Credit overdue risk prediction method and device, electronic equipment and medium

Abstract

The invention provides a credit overdue risk prediction method and device, electronic equipment and a medium, and relates to the technical field of graph data processing, the method comprises the following steps: obtaining target graph data constructed by taking a target object as a central node, the target graph data comprising at least one entity node having an association relationship with the target object, each entity node is connected with a center node corresponding to the target object through a target edge, and the target edge has a target edge text representing an association relationship; determining a target edge weight corresponding to at least one target edge text by matching at least one target edge text in the target map data with a keyword score table, the keyword score table comprising at least one keyword and a risk score corresponding to each keyword; and predicting the credit overdue risk of the target object based on the target map data and the target edge weight corresponding to the at least one target edge text. The method can improve the recognition accuracy of the credit overdue risk corresponding to the target object.

Description

Technical Field

[0001] The present disclosure relates to the technical field of graph data processing, and in particular to a method, device, electronic device and medium for predicting credit overdue risk. Background Art

[0002] In the credit risk control scenario, it is often necessary to assess the user's credit overdue risk. In related technologies, a graph model (or graph data) is formed by modeling users and their associated entities. The graph model consists of entity nodes, edges, and edge texts. By processing the graph model, the user's overall overdue risk is mined. Therefore, how to mine and process information on the graph model is an urgent problem to be solved in order to improve the accuracy of credit overdue risk identification. Summary of the invention

[0003] The present disclosure provides a method, device, electronic device and medium for predicting credit overdue risk, which can improve the prediction accuracy of credit overdue risk. The technical solution is as follows:

[0004] According to one aspect of the present disclosure, a method for predicting credit overdue risk is provided, the method comprising:

[0005] Obtain target graph data constructed with the target object as a central node, wherein the target graph data includes at least one entity node having an association relationship with the target object, each of the entity nodes is connected to the central node corresponding to the target object through a target edge, and the target edge has a target edge text representing the association relationship;

[0006] Determining a target edge weight corresponding to the at least one target edge text by matching at least one target edge text in the target graph data with a keyword score table, wherein the keyword score table includes at least one keyword and a risk score corresponding to each keyword, wherein the risk score represents credit risk information indicated by the keyword;

[0007] Based on the target graph data and the target edge weight corresponding to the at least one target edge text, the credit overdue risk of the target object is predicted.

[0008] According to another aspect of the present disclosure, a credit overdue risk prediction device is provided, the device comprising:

[0009] A first acquisition module is used to acquire target graph data constructed with the target object as a central node, wherein the target graph data includes at least one entity node having an association relationship with the target object, each of the entity nodes is connected to the central node corresponding to the target object through a target edge, and the target edge has a target edge text representing the association relationship;

[0010] A matching module, configured to determine a target edge weight corresponding to the at least one target edge text by matching at least one target edge text in the target graph data with a keyword score table, wherein the keyword score table includes at least one keyword and a risk score corresponding to each keyword, wherein the risk score represents credit risk information indicated by the keyword;

[0011] A prediction module is used to predict the credit overdue risk of the target object based on the target graph data and the target edge weight corresponding to the at least one target edge text.

[0012] According to one aspect of the present disclosure, an electronic device is provided, comprising: a processor and a memory storing a program, wherein the program comprises instructions, and when the instructions are executed by the processor, the processor executes the credit overdue risk prediction method as described above.

[0013] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause the computer to execute the credit overdue risk prediction method as described above.

[0014] According to another aspect of the present disclosure, a computer program product is provided, the computer program product comprising computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the above-mentioned credit overdue risk prediction method.

[0015] The beneficial effects brought by the technical solution provided by the embodiments of the present disclosure include at least:

[0016] By pre-setting a keyword score table, the keyword score table includes several keywords and the risk score corresponding to each keyword, and the risk score can represent the credit risk information indicated by the keyword; when processing the target graph data, the target edge weight corresponding to each target edge text can be determined directly by matching each target edge text in the target graph data with the keyword score table. Since the target edge weight is generated by matching with the keyword score table, and the risk score in the keyword score table can represent the credit risk information, the generated target edge weight can reflect the credit risk information (i.e., semantic information) of the target edge text, thereby realizing the extraction of semantic information of the target edge text. Moreover, since the keyword score table is pre-saved, it is realized by text matching and searching in the process of modeling the target edge text, which can improve the real-time modeling of the edge text compared with the method of semantic modeling of the edge text using a pre-trained language model. Furthermore, when identifying the credit overdue risk of the target object based on the target graph data and the target edge weight, the influence of other actual nodes can be referred to based on the weight level, thereby improving the recognition accuracy of the credit overdue risk corresponding to the target object. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Further details, features and advantages of the present disclosure are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:

[0018] Figure 1 A schematic diagram showing an example system in which various methods described herein may be implemented according to an exemplary embodiment of the present disclosure;

[0019] Figure 2 A flowchart of a credit overdue risk prediction method according to an exemplary embodiment of the present disclosure is shown;

[0020] Figure 3 is a schematic diagram of target graph data of an exemplary embodiment of the present disclosure;

[0021] Figure 4 A flowchart showing another credit overdue risk prediction method according to an exemplary embodiment of the present disclosure is shown;

[0022] Figure 5 A flowchart showing another method for predicting credit overdue risk according to an exemplary embodiment of the present disclosure is shown;

[0023] Figure 6 is a schematic diagram of a process of edge text aggregation based on sample graph data according to an exemplary embodiment of the present disclosure;

[0024] Figure 7 is a schematic diagram of the training process of the risk scoring model of an exemplary embodiment of the present disclosure;

[0025] Figure 8is a schematic diagram of a process for generating a keyword score table according to an exemplary embodiment of the present disclosure;

[0026] Fig. 9 A flowchart showing another method for predicting credit overdue risk according to an exemplary embodiment of the present disclosure is shown;

[0027] Fig.10 A system schematic diagram of a credit overdue risk prediction process according to an exemplary embodiment of the present disclosure is shown;

[0028] Fig.11 It is a structural schematic diagram of a credit overdue risk prediction device provided by an embodiment of the present disclosure;

[0029] Fig.12 A structural block diagram of an exemplary electronic device that can be used to implement the embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0030] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein, which are instead provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0031] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0032] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to". The term "based on" means "based at least in part". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments". The relevant definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc. mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units. It should be noted that the modifications of "one" and "multiple" mentioned in the present disclosure are illustrative and not restrictive. Those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more". The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are for illustrative purposes only, and are not used to limit the scope of these messages or information.

[0033] The solution of the present invention is described below with reference to the accompanying drawings, and the technical solution provided by the embodiment of the present invention is explained in detail through specific embodiments and their application scenarios.

[0034] In the credit risk control scenario, graph models (or graph data) are a very effective method for mining credit risks. The overall risk of target users is identified by modeling target users and their associated entities. All target users and corresponding associated entities constitute a graph. The association relationship between the target user and the associated entity is the edge of the graph. These association relationships are usually accompanied by some attributes. If the attribute is text, then this edge is called a text edge. In the modeling of text edges in the graph model, the statistical method based on calculating the weight of the text edge is not accurate enough, and the method based on the language model to calculate the text edge weight requires high computing resources and is time-consuming. In order to solve the problems existing in the related art, the present disclosure proposes a method that combines fine-tuning of pre-trained language models and keyword scoring matching, which achieves a balance between deep semantic modeling of the model and computing resource consumption. Moreover, after using this method to weight the text edges of the graph, the risk identification performance of the credit risk control model can also be improved.

[0035] Figure 1 A schematic diagram of an exemplary system in which various methods described herein may be implemented according to an exemplary embodiment of the present disclosure is shown. The system includes at least one first device 110 and a second device 120 .

[0036] The first device 110 may be a device that provides object data. The object data is used to generate graph data with the object as the central node. In the embodiment of the present disclosure, the object data may be address book data, purchase information data, social network data, etc.

[0037] The second device 120 is a device for identifying overdue credit risks. In the application stage, the second device 120 can generate graph data based on the object data uploaded by the first device, calculate edge weights on the graph data, and use the processed graph data to identify overdue credit risks.

[0038] Optionally, the second device 120 may also be a computer device with machine learning capabilities, for example, the computer device may be a terminal or a server. The second device 120 may also train a risk scoring model and a credit risk control model based on the collected graph data and credit overdue labels.

[0039] Please refer to Figure 2 , which shows a flow chart of a credit overdue risk prediction method according to an exemplary embodiment of the present disclosure. Figure 1 The second device 120 shown in the figure is used as an example for explanation. Figure 2 As shown, the method includes:

[0040] Step 201, obtain target graph data constructed with the target object as the central node, the target graph data includes at least one entity node having an association relationship with the target object, each entity node is connected to the central node corresponding to the target object through a target edge, and the target edge has a target edge text representing the association relationship.

[0041] In order to predict the overdue risk of the target object in the credit scenario, the object data of the target object can be collected and obtained with the authorization of the target object or the target user, with the target object as the central node, the object in the object data that has an association relationship with the target object as the entity node, the entity node and the central node corresponding to the target object are connected through the target edge, and the association relationship between the entity node and the target object is extracted as text as the target edge text corresponding to the target edge, so as to model the target graph data with the target object as the central node, so that the overall overdue risk of the target object can be identified based on the modeled target graph data later. The corresponding target graph data includes at least one entity node that has an association relationship with the target object, each entity node is connected to the central node corresponding to the target object through the target edge, and the target edge has a target edge text that represents the association relationship.

[0042] Among them, at least one entity node that has an association relationship with the target object may be other entity object nodes of the same type. For example, if the target object is a target user, the corresponding entity node may be other user nodes that have an association relationship with the target user; optionally, at least one entity node that has an association relationship with the target object may also be other entity object nodes of different types. For example, if the target object is a target user, the corresponding entity node may be an entity node such as a company, group, organization that has an association relationship with the target user; optionally, at least one entity node that has an association relationship with the target object may be both an entity object node of the same type and an entity object node of different types. For example, if the target object is a target user, the corresponding entity node may include other user nodes that have an association relationship with the target user, and may also include entity nodes such as a company, group, organization that has an association relationship with the target user.

[0043] The association relationship between the target object and each entity node can be determined by the object data of the target object obtained. For example, if the object data is the address book data of the target object, the entity node can be the communication object in the address book data, the association relationship can be the address book relationship between the target object and each communication object, and the address book relationship can be the remark name of the target object to each communication object. For example, if the remark name of the target object A to the communication object B is "Credit Card Xiao Zhang", the target edge text on the target edge between the target object A and the communication object B can be "Credit Card Xiao Zhang"; if the object data is the purchase record information of the target object, the entity node can be the purchased goods in the purchase record information, and the association relationship can be the purchase time of the target object to purchase the goods. For example, if the target object A purchased a product record on January 1, 2019, the target edge text on the target edge between the target object A and the purchased product can be "January 1, 2019". It should be noted that the object data obtained in the embodiment of the present disclosure is obtained after authorization by the target object, and the object data is data that can characterize the overdue risk of the target object in the credit risk control scenario. The embodiments of the present disclosure only use address book data and purchase record data as examples, and do not limit the specific object data.

[0044] Please refer to Figure 3 , which is a schematic diagram of target graph data of an exemplary embodiment of the present disclosure. Figure 3 As shown, taking the target graph data 301 constructed by the address book data of the target object as an example, the target graph data 301 includes a central node 302 corresponding to user A and 6 entity nodes 303 (user B, user C, user D, user E, user F and user G) that have an association relationship with user A, and each entity node 303 is connected to the central node 302 through a target edge, and each target edge has a target edge text. For example, the edge text between the user A node and the user B node is "Manager Zhang", the edge text between the user A node and the user C node is "Decoration and Transportation", and the edge text between the user A node and the user D node is "Music Teacher Zhang"; the edge text between the user A node and the user E node is "Xiao Ming's Mother"; the edge text between the user A node and the user F node is "Credit Card Xiao Zhang"; the edge text between the user A node and the user G node is "Credit Manager Li".

[0045] Step 202, by matching at least one target edge text in the target graph data with a keyword score table, a target edge weight corresponding to the at least one target edge text is determined, wherein the keyword score table includes at least one keyword and a risk score corresponding to each keyword, and the risk score represents the credit risk information indicated by the keyword.

[0046] In the related art, when processing graph data, it is generally necessary to perform semantic modeling on the edge text in the graph data. Method 1 is to model the edge text based on statistical word vectors, specifically by counting the number of occurrences of each word after word segmentation to obtain the vector representation of each edge text; Method 2 is to use a pre-trained language model to model the edge text, specifically by directly inputting all edge texts in the graph data into the pre-trained language model and outputting the vector representation of the edge text. However, using method 1 to model edge text cannot capture the semantic information of the edge text; although method 2 can capture the deep semantic information of the edge text, the pre-trained language model usually has a large number of parameters, consumes a lot of computing resources to run the model, and runs the model slowly. When there are many edge texts in the graph data, it cannot meet the high real-time modeling requirements.

[0047] In view of the problems in the related art, in order to capture the semantic information of the text, the real-time modeling requirements of the edge text can also be met. The disclosed embodiment pre-provides a keyword score table, which includes several keywords and the risk score corresponding to each keyword, and the size of the risk score can represent the credit risk information indicated by the keyword; illustratively, the higher the risk score, the greater the credit risk indicated by the keyword, and the lower the risk score, the lower the credit risk indicated by the keyword, that is, the keyword score table pre-stores the semantic information represented by several keywords in the credit risk control scenario. When processing the target graph data, the target edge weight corresponding to each target edge text can be determined directly by matching each target edge text in the target graph data with the keyword score table. Since the target edge weight is generated by matching with the keyword score table, and the risk score in the keyword score table can represent the credit risk information, the generated target edge weight can reflect the credit risk information (i.e., semantic information) of the target edge text, thereby realizing the extraction of semantic information of the target edge text. In addition, since the keyword score table is pre-saved, it is implemented through text matching and searching in the process of modeling the target edge text. Compared with the method of using a pre-trained language model to perform semantic modeling on the edge text, the real-time modeling of the edge text can be improved.

[0048] Among them, in the credit risk control scenario, the keywords in the keyword score table can be collected by collecting several segmented words or words that can reflect credit risk information; or, the keywords in the keyword score table can also be obtained by collecting graph data in several credit risk control scenarios and segmenting the edge text in the graph data; and the risk score corresponding to each keyword in the keyword score table can be set by business personnel based on actual experience; or, a risk scoring model can be pre-trained, and the risk value of each keyword is scored by the risk scoring model. Exemplarily, the keyword score table can be shown in Table 1.

[0049] Table 1

[0050] word Risk Score manager 0.231 Mother 0.523 teacher 0.108 loan 0.754 credit card 0.841 … …

[0051] As can be seen from Table 1, the risk score corresponding to the keyword "credit card" is 0.841; and the higher the risk score, the higher the credit risk indicated by the keyword.

[0052] like Figure 3 As shown, after the target edge text in the target graph data 301 is subjected to variable weight calculation, the processed target graph data 304 can be obtained, and each target edge of the processed target graph data 304 is replaced with the target edge weight corresponding to the target edge text. For example, the edge weight of the edge between the user A node and the user B node is 0.231, the edge weight of the edge between the user A node and the user C node is 0.587, the edge weight of the edge between the user A node and the user D node is 0.108, the edge weight of the edge between the user A node and the user E node is 0.523, the edge weight of the edge between the user A node and the user F node is 0.841, and the edge weight of the edge between the user A node and the user G node is 0.754.

[0053] Step 203: predicting the credit overdue risk of the target object based on the target graph data and the target edge weight corresponding to at least one target edge text.

[0054] After modeling the target edge text in the target graph data and obtaining the target edge weight, the credit overdue risk of the target object can be predicted based on the target graph data and the target edge weight corresponding to each target edge text. Since the target edge weight can reflect the impact of each entity node on the credit risk of the target object, the subsequent credit risk control model can refer to the impact of other entity nodes based on the weight when identifying the credit overdue risk of the target object based on the target graph data and the target edge weight, thereby improving the accuracy of identifying the corresponding credit overdue risk of the target object.

[0055] A possible implementation method for predicting the credit overdue risk of the target object based on the target graph data and the target edge weight is: replace the target edge text in the target graph data with the target edge weight to obtain the fine-tuned target graph data, and input the fine-tuned target graph data into the credit risk control model, and the credit risk control model predicts the credit overdue risk of the target object based on the fine-tuned target graph data.

[0056] Optionally, when predicting credit overdue risk based on the fine-tuned target graph data, in order to reduce the amount of information processing of the credit risk control model, an edge weight threshold may be set, and the fine-tuned target graph data may be processed based on the edge weight threshold. Specifically, the target edges and entity nodes whose target edge weights are lower than the edge weight threshold are removed, and the target edges and entity nodes whose target edge weights are higher than the edge weight threshold are retained to obtain the processed target graph data, and then the credit overdue risk is identified based on the processed target graph data, thereby avoiding the influence of entity nodes with lower edge weights on the accuracy of credit overdue risk identification.

[0057] In summary, the disclosed embodiment provides a method for predicting credit overdue risk: by pre-setting a keyword score table, the keyword score table includes a number of keywords and a risk score corresponding to each keyword, and the risk score can represent the credit risk information indicated by the keyword; so that when processing the target graph data, the target edge weight corresponding to each target edge text can be directly determined by matching each target edge text in the target graph data with the keyword score table. Since the target edge weight is generated by matching with the keyword score table, and the risk score in the keyword score table can represent the credit risk information, the generated target edge weight can reflect the credit risk information (i.e., semantic information) of the target edge text, thereby realizing the extraction of semantic information of the target edge text. Moreover, since the keyword score table is pre-saved, it is realized by text matching and searching in the process of modeling the target edge text, which can improve the real-time modeling of the edge text compared with the method of using a pre-trained language model to perform semantic modeling on the edge text. Furthermore, when identifying the credit overdue risk of the target object based on the target graph data and the target edge weight, the influence of other actual nodes can be referred to based on the weight, thereby improving the accuracy of identifying the corresponding credit overdue risk of the target object.

[0058] When determining edge weights, there are multiple situations for the matching results between edge text and the keyword score table. For a single edge text, the following situations may exist: no keyword is matched; a single keyword can be matched; multiple keywords are matched; in order to ensure the smooth calculation of edge weights, corresponding edge weight calculation methods are set for each situation.

[0059] Please refer to Figure 4 , which shows a flowchart of another credit overdue risk prediction method according to an exemplary embodiment of the present disclosure. Figure 1 The computer device 120 shown in the figure is used as an example for explanation. Figure 4 As shown, the method includes:

[0060] Step 401, obtain target graph data constructed with the target object as the central node, the target graph data includes at least one entity node having an association relationship with the target object, each entity node is connected to the central node corresponding to the target object through a target edge, and the target edge has a target edge text representing the association relationship.

[0061] The implementation of step 401 may refer to step 201 , and this embodiment will not be described in detail here.

[0062] In a possible implementation, the target graph data is generated from the address book data of the target object, the entity node is the address book object of the target object, and the target edge text is the remark label information of the address book object.

[0063] Step 402: for each target edge text in the target graph data, if there is a keyword matching the target edge text in the keyword score table, determine the target edge weight corresponding to the target edge text based on the risk score of the matched keyword.

[0064] Before matching the target edge text in the target graph data with the keyword score table, each target edge text may be segmented to obtain at least one target segmentation corresponding to each target edge text, and then the target segmentation is matched with the keywords in the keyword score table to determine whether there is a keyword matching any target segmentation in the keyword score table. For example, if the target edge text is "Music Teacher Zhang", the target edge text may be segmented into "Music" and "Teacher Zhang", and then the keyword score table may be searched for a keyword matching "Music", and the keyword score table may be searched for a keyword matching "Teacher Zhang".

[0065] Optionally, the target side text may not be segmented, and the target side text may be directly matched with the keywords in the keyword scoring table to determine whether there are keywords in the keyword scoring table that match the target side text. This can also be understood as determining whether there are keywords in the keyword scoring table in the target side text.

[0066] For each target edge text in the target graph data, when matching, there are situations where keywords can be matched and situations where keywords cannot be matched; if there are keywords matching the target edge text in the keyword score table, the target edge weight of the target edge text can be determined directly based on the risk score of the matched keyword.

[0067] Considering the differences in the length of the target edge text, some target edge texts can match multiple keywords, corresponding to multiple risk scores, and some target edge texts can match a single keyword, corresponding to one risk score; and each target edge text corresponds to a target edge weight. In the case of matching a keyword, there are different ways to determine the target edge weight according to the number of matched keywords. Correspondingly, in an exemplary example, step 402 may also include step 402A and step 402B.

[0068] Step 402A: If the target edge text matches a single keyword, the risk score corresponding to the keyword is determined as the target edge weight corresponding to the target edge text.

[0069] In the case where the target edge text matches a single keyword, the risk score corresponding to the keyword can be directly obtained as the target edge weight corresponding to the target edge text. For example, if the target edge text is "Lobby Manager", based on the keyword score table shown in Table 1, the keyword "Manager" can be matched, and the risk score "0.231" corresponding to "Manager" is determined as the target edge weight corresponding to the target edge text.

[0070] Step 402B, if the target edge text matches at least two keywords, determine the target risk value based on the risk scores corresponding to the at least two keywords, and determine the target risk value as the target edge weight corresponding to the target edge text, and the target risk value is any one of the average, maximum or minimum values ​​of the at least two risk scores.

[0071] In the case where the target edge text matches multiple keywords (i.e., at least two keywords), it is necessary to determine a target risk value based on at least two risk scores corresponding to the at least two keywords, and then determine the target risk value as the target edge weight corresponding to the target edge text.

[0072] When determining the target edge weight based on multiple risk scores, the selected calculation strategy can be determined by the modeling requirements. For example, the maximum value among multiple risk scores can be determined as the target risk value (i.e., target edge weight); the minimum value among multiple risk scores can be determined as the target risk value (i.e., target edge weight); or the average value of multiple risk scores can be determined as the target risk value (i.e., target edge weight); or, a weighted average value can be calculated for multiple risk scores to determine the target risk value (i.e., target edge weight), and the weight of each risk score in the weighted average calculation process can be determined by the proportion of the number of words of the keyword in the target edge text.

[0073] For example, if the target edge text is "loan manager", and it matches the keywords "loan" and "manager", "loan" corresponds to a risk score of 0.754, and "manager" corresponds to a risk score of "0.231", then the target edge weight corresponding to the target edge text can take the maximum risk score of "0.754".

[0074] Optionally, the process of determining the target edge weight corresponding to the target text edge according to the risk scores corresponding to the matched multiple keywords may further include step 402B1 and step 402B2.

[0075] Step 402B1, for each matched keyword, determine the weight corresponding to each keyword based on the proportion of the number of words occupied by each keyword in the target edge text.

[0076] Step 402B2: Determine the target risk value based on the weighted weight and risk score corresponding to each keyword.

[0077] Considering that the word segments with more characters in the text can better represent the summary information of the keywords, when multiple keywords are matched, the reference proportion of the risk score when calculating the edge weight of the edge text can be determined according to the proportion of the number of characters occupied by the keywords in the target edge text. In a corresponding possible implementation, for each matched keyword, the proportion of the number of characters occupied by each keyword in the target edge text is first obtained, and the weighted weight corresponding to each keyword is determined, and then when the risk scores of multiple keywords are integrated, the target risk value is determined according to the weighted weight and the risk score.

[0078] For example, if the target edge text is "credit card manager", and the keywords "credit card" and "manager" are matched, the corresponding risk score of "credit card" is "0.841", and the corresponding risk score of "manager" is "0.231". When determining the target edge weight corresponding to the target edge text, the word ratio of the keyword "credit card" in the target edge text is obtained: 3 / 5, and the word ratio of the keyword "manager" in the target edge text is obtained: 2 / 5, then the target edge weight = (0.841×3 / 5+0.231×2 / 5) = 0.597. In other possible scenarios, corresponding relevance weights can be set for keywords according to the frequency of occurrence of keywords in the credit overdue scenario, and the relevance weight is positively correlated with the frequency of occurrence, that is, the higher the frequency of occurrence, the higher the relevance weight; so that when the risk scores of multiple keywords are integrated, the relevance of the keywords to the credit overdue scenario can also be considered. A target vocabulary related to the credit overdue scenario is pre-set, and the target vocabulary includes several words and the correlation weight corresponding to each word; when comprehensively determining the target edge weight corresponding to the target edge text according to the risk scores of the matched multiple keywords, it is possible to first determine whether the keyword exists in the target vocabulary. If so, obtain the corresponding correlation weight, and then comprehensively determine the target edge weight according to the correlation weight and the risk score.

[0079] Exemplarily, the target vocabulary may be as shown in Table 2.

[0080] Table 2

[0081] vocabulary Relevance Weight credit card 0.8 loan 0.9 manager 0.3 … …

[0082] Based on the target vocabulary shown in Table 2, if the target edge text is "credit card manager", and the keywords "credit card" and "manager" are matched, the corresponding risk score of "credit card" is "0.841", and the corresponding risk score of "manager" is "0.231"; and it is determined that "credit card" and "manager" are both words related to the credit overdue scenario, then from Table 2, it can be obtained that the corresponding association weight of "credit card" is "0.8", and the corresponding association weight of "manager" is "0.3", then the target edge weight = (0.841×0.8+0.231×0.3) = 0.7421.

[0083] Step 403: for each target edge text in the target graph data, if there is no keyword matching the target edge text in the keyword score table, a preset score is determined as the target edge weight corresponding to the target edge text.

[0084] For target edge texts that do not match keywords, in order to ensure the smooth calculation of edge weights, a preset score is provided, and the preset score is determined as the target edge weight corresponding to the target edge text. The preset score can be a default value set by business personnel based on experience; the average of all risk scores in the keyword score table can also be determined as the preset score; or the preset score can also be the risk score output by inputting "None" into the risk scoring model.

[0085] Step 404 , predicting the credit overdue risk of the target object based on the target graph data and the target edge weight corresponding to at least one target edge text.

[0086] The implementation of step 404 may refer to step 203, and this embodiment will not be described in detail here.

[0087] In this embodiment, for different matching results between the target edge text and the keyword score table, a method for determining the target edge weight is set respectively to meet the edge weight determination requirements under different matching results and ensure the smooth modeling of the edge text in the target graph data.

[0088] It can be seen from the above embodiments that when predicting the credit overdue risk of the target object, the keywords and risk scores in the keyword score table play a relatively important role. In order to make the keyword score table more in line with the modeling requirements in the credit risk control scenario, a risk scoring model is pre-trained to generate the keyword score table.

[0089] Please refer to Figure 5 , which shows a flowchart of another credit overdue risk prediction method according to an exemplary embodiment of the present disclosure. Figure 1 The computer device 120 shown in the figure is used as an example for explanation. Figure 5 As shown, the method includes:

[0090] Step 501, obtaining at least one sample graph data, each sample graph data is composed of at least two sample nodes, at least one sample edge and sample edge text on each sample edge, and at least two sample nodes include at least one node corresponding to a sample object.

[0091] Step 502: Obtain a credit overdue label corresponding to the sample object. The credit overdue label is used to indicate whether the sample object has any credit overdue behavior.

[0092] When training a risk scoring model, in order to enable the risk scoring model to have the ability to identify credit risk information represented by text, in a possible implementation, a number of sample graph data constructed with sample objects as central nodes can be obtained, and credit overdue labels corresponding to the sample objects can be obtained. The credit overdue labels are used to indicate whether the sample objects have credit overdue behavior, so that the subsequent risk scoring model can use the credit overdue labels as supervision standards to identify credit risk information in the text.

[0093] The sample graph data can be constructed in a similar manner to the target graph data. The sample graph data can also be constructed based on the address book data. Each sample graph data is composed of at least two sample nodes, at least one sample edge, and sample edge text on each sample edge. At least two sample nodes include at least one node corresponding to a sample object. The sample edge text represents the association relationship between the two nodes.

[0094] For example, the credit overdue label can be represented by "0" and "1". The credit overdue label of "0" indicates that the sample object has no credit overdue behavior, and the credit overdue label of "1" indicates that the sample object has credit overdue behavior. It should be noted that if the sample object has a loan behavior but has not yet reached the repayment period, it also means that there is no credit overdue behavior.

[0095] Since the risk scoring model can only learn the credit risk information corresponding to the word if the text appears a sufficient number of times, it is necessary to construct a sufficient number of sample graph data during the model training phase and set the corresponding credit overdue label for each sample object in the sample graph data.

[0096] It should be noted that when collecting information to construct sample graph data, the collected information is authorized by the user and privacy is processed during processing.

[0097] Step 503: training a risk scoring model based on at least one sample graph data and a credit overdue label corresponding to each sample object.

[0098] After constructing the training set data, the risk scoring model can be trained based on at least one sample graph data and the credit overdue label corresponding to each sample object, so that the trained risk scoring model can identify the credit risk information in the input text.

[0099] Among them, the risk scoring model can use a pre-trained language model, such as pre-trained Bert, Transformer and other models and their variant structures. The model training task is a 0-1 binary classification task.

[0100] Since the risk scoring model is mainly used to identify credit risk information in text, before training the risk scoring model, the sample graph data needs to be processed to obtain a training data set of the type of text + credit overdue label. Correspondingly, in an exemplary example, step 503 may include step 503A and step 503B.

[0101] Step 503A: perform edge text aggregation on the sample graph data based on the sample objects to obtain a text string corresponding to at least one sample object, where the text string is a collection of sample edge texts on sample edges associated with the sample objects in the sample graph data.

[0102] Considering that the information in the sample graph data that can reflect whether the sample object has overdue risk is generally the entity node that has an association relationship with the sample object, when extracting text for risk prediction from the sample graph data, the sample graph data can be edge-text aggregated based on the sample object, that is, the sample edge text on the sample edge associated with the sample object in the sample graph data is aggregated to obtain the text string corresponding to the sample object as a subsequent training sample.

[0103] Please refer to Figure 6 , which is a schematic diagram of the process of edge text aggregation based on sample graph data according to an exemplary embodiment of the present disclosure. Figure 6 As shown, when edge text aggregation is performed on the sample graph data 601 with the sample object as the user A node, the set of edge texts on multiple sample edges associated with the user A node can be used as the text string 602 after the sample object aggregation; the text string 602 is "User A: Manager Zhang, decorator and mover, music teacher Zhang, Xiao Ming's mother".

[0104] Since a sample object may have multiple associated sample edges, when performing edge text aggregation, the order of sample edge texts corresponding to different sample edges, whether text deduplication is required, and whether separators are used between different sample edge texts can all be customized according to the requirements of the risk scoring model.

[0105] Specifically, when aggregating sample edge texts on multiple sample edges associated with a sample object, the order of each sample edge text in the aggregated text string is randomly determined; or, the order of the sample edge texts can be defined by business personnel, and the order of the sample edge texts does not affect the subsequent training and scoring of the risk scoring model; if the same sample edge text exists in the sample edge texts associated with the same sample object, the same sample edge text can be retained, that is, the sample edge text is not deduplicated; or, the sample edge text can also be deduplicated, and when deduplication is selected, the corresponding number of text occurrences can be set for the sample edge text, so that the subsequent risk scoring model can also refer to the number of occurrences of the sample edge text when identifying the credit overdue risk represented by the text. In addition, when aggregating multiple sample edge texts, no delimiter can be set between different sample text edges; or delimiters can be set to separate different sample edge texts, for example, delimiters such as "," " / " "," ";" can be used; so that the risk scoring model can accurately distinguish different sample edge texts.

[0106] Step 503B: training a risk scoring model based on the text string corresponding to at least one sample object and the credit overdue label corresponding to each sample object.

[0107] After edge text aggregation processing is performed on multiple sample graph data, text strings corresponding to several sample objects can be obtained, and the credit overdue label corresponding to each sample object is merged with the text string and used together as training data to train the risk scoring model.

[0108] Exemplarily, the training data after the text string and the credit overdue label are combined can be shown in Table III.

[0109] Table 3

[0110] Sample Object Aggregate text Credit overdue label User A Manager Zhang, renovation and moving, music teacher Zhang, Xiao Ming’s mother 0 User B Loan Manager Li, Credit Card Xiao Zhang 1 User C Zhang A, Li B, Wang C, Teacher Chen 0 User D Lobby Manager, Mr. Ma, Xiao Li 0 … … …

[0111] It can be seen from Table 3 that the aggregated text corresponding to user A, "Manager Zhang, decoration mover, music teacher Zhang, Xiao Ming's mother", is the text string after edge text aggregation based on user A. The credit overdue label corresponding to user A is "0", indicating that user A has no credit overdue behavior.

[0112] The function of the risk scoring model is to identify the credit risk information represented by the input text, that is, the output of the risk scoring model is the risk score of the credit risk represented by the text. When training the risk scoring model, the credit overdue label is used as the supervision value for model training. Optionally, step 503B may also include step 503B1 and step 503B2.

[0113] Step 503B1: for each sample object, input the text string corresponding to the sample object into the risk scoring model to obtain the predicted score output by the risk scoring model, where the predicted score represents the credit risk information indicated by the text string.

[0114] During the risk scoring model training process, the text string corresponding to the sample object is input into the risk scoring model, and the risk scoring model uses semantic understanding to predict the level of credit risk represented by the text string, and outputs a corresponding predicted score for the text string, which represents the credit risk information indicated by the text string.

[0115] Step 503B2: training a risk scoring model based on the loss value between the predicted score and the credit overdue label corresponding to the sample object.

[0116] Since each sample object is marked with an overdue risk label to characterize whether the sample object has an overdue risk, in order to enable the risk scoring model to learn the credit risk information in the text string, the model parameters of the risk scoring model can be updated based on the loss value between the predicted score and the credit overdue label corresponding to the sample object. Specifically, when calculating the loss value, the cross entropy loss between the predicted score and the credit overdue label can be used as the loss function.

[0117] The risk scoring model is trained multiple times using the text strings and credit overdue labels corresponding to the sample objects. When the loss value is less than the preset value, the risk scoring model training is determined to be completed. It can be used for subsequent identification of credit risk information on text or keywords to achieve the purpose of keyword scoring.

[0118] Please refer to Figure 7 , which is a schematic diagram of the training process of the risk scoring model of the exemplary embodiment of the present disclosure. Figure 7 As shown, after obtaining the training data (the text string 701 and the credit overdue label 704 corresponding to each sample object), the text string 701 corresponding to each sample object can be input into the risk scoring model 702, and the risk scoring model 702 identifies the credit risk of the input text and outputs the predicted score 703 corresponding to each text string; and according to the cross entropy loss between the predicted score 703 and the credit overdue label 704, the risk scoring model 702 is fine-tuned or trained; so that after multiple rounds of training, a risk scoring model after training is obtained.

[0119] Step 504: Generate a keyword score table based on at least one sample graph data and the trained risk scoring model.

[0120] After the trained risk scoring model is obtained, the risk scoring model can be used to score the credit risk of keywords to obtain the risk score corresponding to each keyword.

[0121] For the keywords in the keyword score table, the sample edge text in the sample graph data used in model training can also be reused. Since the amount of sample graph data required for model training is relatively comprehensive, the extracted keywords can cover most of the keywords in the credit risk scenario. In a possible implementation, a keyword score table can be generated based on several sample graph data and a trained risk scoring model.

[0122] Since the keyword score table is composed of keywords and risk scores, and there are differences in the edge texts in the graph data during actual application, directly inputting the sample edge texts in the sample graph data into the risk scoring model for scoring may cause the subsequent keyword matching to fail. Therefore, before scoring, the sample edge texts in the sample graph data need to be segmented. Correspondingly, in an exemplary example, step 504 may include steps 504A to 504C.

[0123] Step 504A: perform word segmentation processing on the sample edge text in at least one sample graph data to obtain at least one keyword.

[0124] Among them, word segmentation is to divide a continuous text string into a meaningful sequence of words. In the embodiment of the present disclosure, the word segmentation process is to divide the sample edge text in the sample graph data into multiple words. The word segmentation process can adopt a rule-based method, that is, using predefined dictionaries and rules for word segmentation, such as the forward maximum matching method (Maximum Matching Algorithm, MM) and the reverse maximum matching method (Reverse Maximum Matching, RMM); it can also adopt a statistical method, that is, using a large-scale corpus and performing word segmentation through statistical information, such as the Hidden Markov Model (HMM), the Conditional Random Field (CRF); it can also adopt a deep learning-based method, that is, using a neural network model, such as a recurrent neural network (RNN), a long short-term memory network (LSTM), a Transformer, etc. The embodiment of the present disclosure does not limit the word segmentation process.

[0125] In one possible implementation, a preset word segmentation processing method is used to perform word segmentation processing on the sample edge text in the sample graph data to divide the sample edge text into multiple meaningful words, obtain a word segmentation result, and directly determine the word segmentation result as at least one keyword for use in a subsequent scoring model to score the risk score.

[0126] Optionally, considering that only when a word appears enough times in the sample edge text can the risk scoring model learn the risk information of the word, in order to ensure the accuracy of the scoring, a certain number of high-frequency words are selected from the word segmentation results according to the actual modeling requirements as at least one keyword. Exemplarily, the high-frequency words can be the words with the highest frequency of appearance in the word segmentation results, and the value of K can be 3000, that is, the TOP3000 words are selected as keywords.

[0127] Exemplarily, the keyword word list obtained after word segmentation processing is performed on the sample text edges in the sample graph data can be shown in Table 4.

[0128] Table 4

[0129] Keywords Word frequency manager 456 Mother 235 teacher 157 loan 42 credit card 23 … …

[0130] The keywords shown in Table 4 are the words obtained after word segmentation and high-frequency word screening of the sample text, and the word frequency is the frequency of occurrence of the keyword in the word segmentation results.

[0131] Step 504B: input at least one keyword into the trained risk scoring model to obtain a risk score corresponding to each keyword.

[0132] After obtaining a number of keywords, the keywords can be input into the trained risk scoring model for credit risk identification, and the risk scoring model outputs the risk score corresponding to each keyword.

[0133] Step 504C: associate at least one keyword with the risk score corresponding to each keyword to obtain a keyword score table.

[0134] After obtaining several keywords and the risk scores corresponding to each keyword, the scoring results can be saved, that is, the keywords and their corresponding risk scores are associated to generate a keyword score table, which is convenient for the subsequent reuse of the keyword score table to calculate the edge weight of the edge text. This ensures the reusability of the scoring results while retaining the deep semantic modeling capabilities.

[0135] Please refer to Figure 8 , which is a schematic diagram of the generation process of the keyword score table of the exemplary embodiment of the present disclosure. Figure 8 As shown, several keywords 801 obtained after word segmentation processing are input into the trained risk scoring model 802, and the risk score 803 corresponding to each keyword is output, and then the keyword 801 and the risk score 803 are associated and saved to obtain a keyword score table 804. It should be noted that when the risk scoring model is scoring, it is not necessary to input the word frequency corresponding to the keyword, and the word frequency corresponding to each keyword may not be included in the keyword score table.

[0136] Step 505, obtain target graph data constructed with the target object as the central node, the target graph data includes at least one entity node having an association relationship with the target object, each entity node is connected to the central node corresponding to the target object through a target edge, and the target edge has a target edge text representing the association relationship.

[0137] Step 506, determining a target edge weight corresponding to at least one target edge text by matching at least one target edge text in the target graph data with a keyword score table, wherein the keyword score table includes at least one keyword and a risk score corresponding to each keyword, and the risk score represents the credit risk information indicated by the keyword.

[0138] Step 507: predicting the credit overdue risk of the target object based on the target graph data and the target edge weight corresponding to at least one target edge text.

[0139] The implementation of steps 505 to 507 may refer to the above embodiment, and will not be described in detail in this embodiment.

[0140] In this embodiment, the risk scoring model is trained by aggregating edge text and credit overdue labels, so that the trained risk scoring model can identify the credit risk of the input text; so that the risk scoring model can be used to identify and score the credit risk of keywords in the future, and the risk score corresponding to the keyword can be obtained, so that the risk score in the keyword score table can reflect the credit risk information of the text; and then when the keyword score table is used to look up the edge text for weighting, the text semantic information and real-time issues required for edge text modeling can be taken into account; in addition, since the risk scoring model does not change the score of the keyword, the reusability of the keyword scoring results is guaranteed.

[0141] In other possible application scenarios, after completing the edge weight calculation for the target graph data, the target graph data with updated edge weights can also be used as new training samples to fine-tune the credit risk control model to identify the credit overdue risk of the target object.

[0142] Please refer to Fig. 9 , which shows a flowchart of another credit overdue risk prediction method according to an exemplary embodiment of the present disclosure. Figure 1 The computer device 120 shown in the figure is used as an example for explanation. Fig. 9 As shown, the method includes:

[0143] Step 901, obtain target graph data constructed with the target object as the central node, the target graph data includes at least one entity node having an association relationship with the target object, each entity node is connected to the central node corresponding to the target object through a target edge, and the target edge has a target edge text representing the association relationship.

[0144] Step 902, by matching at least one target edge text in the target graph data with a keyword score table, a target edge weight corresponding to the at least one target edge text is determined, wherein the keyword score table includes at least one keyword and a risk score corresponding to each keyword, and the risk score represents the credit risk information indicated by the keyword.

[0145] The implementation of step 901 and step 902 can refer to the above embodiment, and this embodiment will not be described in detail here.

[0146] Step 903: Obtain a target overdue tag corresponding to the target object. The target overdue tag is used to indicate whether the target object has overdue behavior.

[0147] Step 904: Replace the target edge text in the target graph data with the target edge weight to obtain the fine-tuned target graph data.

[0148] Among them, the credit risk control model is used to identify the credit overdue risk of the target object based on the input graph data. In the credit risk control model training stage, the training data taken includes the graph data corresponding to the user and the associated credit overdue label. Since the above embodiment provides a method for calculating the edge weight of the edge text in the graph data, the credit risk control model can also be trained using the graph data with edge weights as training data, so that the credit risk control model can better learn and understand the impact of each node in the graph data on the user's credit overdue behavior.

[0149] In one possible implementation, after obtaining the target edge weights of each target edge text in the target graph data, the target edge text in the target graph data can be directly replaced with the target edge weight to obtain fine-tuned target graph data and a target overdue label corresponding to the merged target object. The target overdue label is used to indicate whether the target object has overdue behavior, and then the fine-tuned target graph data and target overdue label are used as training data to train the credit risk control model.

[0150] Step 905: Based on the fine-tuned target graph data and the target overdue label, a credit risk control model is trained. The credit risk control model is used to identify the credit overdue risk of the target object.

[0151] In an illustrative example, step 905 may also include step 905A and step 905B.

[0152] Step 905A, input the fine-tuned target graph data into the credit risk control model to obtain the credit overdue score output by the credit risk control model.

[0153] Step 905B, training a credit risk control model based on the loss between the credit overdue score and the target overdue label.

[0154] During the credit risk control model training phase, the fine-tuned target graph data can be input into the credit risk control model for credit risk identification, and the credit overdue score output by the credit risk control model can be obtained. Then, based on the cross entropy loss between the credit overdue score and the target overdue label, the model parameters of the credit risk control model can be fine-tuned. After multiple rounds of training, until the loss value is less than the preset value, it is determined that the credit risk control model training is completed. The trained credit risk control model can be used to identify the credit overdue risk of the target object.

[0155] This embodiment provides another application scenario of edge weights. The target graph data with updated edge weights can also be used as new training samples to fine-tune the credit risk control model, so that the credit risk control model can better learn and understand the impact of each node in the graph data on the user's credit overdue behavior, thereby improving the model performance of the credit risk control model.

[0156] In another possible application scenario, the sample graph data may be directly fine-tuned to train the credit risk control model. The corresponding process of training the credit risk control model may include the following steps 1 to 7:

[0157] Step 1: Obtain at least one sample graph data, each sample graph data is composed of at least two sample nodes, at least one sample edge and sample edge text on each sample edge, and at least two sample nodes include at least one node corresponding to a sample object.

[0158] Step 2: Obtain the credit overdue label corresponding to the sample object. The credit overdue label is used to indicate whether the sample object has any credit overdue behavior.

[0159] Step 3: Train a risk scoring model based on at least one sample graph data and a credit overdue label corresponding to each sample object.

[0160] Step 4: Generate a keyword score table based on at least one sample graph data and the trained risk scoring model.

[0161] Step 5: Determine the sample edge weight corresponding to the sample edge text by matching at least one sample edge text in the sample graph data with the keyword score table.

[0162] Step 6: Replace the sample edge text in the sample graph data with the sample edge weight to obtain the fine-tuned sample graph data.

[0163] Step 7: Train the credit risk control model based on the sample graph data and credit overdue labels.

[0164] After collecting the sample graph data and the associated credit overdue labels, on the one hand, the sample graph data and the credit overdue labels can be used to train the risk scoring model, and the trained risk scoring model can be used to generate a keyword score table; on the other hand, after generating the keyword score table, the keyword score table can be used to calculate the edge weights of the sample graph data to fine-tune the sample graph data and obtain the sample graph data containing the sample edge weights, so that the fine-tuned sample graph data and the associated credit overdue labels can be used to train the credit risk control model later. This allows the training of two models to be achieved by reusing the same sample graph data.

[0165] Based on the implementation method described in the above embodiment, please refer to Fig.10 , which shows a system schematic diagram of a credit overdue risk prediction process according to an exemplary embodiment of the present disclosure. Fig.10 As shown, this principle mainly includes the following stages:

[0166] User node text aggregation stage 1001: User A is connected to other nodes through association relationships. Other nodes can all be user nodes, or actually other types of nodes. There is a corresponding edge text on each edge between user A and other nodes. The texts of all edges associated with user A are aggregated to user A. The order of different edge texts, whether text deduplication is required, and the separators between different edge texts can be customized according to model requirements. Therefore, edge text aggregation is performed on user A to obtain a new text string. Similarly, user node text aggregation is performed on other user nodes to obtain corresponding node aggregation text strings (i.e., the text strings corresponding to each sample object in the above embodiment).

[0167] Overdue performance label merging stage 1002: For users with post-loan overdue performance labels, the text string of the user node aggregation user and the overdue performance label are merged. The overdue label (i.e., the credit overdue label in the above embodiment) is 1, which means that the user is overdue, and the overdue label is 0, which means that the user is not overdue. It is emphasized that users have post-loan overdue performance labels because some users have not yet reached the repayment period after the loan, and therefore have no repayment overdue performance. The user edge text string after merging the labels can be used for subsequent fine-tuning of the pre-trained language model (i.e., the risk scoring model in the above embodiment).

[0168] Pre-trained language model fine-tuning stage 1003: Use the user node aggregate edge text string with overdue performance labels to fine-tune the pre-trained language model so that the fine-tuned pre-trained language model can identify risk information in the text. Specifically, the pre-trained language model can use pre-trained Bert, Transformer and other models and their variant structures. The model fine-tuning task is a 0-1 binary classification task, and the loss function of the model fine-tuning is the cross entropy loss. The trained model is used for subsequent text keyword scoring.

[0169] Edge text segmentation statistics and scoring stage 1004: Segment the text on each edge in the sample graph data, then count the word frequencies of all segmented texts, sort them in descending order, and form a word list. Then select a certain number of high-frequency words according to actual modeling requirements, such as top3000, and use them as input to the fine-tuned pre-trained language model for scoring, and save the scoring results (that is, obtain a keyword score table containing keywords and risk scores).

[0170] Edge weight calculation stage 1005: When weighting the edges in the graph data, match the word segmentation results of the text on each edge in the graph against the saved high-frequency word scoring results, and weight each edge according to the matching results. For example, if the word segmentation result of a certain edge text can only match the scoring result of one high-frequency word, then the scoring result of this high-frequency word is used as the weight of this edge. If the word segmentation result of a certain edge text can match the scoring results of two or more high-frequency words, then the average / maximum / minimum value of the scores of these matched high-frequency words can be used as the weight of this edge according to the modeling requirements. If a certain edge text does not match a word in the high-frequency word list, then this edge can be weighted as the average score of all words in the high-frequency word list.

[0171] The reason for scoring keywords (high-frequency words) is that the keyword scoring results are reusable. Scoring can be obtained by searching the word list directly according to the keywords, and the keywords can reflect the summary information of the edge text. Since the text of each edge may be different, if the entire text of each edge is scored, the edge text needs to be input into the fine-tuned pre-trained language model for scoring. Since the real-time scoring of the pre-trained language model requires high computing resources and is time-consuming, if a user is associated with many edges, the real-time scoring cannot be met. In addition, the reason for scoring only high-frequency keywords is that only when the number of times a word appears in all edge texts is sufficient to ensure that the pre-trained language model can learn the risk information corresponding to this word.

[0172] Credit risk prediction stage 1006: Input the fine-tuned graph data into the credit risk control model for credit risk prediction to predict the user's credit overdue risk.

[0173] Please refer to Fig.11 , which is a schematic diagram of the structure of a credit overdue risk prediction device provided by an embodiment of the present disclosure. For example, Fig.11 As shown, the device 1100 includes:

[0174] The first acquisition module 1101 is used to acquire target graph data constructed with the target object as the central node, wherein the target graph data includes at least one entity node having an association relationship with the target object, each of the entity nodes is connected to the central node corresponding to the target object through a target edge, and the target edge has a target edge text representing the association relationship;

[0175] A matching module 1102 is used to determine a target edge weight corresponding to the at least one target edge text by matching the at least one target edge text in the target graph data with a keyword score table, wherein the keyword score table includes at least one keyword and a risk score corresponding to each keyword, wherein the risk score represents credit risk information indicated by the keyword;

[0176] The prediction module 1103 is used to predict the credit overdue risk of the target object based on the target graph data and the target edge weight corresponding to the at least one target edge text.

[0177] Optionally, the matching module 1102 is further configured to:

[0178] For each of the target edge texts in the target graph data, if there is a keyword matching the target edge text in the keyword score table, determine the target edge weight corresponding to the target edge text based on the risk score of the matched keyword;

[0179] For each of the target edge texts in the target graph data, if there is no keyword matching the target edge text in the keyword score table, a preset score is determined as the target edge weight corresponding to the target edge text.

[0180] Optionally, the matching module 1102 is further configured to:

[0181] If the target edge text matches a single keyword, the risk score corresponding to the keyword is determined as the target edge weight corresponding to the target edge text;

[0182] If the target edge text matches at least two of the keywords, a target risk value is determined based on the risk scores corresponding to the at least two keywords, and the target risk value is determined as the target edge weight corresponding to the target edge text, and the target risk value is any one of the average, maximum or minimum values ​​of at least two risk scores.

[0183] Optionally, the device further comprises:

[0184] A second acquisition module is used to acquire at least one sample graph data, each of which is composed of at least two sample nodes, at least one sample edge and sample edge text on each sample edge, and the at least two sample nodes include at least one node corresponding to a sample object;

[0185] A third acquisition module is used to acquire a credit overdue label corresponding to the sample object, wherein the credit overdue label is used to indicate whether the sample object has a credit overdue behavior;

[0186] A first training module, configured to train a risk scoring model based on the at least one sample graph data and the credit overdue label corresponding to each sample object;

[0187] A generation module is used to generate the keyword score table based on the at least one sample graph data and the trained risk scoring model.

[0188] Optionally, the generating module is further used to:

[0189] Performing word segmentation processing on the sample edge text in the at least one sample graph data to obtain the at least one keyword;

[0190] Inputting the at least one keyword into the trained risk scoring model to obtain the risk score corresponding to each keyword;

[0191] The at least one keyword is associated with the risk score corresponding to each keyword to obtain the keyword score table.

[0192] Optionally, the first training module is further used to:

[0193] Perform edge text aggregation on the sample graph data based on the sample object to obtain a text string corresponding to at least one of the sample objects, wherein the text string is a set of sample edge texts on sample edges associated with the sample object in the sample graph data;

[0194] The risk scoring model is trained based on the text string corresponding to at least one of the sample objects and the credit overdue label corresponding to each of the sample objects.

[0195] Optionally, the first training module is further used to:

[0196] For each of the sample objects, inputting the text string corresponding to the sample object into the risk scoring model to obtain a predicted score output by the risk scoring model, wherein the predicted score represents the credit risk information indicated by the text string;

[0197] The risk scoring model is trained based on the loss value between the predicted score and the credit overdue label corresponding to the sample object.

[0198] Optionally, the device further comprises:

[0199] A fourth acquisition module is used to acquire a target overdue tag corresponding to the target object, wherein the target overdue tag is used to indicate whether the target object has an overdue behavior;

[0200] A data processing module, used for replacing the target edge text in the target graph data with the target edge weight to obtain fine-tuned target graph data;

[0201] The second training module is used to train a credit risk control model based on the fine-tuned target graph data and the target overdue label, wherein the credit risk control model is used to identify the credit overdue risk of the target object.

[0202] Optionally, the second training module is further used to:

[0203] Inputting the fine-tuned target graph data into the credit risk control model to obtain a credit overdue score output by the credit risk control model;

[0204] The credit risk control model is trained based on the loss between the credit overdue score and the target overdue label.

[0205] Optionally, the target graph data is generated from the address book data of the target object, the entity node is the address book object of the target object, and the target edge text is the remark label information of the address book object.

[0206] The exemplary embodiment of the present disclosure also provides an electronic device, comprising: at least one processor; and a memory connected to the at least one processor in communication. The memory stores a computer program that can be executed by the at least one processor, and the computer program is used to cause the electronic device to perform the method according to the embodiment of the present disclosure when executed by the at least one processor.

[0207] The exemplary embodiments of the present disclosure also provide a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor of a computer, is used to cause the computer to perform the method according to the embodiments of the present disclosure.

[0208] The exemplary embodiments of the present disclosure further provide a computer program product, including a computer program, wherein when the computer program is executed by a processor of a computer, the computer is used to enable the computer to perform the method according to the embodiments of the present disclosure.

[0209] refer to Fig.12 , a block diagram of an electronic device 1200 that can be used as a server or client of the present disclosure will now be described, which is an example of a hardware device that can be applied to various aspects of the present disclosure. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0210] like Fig.12 As shown, the electronic device 1200 includes a computing unit 1201, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1202 or a computer program loaded from a storage unit 1208 into a random access memory (RAM) 1203. In the RAM 1203, various programs and data required for the operation of the device 1200 can also be stored. The computing unit 1201, the ROM 1202, and the RAM 1203 are connected to each other via a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0211] Multiple components in the electronic device 1200 are connected to the I / O interface 1205, including: an input unit 1206, an output unit 1207, a storage unit 1208, and a communication unit 1209. The input unit 1206 can be any type of device that can input information to the electronic device 1200, and the input unit 1206 can receive input digital or character information, and generate key signal input related to user settings and / or function control of the electronic device. The output unit 1207 can be any type of device that can present information, and can include but is not limited to a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 1208 can include but is not limited to a disk, an optical disk. The communication unit 1209 allows the electronic device 1200 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and can include but is not limited to a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, such as a Bluetooth device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0212] The computing unit 1201 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1201 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1201 performs the various methods and processes described above. For example, in some embodiments, Figure 2 , Figure 4 , Figure 5 , Fig. 9 The method shown may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as the storage unit 1208. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 1200 via the ROM 1202 and / or the communication unit 1209. In some embodiments, the computing unit 1201 may be configured to execute the computer program by any other suitable means (e.g., by means of firmware). Figure 2 , Figure 4 , Figure 5 , Fig. 9 The method shown.

[0213] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0214] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0215] As used in this disclosure, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0216] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0217] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0218] A computer system may include clients and servers. Clients and servers are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship to each other.

Claims

1. A method for predicting credit overdue risk, characterized in that: The method comprises: Obtain target graph data constructed with the target object as a central node, wherein the target graph data includes at least one entity node having an association relationship with the target object, each of the entity nodes is connected to the central node corresponding to the target object through a target edge, and the target edge has a target edge text representing the association relationship; Determining a target edge weight corresponding to the at least one target edge text by matching at least one target edge text in the target graph data with a keyword score table, wherein the keyword score table includes at least one keyword and a risk score corresponding to each keyword, wherein the risk score represents credit risk information indicated by the keyword; Based on the target graph data and the target edge weight corresponding to the at least one target edge text, the credit overdue risk of the target object is predicted.

2. The method according to claim 1, characterized in that: The step of determining a target edge weight corresponding to the at least one target edge text by matching the at least one target edge text in the target graph data with a keyword score table includes: For each of the target edge texts in the target graph data, if there is a keyword matching the target edge text in the keyword score table, determine the target edge weight corresponding to the target edge text based on the risk score of the matched keyword; For each of the target edge texts in the target graph data, if there is no keyword matching the target edge text in the keyword score table, a preset score is determined as the target edge weight corresponding to the target edge text.

3. The method according to claim 2, characterized in that The step of determining the target edge weight corresponding to the target edge text based on the risk score of the matched keyword includes: If the target edge text matches a single keyword, the risk score corresponding to the keyword is determined as the target edge weight corresponding to the target edge text; If the target edge text matches at least two of the keywords, a target risk value is determined based on the risk scores corresponding to the at least two keywords, and the target risk value is determined as the target edge weight corresponding to the target edge text, and the target risk value is any one of the average, maximum or minimum values ​​of at least two risk scores.

4. The method according to claim 3, characterized in that The determining a target risk value based on the risk scores corresponding to at least two of the keywords includes: For each of the matched keywords, based on the proportion of the number of words occupied by each of the keywords in the target edge text, determine the weight corresponding to each of the keywords; The target risk value is determined based on the weighted weight and the risk score corresponding to each of the keywords.

5. The method according to any one of claims 1 to 4, characterized in that: The method further comprises: Obtaining a target overdue tag corresponding to the target object, where the target overdue tag is used to indicate whether the target object has an overdue behavior; Replacing the target edge text in the target graph data with the target edge weight to obtain fine-tuned target graph data; Based on the fine-tuned target graph data and the target overdue label, a credit risk control model is trained, and the credit risk control model is used to identify the credit overdue risk of the target object.

6. The method according to claim 5, characterized in that The step of training a credit risk control model based on the fine-tuned target graph data and the target overdue label includes: Inputting the fine-tuned target graph data into the credit risk control model to obtain a credit overdue score output by the credit risk control model; The credit risk control model is trained based on the loss between the credit overdue score and the target overdue label.

7. The method according to any one of claims 1 to 4, characterized in that: The target graph data is generated from the address book data of the target object, the entity node is the address book object of the target object, and the target edge text is the remark label information of the address book object.

8. A credit overdue risk prediction device, characterized in that: The device comprises: A first acquisition module is used to acquire target graph data constructed with the target object as a central node, wherein the target graph data includes at least one entity node having an association relationship with the target object, each of the entity nodes is connected to the central node corresponding to the target object through a target edge, and the target edge has a target edge text representing the association relationship; A matching module, configured to determine a target edge weight corresponding to the at least one target edge text by matching at least one target edge text in the target graph data with a keyword score table, wherein the keyword score table includes at least one keyword and a risk score corresponding to each keyword, wherein the risk score represents credit risk information indicated by the keyword; A prediction module is used to predict the credit overdue risk of the target object based on the target graph data and the target edge weight corresponding to the at least one target edge text.

9. An electronic device, comprising: processor; as well as Memory for storing programs, The program includes instructions, which, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-7.

Images