and outputs the target user's credit score and scoring basis.

[0089] The beneficial effect of this embodiment is that, based on the number of available financial data dimensions of the target user's financial data and the total number of dimensions of the input data, the first weight of financial data, the second weight of communication behavior data, and the third weight of consumption habit data are dynamically determined, so that the determined weights are more in line with the personalized characteristics of the target user, thereby improving the accuracy of the credit score determined based on the above weights.

[0090] The first weight is equivalent to the weight ratio of financial data in the aforementioned embodiments, the second weight is equivalent to the weight ratio of communication behavior data in the aforementioned embodiments, and the third weight is equivalent to the weight ratio of consumption habit data in the aforementioned embodiments.

[0091] Optionally, based on the number of available financial data dimensions and the total number of dimensions of the input data, determine the first weight of the financial data, the second weight of the communication behavior data, and the third weight of the consumption habit data, including:

[0092] The weighting of non-financial data is determined based on the number of available financial data dimensions and the total number of input data dimensions.

[0093] The primary weight of financial data is determined based on the weighting proportion of non-financial data.

[0094] Determine the first data volume of communication behavior data and the second data volume of consumption habit data;

[0095] Based on the first and second data volumes, the data weight ratios of non-financial data are divided to obtain the first weight of communication behavior data and the second weight of consumption habit data, where the weights are positively correlated with the corresponding data volumes.

[0096] The method for determining the data weight ratio of non-financial data can be found in the aforementioned formula (1). The method for determining the first weight can be found in the method for determining the data weight ratio of financial data in the aforementioned embodiment. The process for determining the second and third weights can be found in [the relevant documentation]. Figure 2 Corresponding implementation examples.

[0097] Optionally, communication behavior data includes base station location trajectories, and consumption habit data includes package consumption data;

[0098] The method also includes:

[0099] Spatial clustering of base station location trajectories yields numerical characteristics of daily activity radius dispersion.

[0100] Calculate the coefficient of variation within a sliding window for the package consumption data to obtain the time-series characteristics of the package consumption data;

[0101] The method for obtaining the numerical characteristics of the daily activity radius dispersion and the time-series characteristics corresponding to the package consumption data can be found in the content of feature standardization processing of the input data in the aforementioned embodiments.

[0102] The interpretable credit scoring model is used to process the input data, first weight, second weight, and third weight to obtain and output the target user's credit score and scoring criteria, including:

[0103] The interpretable credit scoring model is used to process the input data, the first weight, the second weight, the third weight, the numerical features of the daily activity radius dispersion, and the time series features corresponding to the package consumption data, to obtain and output the credit score and scoring basis of the target user.

[0104] Optionally, an interpretable credit scoring model is used to process the input data, the first weight, the second weight, and the third weight to obtain and output the target user's credit score and scoring basis, including:

[0105] Obtain fused data from communication behavior data and financial data;

[0106] The interpretable credit scoring model is used to process the input data, first weight, second weight, third weight, and fused data to obtain and output the target user's credit score and scoring basis.

[0107] For a detailed implementation of this embodiment, please refer to [link / reference]. Figure 3 Corresponding implementation examples.

[0108] Optionally, before processing the input data, first weight, second weight, and third weight using the interpretability credit scoring model, the following steps are also included:

[0109] When a negative financial event is identified in a target user based on financial data, the primary weight corresponding to the financial data is increased by a preset amount.

[0110] The preset range can be 20% or other ranges. For negative financial events, please refer to Example 2 above.

[0111] This application also provides an apparatus for constructing a multi-source data scoring model based on dynamic feature weights. Please refer to [link to relevant documentation]. Figure 6 The device may specifically include the following units.

[0112] The input unit 601 is used to obtain input data from the target user, including financial data, communication behavior data, and consumption habit data.

[0113] The determining unit 602 is used to determine the first weight of financial data, the second weight of communication behavior data, and the third weight of consumption habit data based on the number of available financial data dimensions and the total number of dimensions of input data.

[0114] The processing unit 603 is used to process the input data, the first weight, the second weight, and the third weight using an interpretable credit scoring model to obtain and output the credit score and scoring basis of the target user.

[0115] Optionally, the determining unit 602 determines the first weight of the financial data, the second weight of the communication behavior data, and the third weight of the consumption habit data based on the number of available financial data dimensions and the total number of dimensions of the input data, including:

[0116] The weighting of non-financial data is determined based on the number of available financial data dimensions and the total number of input data dimensions.

[0117] The primary weight of financial data is determined based on the weighting proportion of non-financial data.

[0118] Determine the first data volume of communication behavior data and the second data volume of consumption habit data;

[0119] Based on the first and second data volumes, the data weight ratios of non-financial data are divided to obtain the first weight of communication behavior data and the second weight of consumption habit data, where the weights are positively correlated with the corresponding data volumes.

[0120] Optionally, communication behavior data includes base station location trajectories, and consumption habit data includes package consumption data;

[0121] Processing unit 603 is also used for:

[0122] Spatial clustering of base station location trajectories yields numerical characteristics of daily activity radius dispersion.

[0123] Calculate the coefficient of variation within a sliding window for the package consumption data to obtain the time-series characteristics of the package consumption data;

[0124] Processing unit 603 uses an interpretable credit scoring model to process input data, first weight, second weight, and third weight, to obtain and output the target user's credit score and scoring basis, including:

[0125] The interpretable credit scoring model is used to process the input data, the first weight, the second weight, the third weight, the numerical features of the daily activity radius dispersion, and the time series features corresponding to the package consumption data, to obtain and output the credit score and scoring basis of the target user.

[0126] Optionally, the processing unit 603 uses an interpretable credit scoring model to process the input data, the first weight, the second weight, and the third weight to obtain and output the target user's credit score and scoring basis, including:

[0127] Obtain fused data from communication behavior data and financial data;

[0128] The interpretable credit scoring model is used to process the input data, first weight, second weight, third weight, and fused data to obtain and output the target user's credit score and scoring basis.

[0129] Optionally, before processing the input data, the first weight, the second weight, and the third weight using the interpretability credit scoring model, the processing unit 603 is further configured to:

[0130] When a negative financial event is identified in a target user based on financial data, the primary weight corresponding to the financial data is increased by a preset amount.

[0131] The working principle of the above-mentioned multi-source data scoring model construction device based on dynamic feature weights can be found in the relevant steps of the multi-source data scoring model construction method based on dynamic feature weights in the foregoing embodiments, and will not be repeated here.

[0132] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0133] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0134] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0135] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0136] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for constructing a multi-source data scoring model based on dynamic feature weights, characterized in that, The method comprises: obtaining input data of a target user, the input data comprising financial data, communication behavior data and consumption habit data; determining a first weight of the financial data, a second weight of the communication behavior data and a third weight of the consumption habit data according to a number of available financial data dimensions of the financial data and a total number of dimensions of the input data; processing the input data, the first weight, the second weight and the third weight by using an interpretable credit scoring model to obtain and output a credit score and a scoring basis of the target user.

2. The method of claim 1, wherein, The method further comprises: determining a first data amount of the communication behavior data and a second data amount of the consumption habit data; dividing the data weight proportion of the non-financial data according to the first data amount and the second data amount to obtain the first weight of the communication behavior data and the second weight of the consumption habit data, wherein the weight and the corresponding data amount are positively correlated. The communication behavior data comprises base station location trajectory, and the consumption habit data comprises package consumption data. The method further comprises:

3. The method of claim 1, wherein, performing spatial clustering on the base station location trajectory to obtain a daily activity radius dispersion numerical feature; calculating a coefficient of variation within a sliding window for the package consumption data to obtain a time sequence feature corresponding to the package consumption data. The method further comprises: obtaining fusion data of the communication behavior data and the financial data; processing the input data, the first weight, the second weight, the third weight, the daily activity radius dispersion numerical feature and the time sequence feature corresponding to the package consumption data by using the interpretable credit scoring model to obtain and output the credit score and the scoring basis of the target user. The method further comprises:

4. The method of claim 1, wherein, obtaining fusion data of the communication behavior data and the financial data; processing the input data, the first weight, the second weight, the third weight and the fusion data by using the interpretable credit scoring model to obtain and output the credit score and the scoring basis of the target user. The method further comprises:

5. The method of claim 1, wherein, when it is determined according to the financial data that the target user has a negative financial event, increasing the first weight corresponding to the financial data by a preset amplitude. The method comprises: 6.A dynamic feature weight based multi-source data scoring model construction device, characterized in that, ​ The input unit is configured to obtain input data of a target user, the input data comprising financial data, communication behavior data and consumption habit data; The determination unit is configured to determine a first weight of the financial data, a second weight of the communication behavior data and a third weight of the consumption habit data according to a number of available financial data dimensions of the financial data and a total number of dimensions of the input data; The processing unit is configured to process the input data, the first weight, the second weight and the third weight by using an interpretable credit scoring model, and obtain and output a credit score and a scoring basis of the target user.

7. The apparatus of claim 6, wherein, The determination unit determines the first weight of the financial data, the second weight of the communication behavior data and the third weight of the consumption habit data according to the number of available financial data dimensions of the financial data and the total number of dimensions of the input data, comprising: determining a data weight proportion of non-financial data according to the number of available financial data dimensions of the financial data and the total number of dimensions of the input data; determining the first weight of the financial data according to the data weight proportion of the non-financial data; determining a first data amount of the communication behavior data and a second data amount of the consumption habit data; dividing the data weight proportion of the non-financial data according to the first data amount and the second data amount to obtain the first weight of the communication behavior data and the second weight of the consumption habit data, wherein the weight and the corresponding data amount are positively correlated.

8. The apparatus of claim 6, wherein, The communication behavior data comprises base station location trajectory, and the consumption habit data comprises package consumption data; The processing unit is further configured to: perform spatial clustering on the base station location trajectory to obtain a daily activity radius dispersion numerical feature; calculate a coefficient of variation within a sliding window of the package consumption data to obtain a time sequence feature corresponding to the package consumption data; The processing unit processes the input data, the first weight, the second weight and the third weight by using an interpretable credit scoring model, and obtains and outputs a credit score and a scoring basis of the target user, comprising: processing the input data, the first weight, the second weight, the third weight, the daily activity radius dispersion numerical feature and the time sequence feature corresponding to the package consumption data by using an interpretable credit scoring model, and obtaining and outputting a credit score and a scoring basis of the target user.

9. The apparatus of claim 6, wherein, The processing unit processes the input data, the first weight, the second weight and the third weight by using an interpretable credit scoring model, and obtains and outputs a credit score and a scoring basis of the target user, comprising: obtaining fusion data of the communication behavior data and the financial data; processing the input data, the first weight, the second weight, the third weight and the fusion data by using an interpretable credit scoring model, and obtaining and outputting a credit score and a scoring basis of the target user.

10. The apparatus of claim 6, wherein, Before the processing unit processes the input data, the first weight, the second weight and the third weight by using an interpretable credit scoring model, the processing unit is further configured to: When it is determined according to the financial data that the target user has a negative financial event, the first weight corresponding to the financial data is increased by a preset amplitude.

Images