A risk control rejection user data processing system

Abstract

This application relates to the field of data processing, specifically to a risk control system for processing rejected user data. It includes: a rejected user pooling module, used to add a first user to a rejected user pool after their loan application is rejected, thus obtaining information on the first rejected user; a new external data matching module, used to perform matching analysis on the first rejected user information and various external data entries in the external database to obtain matching analysis results; a new external data parsing module, used to parse the new external data in response to the matching analysis result indicating the existence of new external data, thus obtaining a first user tag generation result; and a user tag result storage and solidification module, used to store and solidify the first user tag generation result. This system generates behavioral tags for rejected users by using repayment behavior information contained in the new external data added after rejection.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a risk control rejection user data processing system. Background Art

[0002] In the modern financial environment, risk control (R&D) is the core link for financial institutions to ensure the sound operation of credit business. With the development of financial technology, intelligent R&D systems have gradually become mainstream, which rely on machine learning models to predict users' credit risks. These models are usually trained based on a large amount of historical data, aiming to identify risk characteristics with generalization capabilities, thereby helping institutions make more accurate decisions. In actual operation, the R&D system divides all applicants into two categories: "approved users" and "rejected users". For "approved users", since they have obtained loans and have made actual repayments, their real repayment records can be collected as label data for subsequent model optimization and adjustment. However, since "rejected users" have not obtained loans, they lack data on actual repayment behavior, which makes it difficult for traditional R&D models to directly use the information of these users for effective training or evaluation.

[0003] Specifically, traditional risk control models mainly rely on internal data of "approved users" for training, while ignoring the potentially large number of "rejected users". This practice leads to bias in sample selection, because "approved users" often represent a group of people with lower risks, while "rejected users" include higher-risk groups. The resulting model may be biased towards low-risk characteristics and fail to fully reflect the true risk distribution of the entire applicant population. In addition, due to the lack of information on "rejected users", the model can only learn based on the behavior patterns of some users, which limits the generalization ability and prediction accuracy of the model. Especially when the market environment changes rapidly or traffic fluctuates greatly, the adaptability and stability of the model will be affected, and it will be difficult to adapt to the rapidly changing market environment and individual behavior patterns, which in turn affects the quality of the final risk control decision.

[0004] Therefore, a risk control rejection user data processing system is desired. Summary of the invention

[0005] This application is made in consideration of the above problems. One purpose of this application is to provide a risk control rejection user data processing system.

[0006] An embodiment of the present application provides a risk control rejection user data processing system, which includes:

[0007] A rejected user pool module, configured to add a first user to a rejected user pool to obtain first rejected user information after the first user's loan application is rejected;

[0008] A new external data matching module is added, which is used to perform matching analysis on the first rejected user information and each piece of external data in the external database to obtain a matching analysis result:

[0009] A newly added external data parsing module is used for, in response to the matching analysis result indicating that there is newly added external data, performing data parsing on the newly added external data to obtain a first user tag generation result;

[0010] The user tag result storage and solidification module is used to store and solidify the first user tag generation result.

[0011] For example, according to the risk control rejection user data processing system of an embodiment of the present application, the newly added external data matching module includes:

[0012] A first embedding coding unit, configured to embed the first rejected user information by using a user information embedding matrix to obtain a first rejected user information embedding coding feature vector;

[0013] An external data embedding coding unit, used to embed the pieces of external data by using the user information embedding matrix to obtain a sequence of external data embedding coding feature vectors;

[0014] A user information dynamic query matching unit is used to perform fast query dynamic matching processing on the sequence of the first rejected user information embedded coding feature vector and the external data embedded coding feature vector to obtain the user information query response semantic coding feature; wherein the user information dynamic query matching unit includes:

[0015] A fine-grained matching search window calculation subunit, used to calculate a user information fast query positioning factor of the first rejected user information embedding encoding feature vector relative to each external data embedding encoding feature vector in the sequence of the external data embedding encoding feature vector, and determine a fine-grained matching search window;

[0016] A fine-grained query response encoding subunit, configured to perform fine-grained query response encoding on the sequence of the first rejected user information embedding encoding feature vector and the external data embedding encoding feature vector based on the fine-grained matching search window to obtain the user information query response semantic encoding feature;

[0017] A newly added external data matching subunit is used to match the newly added external data based on the semantic coding features of the user information query response to determine the matching analysis result, and the matching analysis result is used to indicate whether the newly added external data exists in the external database.

[0018] For example, according to the risk control rejection user data processing system of an embodiment of the present application, the fine-grained matching search window calculation subunit includes:

[0019] A cross entropy calculation secondary subunit, used to calculate the cross entropy of the first rejected user information embedding coding feature vector relative to each external data embedding coding feature vector in the sequence of the external data embedding coding feature vector to obtain a set of user information fast query positioning factors;

[0020] A positioning matching secondary subunit, used to quickly query a set of positioning factors based on the user information, and determine a positioning matching external data embedding coding feature vector;

[0021] The search window determination secondary subunit is used to determine the fine-grained matching search window based on the positioning matching external data embedded coding feature vector.

[0022] For example, according to the risk control rejection user data processing system of an embodiment of the present application, the positioning and matching secondary subunit is used to:

[0023] The external data embedding coding feature vector corresponding to the minimum value in the set of the user information quick query positioning factors is used as the positioning matching external data embedding coding feature vector.

[0024] For example, according to the risk control rejection user data processing system of an embodiment of the present application, the vector of the center position of the fine-grained matching search window is the positioning matching external data embedded coding feature vector, and each external data embedded coding feature vector in the fine-grained matching search window is defined as a fine-grained query external data embedded coding feature vector to obtain a sequence of fine-grained query external data embedded coding feature vectors.

[0025] For example, according to the risk control rejection user data processing system of an embodiment of the present application, the fine-grained query response encoding subunit includes:

[0026] A linear transformation secondary subunit, configured to perform a linear transformation on the first rejected user information embedding coding feature vector to obtain a first rejected user information query vector and a first rejected user information value vector;

[0027] The heterogeneous conversion secondary sub-unit is used to use each fine-grained query external data embedded coding feature vector in the sequence of the fine-grained query external data embedded coding feature vector as a key vector, and input the first rejected user information query vector, the first rejected user information value vector and the key vector into a fine-grained query coding module based on a heterogeneous converter structure to obtain a user information query response semantic coding vector as the user information query response semantic coding feature.

[0028] For example, according to the risk control rejection user data processing system of an embodiment of the present application, the newly added external data matching subunit is used to:

[0029] The user information query response semantic encoding vector is passed through a newly added external data matcher based on a classifier to obtain the matching analysis result.

[0030] For example, according to the risk control rejection user data processing system of an embodiment of the present application, the newly added external data parsing module is used to:

[0031] In response to the matching analysis result indicating that there is newly added external data in the external database, selecting a corresponding data parsing script based on the matching of the newly added external data;

[0032] The user repayment behavior information in the data parsing script is parsed to generate the first user tag generation result, and the first user tag generation result is used to represent the first user repayment behavior tag.

[0033] According to the risk control rejected user data processing system of the embodiment of the present application, it can solve the problem that rejected users cannot directly obtain labels. The rejected user samples after labeling can be directly applied to model training and strategy development, thereby improving the reliability of the intelligent risk control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application are briefly introduced below. Obviously, the drawings described below only relate to some embodiments of the present application, and are not intended to limit the present application.

[0035] Figure 1 It shows a schematic diagram of the structure of the risk control rejection user data processing system in an embodiment of the present application;

[0036] Figure 2 A schematic diagram of a label generation process of a risk control rejection user data processing system in an embodiment of the present application is shown;

[0037] Figure 3 It shows a schematic diagram of the structure of a newly added external data matching module of the risk control rejection user data processing system in an embodiment of the present application;

[0038] Figure 4 A schematic diagram showing the structure of a user information dynamic query and matching unit of a risk control rejection user data processing system in an embodiment of the present application is shown; and

[0039] Figure 5 A flow chart of a method for processing risk control rejected user data in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work also fall within the scope of protection of the present application.

[0041] The terms used in this specification are those common terms currently widely used in the art in consideration of the functions of the present application, but these terms may vary according to the intentions of those skilled in the art, precedents, or new technologies in the art. In addition, specific terms may be selected, and in this case, their detailed meanings will be described in the detailed description of the present application. Therefore, the terms used in the specification should not be understood as simple names, but rather as a general description based on the meaning of the terms and the present application.

[0042] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0043] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. At the same time, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0044] In addition, the application architecture diagram in the embodiment of the present application is intended to more clearly illustrate the technical solution in the embodiment of the present application, and does not constitute a limitation on the technical solution provided in the embodiment of the present application. Of course, for other application architectures and business applications, the technical solution provided in the embodiment of the present application is also applicable to similar problems.

[0045] The following is a non-restrictive description of the risk control rejection user data processing system provided according to at least one embodiment of the present application through several examples or embodiments. As described below, different features in these specific examples or embodiments can be combined with each other without conflicting with each other to obtain new examples or embodiments, and these new examples or embodiments also fall within the scope of protection of the present application.

[0046] For "approved users", since they have obtained loans and made actual repayments, their real repayment records can be collected as label data for subsequent model optimization and adjustment. However, since "rejected users" have not obtained loans, they lack data on their actual repayment behavior, which makes it difficult for traditional risk control models to directly use the information of these users for effective training or evaluation.

[0047] In response to the above technical problems, in the technical solution of the present application, a risk control rejected user data processing system is proposed, which can generate user behavior labels for "rejected users". Accordingly, in actual application scenarios, in addition to "internal data", commonly used data also includes "external data", that is, data provided by third-party data suppliers. Due to compliance restrictions, "external data" only has query permissions in the application stage. However, with the increasing frequency of interaction between users and institutions, subsequent applications may occur after an application is rejected, which will generate new "external data". Therefore, the technical concept of the present application is to generate behavior labels for "rejected users" by using the repayment behavior information contained in the newly added "external data" after the rejection.

[0048] Specifically, Figure 1 The schematic diagram of the structure of the risk control rejected user data processing system 800 in the embodiment of the present application is shown. In the technical solution of the present application, the risk control rejected user data processing system 800 includes: a rejected user pool module 810, which is used to add the first user to the rejected user pool to obtain the first rejected user information after the loan application of the first user is rejected; a newly added external data matching module 820, which is used to match and analyze the first rejected user information with each piece of external data in the external database to obtain a matching analysis result; a newly added external data parsing module 830, which is used to respond to the matching analysis result that there is newly added external data, and perform data parsing on the newly added external data to obtain the first user tag generation result; a user tag result storage and solidification module 840, which is used to store and solidify the first user tag generation result.

[0049] In the risk control rejected user data processing system, the risk control model relies on a large amount of labeled data for training. For "rejected users", there is no actual lending and repayment behavior, so there is a lack of direct internal data to generate labels. Figure 2 The risk control rejection user data processing system described in this application is described in detail.

[0050] In the rejected user pool module 810 and the newly added external data matching module 820, data matching is implemented. After the user application is rejected and joins the rejected user pool, the system will match the user information in the rejected user pool with the full amount of data in the external database at a fixed time interval (for example, every hour or every day) to find the newly added external data since the user was rejected. This process is automated to ensure timeliness and comprehensiveness. The automated implementation involves automatically adding the rejected loan application users to the rejected user pool and regularly matching with the external database to find the newly added external data.

[0051] In a specific example, a user management module is provided, which is responsible for maintaining a dynamically updated pool of rejected users. When a user's loan application is rejected, the user's basic information and rejection timestamp will be automatically recorded and stored in this pool. This process can be implemented by writing an automated script, which is triggered at the rejection node of the loan approval process and inserts the user data into the rejected user pool database. Next, a timed task scheduler is provided to set a fixed time interval (for example, every hour or every day) to perform the data matching task. This scheduler can be a time-based trigger that wakes up the data matcher at a preset time point and starts the matching process. The data matcher is the core of the automation implementation, which is responsible for matching the user information in the rejected user pool with the full amount of data in the external database. This module can be implemented by writing a matching algorithm that uses efficient query statements and indexes to speed up the matching process. The algorithm checks whether each rejected user has new external data records generated after the rejection time point. If so, these data will be marked as newly added external data and enter the next step of data parsing. In order to improve the accuracy and efficiency of matching, the system can use big data processing technologies such as Hadoop or Spark to process large-scale data sets. These technologies can process data in a distributed manner, improving the processing speed and reliability of matching tasks. In addition, the system also requires an exception handling mechanism to handle any errors or exceptions that may occur during the matching process. This mechanism can include error logging, retry mechanisms, and alarm systems to ensure that system administrators can understand and solve problems in a timely manner. Finally, all matching results will be stored in an intermediate database or data warehouse for subsequent data parsing and label generation. This storage process also needs to be automated to ensure data consistency and traceability.

[0052] Through the above implementation, it is possible to automatically ensure the dynamic update of the rejected user pool and the efficient matching with the external database, providing accurate input data for subsequent data processing. This automated process not only improves the efficiency of data processing, but also reduces the possibility of human errors and enhances the stability and reliability of the entire risk control system.

[0053] Once the newly added "external data" after the rejection time point is matched, the data parsing link is entered. Then, in the newly added external data parsing module 830, based on the newly added "external data" supplier, the corresponding data parsing script is matched to parse the repayment behavior information therein. In one example, a data parsing trigger mechanism can be provided, which is automatically activated after the newly added external data is detected. This mechanism can be an event listener, which is used to monitor the output and trigger the data parsing process when the newly added data is found. Next, according to the supplier information of the newly added external data, the corresponding data parsing script is automatically selected. In order to achieve this, the system needs to maintain a mapping table between suppliers and parsing scripts, which lists in detail the data format and corresponding parsing script of each supplier. This mapping table can be stored in a database and maintained and updated through a configuration management module. Once the appropriate parsing script is determined, the system will execute the script to parse the external data. The parsing script is specially designed for the data format of different suppliers and can identify and extract key information, such as repayment date, repayment amount, overdue status, etc. These scripts may need to be implemented using regular expressions, XML / JSON parsers or other data extraction tools. During the parsing process, the system needs to deal with various data quality issues, such as missing values, inconsistent formats, or data errors. To this end, the system can integrate a data cleaning module, which is responsible for cleaning and standardizing the data after parsing to ensure the quality and consistency of the data.

[0054] Then, a label is generated based on the repayment behavior information obtained. The generation logic is as follows:

[0055] 1) If there is no repayment after the rejection time, the label is empty;

[0056] 2) If there is repayment after the rejection time and the repayment is normal, the label is good;

[0057] 3) If there is a repayment after the rejection time and the repayment is overdue, the label is bad.

[0058] If multiple newly added "external data" are matched, the data parsing and label generation steps are repeated to generate a label for each newly added "external data".

[0059] Specifically, the newly added external data analysis module 830 is used to: in response to the matching analysis result indicating that there is newly added external data in the external database, select a corresponding data analysis script based on the matching of the newly added external data; parse the user repayment behavior information in the data analysis script to generate the first user label generation result, and the first user label generation result is used to represent the first user repayment behavior label.

[0060] Finally, in the user label result storage solidification module 840, the label generation result is stored and solidified, and all "external data" label generation results are stored with the user as the granularity (for example, if user A matches external data 1 and external data 2, the external data 1 label and the external data 2 label are finally stored).

[0061] It is worth mentioning that the risk control rejected user data processing system can solve the problem that "rejected users" cannot directly obtain labels. The labeled "rejected user" samples can be directly applied to model training and strategy development, thereby improving the reliability of the intelligent risk control system.

[0062] In particular, in the above-mentioned risk control rejected user data processing system, the process of matching and analyzing the first rejected user information and each piece of external data in the external database to obtain the matching analysis result is particularly important, which can provide a basis for subsequent parsing and user label generation by identifying the newly added external data. Specifically, in the process of matching analysis, the technical concept of the present application is to analyze the rejected user information and each piece of external data in the external database by using artificial intelligence-based data processing and semantic understanding algorithms, so as to capture the dynamic semantic query response feature representation between the embedded semantics of the rejected user information and the embedded semantics of each piece of external data, and match the newly added external data based on the dynamic semantic query response feature representation to determine whether there is newly added external data in the external database. In this way, the problem of labeling rejected users can be solved, and a basis can be provided for the comprehensiveness and reliability of user labels, thereby improving the intelligence level and reliability of the risk control system.

[0063] Accordingly, if Figure 3 As shown, in the technical solution of the present application, the newly added external data matching module 820 includes: a first embedded coding unit 821, which is used to embed the first rejected user information by using a user information embedding matrix to obtain a first rejected user information embedded coding feature vector; an external data embedded coding unit 822, which is used to embed the various external data by using the user information embedding matrix to obtain a sequence of external data embedded coding feature vectors; a user information dynamic query matching unit 823, which is used to perform fast query dynamic matching processing on the sequence of the first rejected user information embedded coding feature vector and the external data embedded coding feature vector to obtain a user information query response semantic coding feature.

[0064] More specifically, in the technical solution of the present application, first, the first rejected user information is embedded and encoded by using a user information embedding matrix to obtain a first rejected user information embedded coding feature vector. It should be understood that user information usually contains a variety of different types of data, such as personal information, historical transaction records, credit scores, etc. These different types of user information are often difficult to process and semantically analyze subsequently, and are also not conducive to query matching with external data to determine whether there is new external data. Therefore, in the technical solution of the present application, these different types of user information can be embedded and mapped into a common space through embedded coding, and the embedded coding semantic features of the first rejected user information can be captured for subsequent semantic query and matching analysis.

[0065] Next, consider that similar problems also exist in each piece of external data in the external database, that is, different external data also contain different types of data information, which may come from different sources and have different storage formats, which will affect the subsequent semantic query matching analysis of user information. Therefore, in the technical solution of the present application, each piece of external data is embedded and encoded respectively by using the user information embedding matrix to obtain a sequence of external data embedded coding feature vectors. By embedding and encoding each piece of external data respectively through the user information embedding matrix, these external data can be embedded and mapped into the same space, and the embedded coding semantics of each piece of external data can be captured at the same time, so as to compare and analyze with the first rejected user information embedded coding semantics in the same feature space, thereby reducing the risk of false queries and false matches.

[0066] It should be understood that since the sequence of the first rejected user information embedded coding feature vector and the external data embedded coding feature vector respectively contains the first rejected user information embedded semantic coding features and each external data embedded semantic coding features, in the process of matching the newly added external data, it is necessary to perform a matching query on the first rejected user information embedded semantic coding features and each external data embedded semantic coding features to retrieve the relevant semantic information to determine whether there is newly added external data in the external database. However, since the traditional feature matching query method is based on the overall database information for screening, the response speed and accuracy of the retrieval query are low. Therefore, in the technical solution of the present application, the sequence of the first rejected user information embedded coding feature vector and the external data embedded coding feature vector is further subjected to a fast query dynamic matching process to obtain the user information query response semantic coding features. In particular, the process of fast query dynamic matching processing realizes the process from coarse positioning to fine-grained matching and then to deep semantic coding by integrating technologies such as cross entropy calculation, local context perception, linear transformation, and heterogeneous converter structure, aiming to optimize the speed and accuracy of user information semantic query response.

[0067] Specifically, the process of fast query dynamic matching processing introduces a fast positioning mechanism, that is, using cross entropy calculation to evaluate the semantic distance between the first rejected user information embedded coding feature vector and each of the external data embedded coding feature vectors, so that the semantic difference between the two can be quantified, thereby quickly narrowing the search range to the most likely related area and improving the response speed of query matching. Then, a preliminary screening finds a best matching external data embedded coding feature vector. This method reduces unnecessary computing resource consumption and accelerates the preliminary screening process. Secondly, in order to further improve the matching query accuracy, the system defines a fine-grained matching search window around the best matching external data embedded coding feature vector, which focuses on the local area around the best matching point determined by the preliminary screening. This strategy enhances the model's ability to understand the deep association semantics between the first rejected user information and each piece of external data, enabling the system to capture more subtle matching query response semantic differences, thereby providing more accurate user information query response retrieval results, and providing a basis for subsequent new external data queries and judgments.

[0068] The dynamic semantic query encoding module based on the fast positioning mechanism is a new feature query-response encoding architecture. It realizes the process from coarse positioning to fine-grained matching and then to deep semantic encoding by integrating technologies such as cross-entropy calculation, local context awareness, linear transformation and heterogeneous transformer structure, aiming to optimize the speed and accuracy of query response.

[0069] Accordingly, if Figure 4As shown, the user information dynamic query matching unit 823 includes: a fine-grained matching search window calculation subunit 8231, which is used to calculate the user information fast query positioning factor of the first rejected user information embedded coding feature vector relative to each external data embedded coding feature vector in the sequence of the external data embedded coding feature vector, and determine the fine-grained matching search window; a fine-grained query response encoding subunit 8232, which is used to perform fine-grained query response encoding on the first rejected user information embedded coding feature vector and the sequence of the external data embedded coding feature vector based on the fine-grained matching search window to obtain the user information query response semantic coding feature; a newly added external data matching subunit 8233, which is used to perform newly added external data matching based on the user information query response semantic coding feature to determine the matching analysis result, and the matching analysis result is used to indicate whether there is newly added external data in the external database.

[0070] Among them, the fine-grained matching search window calculation subunit 8231 includes: a cross-entropy calculation secondary subunit, used to calculate the cross entropy of the first rejected user information embedded coding feature vector relative to each external data embedded coding feature vector in the sequence of the external data embedded coding feature vector to obtain a set of user information fast query positioning factors; a positioning matching secondary subunit, used to determine the positioning matching external data embedded coding feature vector based on the set of user information fast query positioning factors; a search window determination secondary subunit, used to determine the fine-grained matching search window based on the positioning matching external data embedded coding feature vector.

[0071] Specifically, in the process of fast query dynamic matching, the cross entropy of the first rejected user information embedded coding feature vector relative to each external data embedded coding feature vector in the sequence of the external data embedded coding feature vector is first calculated to obtain a set of user information fast query positioning factors. That is, by calculating the cross entropy of each external data embedded coding feature vector in the sequence of the first rejected user information embedded coding feature vector and the external data embedded coding feature vector, a numerical set reflecting the semantic distance or mismatch between the two is obtained, that is, a set of user information fast query positioning factors. Cross entropy, as an important indicator to measure the difference between two probability distributions, is used here to quantify the semantic differences between the first rejected user information embedded coding feature vector and each external data embedded coding feature vector. A smaller cross entropy value means a more similar semantic expression, which can quickly narrow the search range to the most likely related area.

[0072] The process can be expressed as follows:

[0073]

[0074]

[0075]

[0076] in, is the sequence of external data embedding encoding feature vectors, are respectively the first, second, and third in the sequence of the external data embedded coding feature vector and External data embedding encoding feature vector, The first rejected user information is embedded in the encoded feature vector The eigenvalues ​​at the positions, is the number of eigenvalues, It is the The external data is embedded in the encoded feature vector The eigenvalues ​​at the positions, is the first rejected user information embedded encoding feature vector and the first The cross entropy between the external data embedding encoding feature vectors, i.e., the user information fast query positioning factor, It is a collection of user information quick query positioning factors. The first, second, and and A user information quick query positioning factor.

[0077] In particular, in the technical solution of the present application, the cross entropy is not just a simple distance measurement tool, but a key indicator to measure the difference between two probability distributions. When applied to the matching between the first rejected user information embedding encoding feature vector and the external data embedding encoding feature vector, by treating each external data embedding encoding feature vector as a possible probability distribution and using the first rejected user information embedding encoding feature vector to evaluate these distributions, a series of scores reflecting the semantic similarity between the two can be obtained-that is, the user information fast query positioning factor is a set. The cross entropy calculation here not only takes into account the direct semantic similarity, but also implicitly captures the potential contextual relationship and semantic structure. This mechanism helps to improve retrieval efficiency, especially when facing large-scale data sets, and can significantly reduce unnecessary computational burden while maintaining high retrieval accuracy.

[0078] Next, the minimum value in the set of fast query positioning factors calculated above is used to determine the best matching external data embedding coding feature vector, that is, to locate the matching external data embedding coding feature vector. The basis for this selection is that the external data embedding coding feature vector corresponding to the minimum value is considered to be the candidate that is semantically closest to the first rejected user information embedding coding feature vector. This strategy ensures that the external data embedding coding feature vectors initially screened have the highest relevance, providing a good starting point for subsequent more detailed searches. From the perspective of machine learning, this method embodies the principle of the greedy algorithm: each iteration makes the best choice that seems to be the best at the moment, hoping to achieve the overall optimality through a series of such choices.

[0079] Correspondingly, the positioning matching secondary subunit is used to: use the external data embedding coding feature vector corresponding to the minimum value in the set of user information fast query positioning factors as the positioning matching external data embedding coding feature vector. This process can be expressed as:

[0080]

[0081] in, To return the minimum value , Embed the encoded feature vector for localization matching external data.

[0082] Next, based on the location matching external data embedding encoding feature vector, a fine-grained matching search window is delineated around the point, and the external data embedding encoding feature vector within this window is designated as the fine-grained query external data embedding encoding feature vector for the next more refined matching operation. The design of the fine-grained matching search window allows the system to concentrate resources on in-depth analysis of a small range of data that may contain precise answers, which not only improves retrieval accuracy but also ensures computational efficiency. The process of defining the fine-grained matching search window introduces an important concept - local context awareness. By constructing a small search area around the initially selected location matching external data embedding encoding feature vector, the system is able to conduct in-depth analysis in a smaller and more targeted data subset. This not only improves search efficiency, but also enhances the model's ability to understand local semantic patterns. In addition, the design of the fine-grained matching search window allows the model to better handle the subtle differences between the first rejected user information embedding encoding feature vector and the external data embedding encoding feature vector. By focusing on fine-grained features, the model can capture such details more accurately, thereby improving the quality of the final output. Correspondingly, the vector at the center position of the fine-grained matching search window is the positioning matching external data embedded coding feature vector, and each external data embedded coding feature vector in the fine-grained matching search window is defined as a fine-grained query external data embedded coding feature vector to obtain a sequence of fine-grained query external data embedded coding feature vectors.

[0083] Accordingly, based on the location matching external data embedded encoding feature vector, a fine-grained matching search window is defined around the point and can be expressed as:

[0084]

[0085]

[0086] in, It means calculating the square of the vector-norm. It means to calculate the logarithmic function value with base 2. Indicates rounding down. is the matching search window factor, ( ) is the starting point of the fine-grained matching search window, ( ) is the end point of the fine-grained matching search window, Embedding sequences of encoded feature vectors for fine-grained queries on external data, The fine-grained query external data embedding encoding feature vectors are respectively embedded for each fine-grained query external data in the sequence of the fine-grained query external data embedding encoding feature vectors.

[0087] Finally, a linear transformation is performed on the first rejected user information embedding encoding feature vector to generate a query vector and a value vector. Subsequently, these vectors are input into a fine-grained query encoding module based on a heterogeneous transformer structure, combined with the key vectors in the fine-grained matching search window. In this module, the interaction between query, key, and value simulates the attention mechanism, captures the complex relationship between features through a multi-head self-attention layer, and finally outputs the query response semantic encoding vector. The heterogeneous transformer structure allows the model to handle different types of input data and introduces additional flexibility and expressiveness in the encoding process, resulting in a richer and more accurate query response representation. Linear transformation is one of the key steps to transform the original first rejected user information embedding encoding feature vector into a more suitable representation for downstream tasks. In this process, the first rejected user information embedding encoding feature vector is decomposed into the first rejected user information query vector and the first rejected user information value vector, which are used to guide different aspects of the attention mechanism: the first rejected user information query vector is used to determine which parts should be focused on; while the first rejected user information value vector carries the actual information content. This design enables the attention mechanism to operate more flexibly and targeted. The application of heterogeneous transformers is to address the challenges of interaction between different types of data. Traditional transformer structures usually assume that the input data has the same modality or format, but in the real world, the characteristics of the two often come from different sources or forms of expression. Heterogeneous transformers ensure that the model can maintain good performance in complex environments by introducing additional flexibility, such as support for variable-length inputs and multimodal data processing. The multi-head self-attention layer within the fine-grained query encoding module allows for parallel processing of feature associations from multiple different angles, thereby generating a richer and more comprehensive semantic encoding vector for user information query responses.

[0088] Correspondingly, the fine-grained query response encoding subunit 8232 includes: a linear transformation secondary subunit, which is used to perform a linear transformation on the first rejected user information embedded encoding feature vector to obtain a first rejected user information query vector and a first rejected user information value vector; a heterogeneous conversion secondary subunit, which is used to use each fine-grained query external data embedded encoding feature vector in the sequence of the fine-grained query external data embedded encoding feature vector as a key vector, and input the first rejected user information query vector, the first rejected user information value vector and the key vector into a fine-grained query encoding module based on a heterogeneous converter structure to obtain a user information query response semantic encoding vector as the user information query response semantic encoding feature. The process can be expressed as follows:

[0089]

[0090]

[0091]

[0092] in, is the first rejected user information embedded encoding feature vector, and are the query weight matrix and query bias vector respectively, and are the value weight matrix and value bias vector respectively, and are the first rejected user information query vector and the first rejected user information value vector respectively, The first one in the sequence of the fine-grained query external data embedding encoding feature vector A fine-grained query of external data embedding encoding feature vectors, For the said A fine-grained query of the length of the external data embedding encoding feature vector, is matrix multiplication, for function, A semantic encoding vector is generated for the user information query response.

[0093] Then, the user information query response semantic encoding vector is passed through a newly added external data matcher based on a classifier to obtain the matching analysis result, and the matching analysis result is used to indicate whether there is newly added external data in the external database. In other words, classification processing is performed using the query response feature representation between the user information semantics and the semantics of each external data, so as to match the newly added external data to determine whether there is newly added external data in the external database. In this way, the problem of labeling rejected users can be solved, and a basis can be provided for the comprehensiveness and reliability of user labels, thereby improving the intelligence level and reliability of the risk control system. Accordingly, the newly added external data matching subunit 8233 is used to: pass the user information query response semantic encoding vector through a newly added external data matcher based on a classifier to obtain the matching analysis result.

[0094] In the first embodiment, the newly added external data matching subunit adopts a trained machine learning classifier. The classifier may be a random forest, support vector machine or neural network model, which is pre-trained with a large amount of historical user information and corresponding labels (such as "new data exists" or "no new data"). When the user information query response semantic encoding vector is generated, it will be input into this classifier. The classifier will calculate the probability that the vector belongs to each category based on the internal algorithm, and select the category with the highest probability as the final matching analysis result. For example, if the vector is closer to samples that are known to have new data, the classifier will return "new external data exists".

[0095] In the second embodiment, a relatively simple strategy is adopted, which is based on threshold judgment. The semantic encoding features of the user information query response generated by the fine-grained query response encoding subunit will be compared with a set of predefined thresholds. These thresholds may be the best segmentation points obtained by analyzing a large amount of historical data, which are used to distinguish between the situation of "there is new data" and the situation of "no new data". If the encoding feature exceeds a certain threshold, it is considered that there is new external data in the external database; otherwise, there is no new external data.

[0096] In the third embodiment, deep learning techniques are used, especially convolutional neural networks (CNN) or recurrent neural networks (RNN), which are particularly effective in processing complex semantic encoding features. Deep learning classifiers can automatically capture complex patterns in input data through multiple layers of nonlinear transformations, thereby more accurately distinguishing between "there is new data" and "no new data". This classifier requires a large amount of labeled data for training, but once trained, it can perform well on new data.

[0097] In the fourth embodiment, the concept of adaptive classification is introduced. As the system receives more new data, the classifier can dynamically adjust its parameters to adapt to the changing data distribution. This can be achieved through online learning or regular model updates. The benefit of an adaptive classifier is that it can continuously improve its performance over time, especially in the face of a rapidly changing data environment.

[0098] In the fifth embodiment, a set of carefully designed rule sets is relied upon, which are summarized based on the knowledge of domain experts and historical data analysis. When the semantic coding feature of the user information query response is generated, it will be compared with each rule in the rule set in turn. If the conditions of a certain rule are met, the corresponding matching analysis results are directly output. The advantage of this method is that it is easy to understand and explain, but the disadvantage is that it is less flexible and difficult to cope with complex and changing data scenarios.

[0099] Through the above five embodiments, it can be seen that the newly added external data matching subunit 8233 can use a variety of different methods to perform classification tasks. Each method has its own unique advantages and applicable scenarios. In practical applications, the most appropriate method can be selected according to specific needs and technical conditions, or several methods can be combined to achieve the best effect. Such diversity not only improves the flexibility and adaptability of the system, but also provides broad space for future development.

[0100] Furthermore, in a preferred example, considering that the sequence of the first rejected user information embedded coding feature vector and the external data embedded coding feature vector respectively represent the embedded coding features of the first rejected user information and each external data, when performing dynamic semantic query encoding based on a fast positioning mechanism, due to the differences in the fast positioning mechanism caused by the differences in the distribution of the embedded coding features of the source data, the user information query response semantic coding vector will also have probabilistic convergence and divergence based on different dynamic semantic queries, thereby affecting the accuracy of the matching analysis results obtained by the newly added external data matcher based on the classifier.

[0101] Therefore, when the user information query response semantic encoding vector is passed through a newly added external data matcher based on a classifier, the applicant of the present application first optimizes the user information query response semantic encoding vector, including the following steps:

[0102] Determine the user information query response semantic encoding vector by the user information query response semantic encoding distribution probability value of the newly added external data matcher based on the classifier , wherein the user information query response semantic encoding distribution probability value Indicates the probability that new external data exists in the external database;

[0103] Each feature value in the user information query response semantic encoding vector is normalized to the maximum value to obtain the user information query response semantic encoding probability feature value ;

[0104] Subtract the user information query response semantic encoding distribution probability value from 1 Calculate the semantic encoding probability feature value of each user information query response for the index The power function of the first user information query response semantic encoding probability convergence eigenvalue is obtained ;in, is the semantic encoding distribution probability value of the user information query response, is the semantic encoding probability feature value of the user information query response, is the probability convergence characteristic value of the semantic encoding of the first user information query response;

[0105] The user information query response semantic encoding distribution probability value Calculate an index for each user information query response semantic encoding probability feature value The power function of the difference is used to obtain the probability convergence eigenvalue of the semantic encoding of the second user information query response ;in, is the semantic encoding distribution probability value of the user information query response, is the semantic encoding probability feature value of the user information query response, is the probability convergence characteristic value of the semantic encoding of the second user information query response;

[0106] The semantic encoding probability feature value of each user information query response minus the user information query response semantic encoding distribution probability value The difference is multiplied to obtain the first user information query response semantic encoding probability limiting feature value ;in, is the semantic encoding distribution probability value of the user information query response, is the semantic encoding probability feature value of the user information query response, is the first user information query response semantic coding probability limiting feature value;

[0107] A semantic encoding probability feature value associated with each of the user information query responses The difference between the user information query response semantic encoding distribution probability value Multiply to obtain the second user information query response semantic encoding probability limit feature value ;in, is the semantic encoding distribution probability value of the user information query response, is the semantic encoding probability feature value of the user information query response, is the second user information query response semantic encoding probability limiting feature value;

[0108] After multiplying the first user information query response semantic coding probability convergence eigenvalue by the second user information query response semantic coding probability convergence eigenvalue, further add the first user information query response semantic coding probability limit eigenvalue and the second user information query response semantic coding probability limit eigenvalue to obtain the optimized eigenvalue of the user information query response semantic coding vector ,in, is the probability convergence eigenvalue of the semantic encoding of the first user information query response, is the probability convergence eigenvalue of the semantic encoding of the second user information query response, is the semantically encoded probability-limited feature value of the first user information query response, is the second user information query response semantic encoding probability limiting feature value, It is the optimized feature value of the semantic encoding vector of the user information query response.

[0109] Based on this, by using the cross entropy form power series of the regression probability of the user information query response semantic coding vector as the probability distribution convergence limit, the regression probability convergence approximation is performed on the basis of the combination of the feature set distribution and the probability density distribution of the user information query response semantic coding vector, so as to avoid the conservative convergence control of the fatal potential divergence of the user information query response semantic coding vector, and further use its probability distribution cross entropy as the objective function to guide the limited recovery strategy, so as to achieve the common agility of the convergence of the probability convergence and divergence of the user information query response semantic coding vector to the probability density distribution space, and improve the accuracy of the matching analysis results obtained by the user information query response semantic coding vector through the newly added external data matcher based on the classifier. In this way, the problem of labeling rejected users can be solved, and a basis can be provided for the comprehensiveness and reliability of user labels, thereby improving the intelligence level and reliability of the risk control system.

[0110] Based on the above embodiments, see Figure 5 As shown, it is a flow chart of a method for processing risk control rejection user data in an embodiment of the present application. Figure 5 As shown, according to the risk control rejected user data processing method of the embodiment of the present application, the steps include: S510, after the loan application of the first user is rejected, the first user is added to the rejected user pool to obtain the first rejected user information; S520, the first rejected user information and each external data in the external database are matched and analyzed to obtain a matching analysis result: S530, in response to the matching analysis result that there is newly added external data, the newly added external data is parsed to obtain a first user label generation result; S540, the first user label generation result is stored and solidified.

[0111] Here, those skilled in the art can understand that the specific operations of each step in the above risk control rejection user data processing method have been referred to above. Figures 1 to 4 The risk control rejection user data processing system 800 has been described in detail, and therefore, its repeated description will be omitted.

[0112] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present application is not limited to any particular form of combination of hardware and software.

[0113] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those of ordinary skill in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an idealized or extremely formal sense, unless explicitly defined as such herein.

[0114] The above is an explanation of the present application and should not be considered as limiting thereof. Although several exemplary embodiments of the present application have been described, those skilled in the art will readily appreciate that many modifications may be made to the exemplary embodiments without departing from the novel teachings and advantages of the present application.

Claims

1. A risk control rejection user data processing system, characterized in that: include: A rejected user pool module, configured to add a first user to a rejected user pool to obtain first rejected user information after the first user's loan application is rejected; A new external data matching module is added, which is used to perform matching analysis on the first rejected user information and each piece of external data in the external database to obtain a matching analysis result: A newly added external data parsing module is used for, in response to the matching analysis result indicating that there is newly added external data, performing data parsing on the newly added external data to obtain a first user tag generation result; A user tag result storage and solidification module, used for storing and solidifying the first user tag generation result; Wherein, the newly added external data matching module includes: A first embedding coding unit, configured to embed the first rejected user information by using a user information embedding matrix to obtain a first rejected user information embedding coding feature vector; An external data embedding coding unit, used to embed the pieces of external data by using the user information embedding matrix to obtain a sequence of external data embedding coding feature vectors; A user information dynamic query matching unit is used to perform fast query dynamic matching processing on the sequence of the first rejected user information embedded coding feature vector and the external data embedded coding feature vector to obtain the user information query response semantic coding feature; wherein the user information dynamic query matching unit includes: A fine-grained matching search window calculation subunit, used to calculate a user information fast query positioning factor of the first rejected user information embedding encoding feature vector relative to each external data embedding encoding feature vector in the sequence of the external data embedding encoding feature vector, and determine a fine-grained matching search window; A fine-grained query response encoding subunit, configured to perform fine-grained query response encoding on the sequence of the first rejected user information embedding encoding feature vector and the external data embedding encoding feature vector based on the fine-grained matching search window to obtain the user information query response semantic encoding feature; A newly added external data matching subunit is used to match the newly added external data based on the semantic coding features of the user information query response to determine the matching analysis result, and the matching analysis result is used to indicate whether the newly added external data exists in the external database; Wherein, the fine-grained matching search window calculation subunit includes: A cross entropy calculation secondary subunit, used to calculate the cross entropy of the first rejected user information embedding coding feature vector relative to each external data embedding coding feature vector in the sequence of the external data embedding coding feature vector to obtain a set of user information fast query positioning factors; A positioning matching secondary subunit, used to quickly query a set of positioning factors based on the user information, and determine a positioning matching external data embedding coding feature vector; A search window determination secondary subunit is used to determine the fine-grained matching search window based on the positioning matching external data embedding encoding feature vector; The calculation process of the set of user information quick query positioning factors can be expressed as follows: in, is the sequence of external data embedding encoding feature vectors, are respectively the first, second, and third in the sequence of the external data embedded coding feature vector and External data embedding encoding feature vector, The first rejected user information is embedded in the encoded feature vector The eigenvalues ​​at the positions, is the number of eigenvalues, It is the The external data is embedded in the encoded feature vector The eigenvalues ​​at the positions, is the first rejected user information embedded encoding feature vector and the first The cross entropy between the external data embedding encoding feature vectors, i.e., the user information fast query positioning factor, It is a collection of user information quick query positioning factors. The first, second, and and User information quick query positioning factor; Among them, the calculation process of locating and matching the external data embedded coding feature vector can be expressed as follows: in, To return the minimum value , Embed the encoded feature vector for positioning matching external data; The calculation process of the fine-grained matching search window can be expressed as follows: in, It means calculating the square of the vector-norm. It means to calculate the logarithmic function value with base 2. Indicates rounding down. is the matching search window factor, ( ) is the starting point of the fine-grained matching search window, ( ) is the end point of the fine-grained matching search window, embedding sequences of encoded feature vectors for fine-grained queries on external data, The fine-grained query external data embedding encoding feature vectors are respectively embedded for each fine-grained query external data in the sequence of the fine-grained query external data embedding encoding feature vectors.

2. The risk control rejection user data processing system according to claim 1, characterized in that: The positioning and matching secondary subunit is used for: The external data embedding coding feature vector corresponding to the minimum value in the set of the user information quick query positioning factors is used as the positioning matching external data embedding coding feature vector.

3. The risk control rejection user data processing system according to claim 2, characterized in that: The vector at the center position of the fine-grained matching search window is the positioning matching external data embedded coding feature vector, and each external data embedded coding feature vector in the fine-grained matching search window is defined as a fine-grained query external data embedded coding feature vector to obtain a sequence of fine-grained query external data embedded coding feature vectors.

4. The risk control rejection user data processing system according to claim 3 is characterized in that: The fine-grained query response encoding subunit includes: A linear transformation secondary subunit, configured to perform a linear transformation on the first rejected user information embedding coding feature vector to obtain a first rejected user information query vector and a first rejected user information value vector; The heterogeneous conversion secondary sub-unit is used to use each fine-grained query external data embedded coding feature vector in the sequence of the fine-grained query external data embedded coding feature vector as a key vector, and input the first rejected user information query vector, the first rejected user information value vector and the key vector into a fine-grained query coding module based on a heterogeneous converter structure to obtain a user information query response semantic coding vector as the user information query response semantic coding feature.

5. The risk control rejection user data processing system according to claim 4, characterized in that: The newly added external data matching subunit is used to: pass the user information query response semantic encoding vector through a newly added external data matcher based on a classifier to obtain the matching analysis result.

6. The risk control rejection user data processing system according to claim 5, characterized in that: The newly added external data parsing module is used to: In response to the matching analysis result indicating that there is newly added external data in the external database, selecting a corresponding data parsing script based on the matching of the newly added external data; The user repayment behavior information in the data parsing script is parsed to generate the first user tag generation result, and the first user tag generation result is used to represent the first user repayment behavior tag.

Images