A method and device for predicting reasons for dissatisfaction based on a knowledge graph

Abstract

An embodiment of the present invention provides a method and device for predicting causes of dissatisfaction based on a knowledge graph. First, the unsatisfied business scenario label targeted by the target network user is determined among the multiple business scenario labels contained in the phenomenon layer of the knowledge graph model used for cause prediction; and based on the business behavior data of the target network user and multiple alternative associated feature labels, the user feature data corresponding to the target associated feature label related to the unsatisfied business scenario label is determined, and the user feature data is input into the above-mentioned knowledge graph model. Then, based on the edge weights between the cause concept labels contained in the cause concept layer in the knowledge graph model and the feature classification labels contained in the basic concept layer in the knowledge graph model, the unsatisfied cause label information of the target network user under the unsatisfied business scenario label is obtained, so that more accurate services can be provided to the target network user based on the unsatisfied cause label information, thereby improving the service usage experience of the target network user.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and device for predicting causes of dissatisfaction based on a knowledge graph. Background Art

[0002] With the advent of the big data era and the rapid development of artificial intelligence, machines have acquired the ability to learn. Machine learning means that machines can automatically summarize logic or rules from past data by selecting appropriate algorithms, and make predictions based on the results of this summary and new data. In the field of telecommunications operations, machine learning plays a very important role. Operators can predict user service satisfaction by collecting historical user behavior data, thereby enabling business personnel to provide users with better services.

[0003] Currently, when predicting user satisfaction with existing technologies, it is only possible to predict whether the user is satisfied, which leads to the problem that business personnel have less reference information when providing personalized assistance to users in the later stage. At the same time, business personnel still need to cater to user demands based on experience when performing later maintenance for users, which leads to the problem of low overall operational efficiency of the company. Summary of the Invention

[0004] The purpose of the embodiments of the present invention is to provide a method and device for predicting causes of dissatisfaction based on a knowledge graph, so as to solve the problem that business personnel have less reference information when providing personalized assistance to users in the later stage. At the same time, business personnel still need to cater to users' demands based on experience when performing later maintenance for users, which leads to the problem of low overall operational efficiency of the company.

[0005] In order to solve the above technical problems, the embodiment of the present invention is implemented as follows:

[0006] In a first aspect, an embodiment of the present invention provides a method for predicting causes of dissatisfaction based on a knowledge graph, comprising:

[0007] Determine the dissatisfied business scenario label targeted by the target network user from among multiple business scenario labels included in the phenomenon layer of the knowledge graph model used for cause prediction;

[0008] Determining user feature data corresponding to a target associated feature tag related to the unsatisfactory service scenario tag based on the service behavior data of the target network user and a plurality of candidate associated feature tags;

[0009] Utilizing the knowledge graph model and based on the user feature data, the reasons for dissatisfaction of the target network user are predicted to obtain the label information of the reasons for dissatisfaction of the target network user under the label of the unsatisfied business scenario; wherein, the label information of the reasons for dissatisfaction is determined based on the edge weight between the reason concept label contained in the reason concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model.

[0010] In a second aspect, an embodiment of the present invention provides a device for predicting causes of dissatisfaction based on a knowledge graph, comprising:

[0011] A scenario label determination module is used to determine the dissatisfied business scenario label targeted by the target network user from among the multiple business scenario labels included in the phenomenon layer in the knowledge graph model used for cause prediction;

[0012] A user feature determination module is configured to determine user feature data corresponding to the target associated feature tag related to the unsatisfactory service scenario tag based on the service behavior data of the target network user and a plurality of candidate associated feature tags;

[0013] A dissatisfaction reason prediction module is used to use the knowledge graph model and based on the user feature data to predict the dissatisfaction reason of the target network user, and obtain the dissatisfaction reason label information of the target network user under the unsatisfactory business scenario label; wherein, the dissatisfaction reason label information is determined based on the edge weight between the cause concept label contained in the cause concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model.

[0014] In a third aspect, an embodiment of the present invention provides a computer device comprising a processor, a communication interface, a memory, and a communication bus; wherein the processor, the communication interface, and the memory communicate with each other through a bus; the memory is used to store computer programs; and the processor is used to execute the programs stored in the memory to implement the steps of the method for predicting causes of dissatisfaction based on a knowledge graph as described in the first aspect.

[0015] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the method for predicting causes of dissatisfaction based on a knowledge graph as described in the first aspect are implemented.

[0016] The dissatisfaction cause prediction method and device based on knowledge graph in the embodiment of the present invention include: determining the unsatisfied business scenario label targeted by the target network user among multiple business scenario labels contained in the phenomenon layer of the knowledge graph model used for cause prediction; and determining the user feature data corresponding to the target associated feature label related to the unsatisfied business scenario label targeted by the target network user based on the business behavior data of the above-mentioned target network user and multiple alternative associated feature labels; using the knowledge graph model and based on the above-mentioned user feature data, predicting the dissatisfaction cause of the target network user to obtain the unsatisfied cause label information of the target network user under the unsatisfied business scenario label; wherein the unsatisfied cause label information is determined based on the edge weight between the cause concept label contained in the cause concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model. In an embodiment of the present invention, first, among the multiple business scenario labels contained in the phenomenon layer in the knowledge graph model used for cause prediction, the unsatisfied business scenario label targeted by the target network user is determined; and based on the business behavior data of the target network user and multiple alternative associated feature labels, the user feature data corresponding to the target associated feature label related to the unsatisfied business scenario label is determined, and the user feature data is input into the above-mentioned knowledge graph model, and based on the edge weights between the cause concept labels contained in the cause concept layer in the knowledge graph model and the feature classification labels contained in the basic concept layer in the knowledge graph model, the unsatisfied cause label information of the target network user under the unsatisfied business scenario label is obtained, so that more accurate services can be provided to the target network user based on the unsatisfied cause label information, thereby improving the service usage experience of the target network user. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A schematic diagram of a first flow chart of a method for predicting causes of dissatisfaction based on a knowledge graph provided in an embodiment of the present invention;

[0019] Figure 2 A second flow chart of the method for predicting causes of dissatisfaction based on a knowledge graph provided in an embodiment of the present invention;

[0020] Figure 3 A schematic diagram of a specific implementation process of the knowledge graph-based prediction of dissatisfaction reasons provided in an embodiment of the present invention;

[0021] Figure 4 A third flow chart of the method for predicting causes of dissatisfaction based on a knowledge graph provided in an embodiment of the present invention;

[0022] Figure 5a A schematic diagram of a specific implementation process of the knowledge graph model construction process in the dissatisfaction cause prediction method based on the knowledge graph provided in an embodiment of the present invention;

[0023] Figure 5b A schematic diagram of the specific structure of the knowledge graph model constructed in the method for predicting causes of dissatisfaction based on the knowledge graph provided in an embodiment of the present invention;

[0024] Figure 6 A schematic diagram of the module composition of a device for predicting causes of dissatisfaction based on a knowledge graph provided in an embodiment of the present invention;

[0025] Figure 7 A schematic diagram of the module composition of the knowledge graph model construction process in the dissatisfaction cause prediction device based on the knowledge graph provided by an embodiment of the present invention;

[0026] Figure 8 A schematic diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0028] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0029] An embodiment of the present invention provides a method and device for predicting causes of dissatisfaction based on a knowledge graph. First, the unsatisfied business scenario label targeted by the target network user is determined among the multiple business scenario labels contained in the phenomenon layer of the knowledge graph model used for cause prediction; and based on the business behavior data of the target network user and multiple alternative associated feature labels, the user feature data corresponding to the target associated feature label related to the unsatisfied business scenario label is determined, and the user feature data is input into the above-mentioned knowledge graph model. Based on the edge weights between the cause concept labels contained in the cause concept layer in the knowledge graph model and the feature classification labels contained in the basic concept layer in the knowledge graph model, the unsatisfied cause label information of the target network user under the unsatisfied business scenario label is obtained, so that more accurate services can be provided to the target network user based on the unsatisfied cause label information, thereby improving the service usage experience of the target network user.

[0030] Figure 1 A first flow chart of a method for predicting causes of dissatisfaction based on a knowledge graph provided in an embodiment of the present invention is provided. Figure 1 The method in can be executed by the business server, such as Figure 1 As shown, the dissatisfaction cause prediction method based on knowledge graph includes at least the following steps:

[0031] S101, determine the dissatisfied business scenario label for the target network user among the multiple business scenario labels contained in the phenomenon layer in the knowledge graph model for cause prediction; wherein, the dissatisfied business scenario label belongs to at least one business scenario label contained in the phenomenon layer in the knowledge graph model for cause prediction; the knowledge graph model includes: an atomic layer, a basic concept layer, a cause concept layer, and a phenomenon layer deployed from bottom to top; in the specific implementation, it is first necessary to determine all business scenarios under the target business service according to actual business needs, and use the above business scenarios as business scenario labels of the phenomenon layer in the knowledge graph model. For example, the business scenario labels may include: "tariff", "home broadband", "wireless" and other business scenario labels. Before using the knowledge graph to predict the reasons for dissatisfaction of the target network user, it is necessary to determine the dissatisfied business scenario label for the target network user among the multiple business scenario labels contained in the phenomenon layer of the knowledge graph model used for cause prediction, that is, to determine under which business scenario label the target network user is dissatisfied. For example, among the business scenario labels such as "tariff", "home broadband", and "wireless", it is determined that the target network user is dissatisfied under the business scenario label "tariff". Then, the dissatisfied business scenario label for the target network user can be the business scenario label "tariff".

[0032] S102, based on the business behavior data of the target network user and multiple alternative associated feature tags, determine the user feature data corresponding to the target associated feature tag related to the unsatisfactory business scenario tag targeted by the target network user; in the specific implementation, first use a preset machine learning algorithm to filter out the target associated feature tag related to the unsatisfactory business scenario tag targeted by the target network user from multiple alternative associated feature tags; further, after filtering out the target associated feature tag related to the unsatisfactory business scenario tag targeted by the target network user, determine the user feature data corresponding to the target associated feature tag related to the unsatisfactory business scenario tag targeted by the target network user based on the business behavior data of the target network user. Among them, the above-mentioned alternative associated feature labels include: the associated feature labels contained in the atomic layer in the knowledge graph model; the above-mentioned preset machine learning algorithm can be a random forest algorithm, or other machine learning algorithms can be selected; specifically, in the process of constructing the knowledge graph model, the machine learning model can be used to screen out multiple alternative associated feature labels related to various business scenario labels related to the target business service from a large number of associated feature labels; correspondingly, when using the trained knowledge graph model to predict the cause of dissatisfaction, the machine learning model can also be used to screen out target associated feature labels related to the dissatisfied business scenario labels of the target network users from multiple alternative associated feature labels; for example, the business scenario labels related to the target business service include: "tariff", "home broadband", "wireless" three business scenario labels, and the machine learning algorithm selects random Taking the forest algorithm as an example, in the process of constructing the knowledge graph model, the random forest algorithm is used to screen out the associated feature labels related to the three business scenario labels of "tariff", "home broadband" and "wireless" (that is, the alternative associated feature labels used for predicting the reasons for dissatisfaction), and the atomic layer in the knowledge graph model is constructed based on the multiple screened associated feature labels; correspondingly, when using the trained knowledge graph model to predict the reasons for dissatisfaction, after determining the unsatisfied business scenario label of the target network user, the random forest algorithm is used to screen out the target associated feature labels related to the unsatisfied business scenario label of the target network user. For example, if the unsatisfied business scenario label of the target network user is the business scenario label of "tariff", then the alternative associated feature labels related to the business scenario label of "tariff" are screened out as the target associated feature labels.

[0033] S103, using a pre-trained knowledge graph model and based on the user feature data corresponding to the above-mentioned target-associated feature labels, predict the reasons for dissatisfaction of the target network users, and obtain the label information of the reasons for dissatisfaction of the target network users under the label of dissatisfied business scenarios; wherein, the above-mentioned pre-trained knowledge graph model is the above-mentioned knowledge graph model for cause prediction; wherein, the label information of the reasons for dissatisfaction is determined based on the edge weights between the reason concept labels contained in the reason concept layer in the knowledge graph model and the feature classification labels contained in the basic concept layer in the knowledge graph model. Specifically, the user feature data corresponding to the target-associated feature tag related to the unsatisfied service scenario tag of the target network user obtained in the above step S102 is first used as the input of the atomic layer in the pre-trained knowledge graph model, and then the user feature data corresponding to each target-associated feature tag is mapped to the feature classification tag contained in the basic concept layer in the knowledge graph model, and the unsatisfied service scenario tag of the determined target network user is mapped to the cause concept tag contained in the cause concept layer in the knowledge graph model, and then based on the edge weight between the cause concept tag contained in the cause concept layer and the feature classification tag contained in the basic concept layer, the cause of dissatisfaction of the target network user in the unsatisfied service scenario is determined, thereby realizing accurate automatic prediction of the cause of dissatisfaction of the target network user, so as to guide the business personnel to optimize the business services that the user is dissatisfied with based on the predicted unsatisfied cause tag information of the target network user.

[0034] In an embodiment of the present invention, first, among the multiple business scenario labels contained in the phenomenon layer in the knowledge graph model used for cause prediction, the unsatisfied business scenario label targeted by the target network user is determined; and based on the business behavior data of the target network user and multiple alternative associated feature labels, the user feature data corresponding to the target associated feature label related to the unsatisfied business scenario label is determined, and the user feature data is input into the above-mentioned knowledge graph model, and then based on the edge weights between the cause concept labels contained in the cause concept layer in the knowledge graph model and the feature classification labels contained in the basic concept layer in the knowledge graph model, the unsatisfied cause label information of the target network user under the unsatisfied business scenario label is obtained, so that more accurate services can be provided to the target network user based on the unsatisfied cause label information, thereby improving the service usage experience of the target network user.

[0035] Furthermore, after determining the user characteristic data corresponding to the target associated characteristic tag related to the target user's dissatisfied service scenario tag based on the target user's service behavior data and multiple candidate associated characteristic tags, the pre-trained knowledge graph model can be used to predict the target user's dissatisfaction reasons based on the user characteristic data corresponding to each target associated characteristic tag. The process of predicting the reasons for user dissatisfaction is as follows: Figure 2 As shown, in the above step S103, the dissatisfaction reasons of the target network users are predicted using the pre-trained knowledge graph model and based on the user feature data corresponding to the above target associated feature label, and the dissatisfaction reason label information of the target network users under the unsatisfied service scenario label is obtained, which specifically includes:

[0036] S1031, taking the correspondence between the identification information of the target-related feature tag and the user feature data corresponding to the target-related feature tag as the input of the atomic layer in the knowledge graph model, and using the above atomic layer to input the correspondence into the basic concept layer in the knowledge graph model; specifically, Figure 3 This is a schematic diagram of a specific implementation process of the prediction of dissatisfaction reasons based on the knowledge graph provided by an embodiment of the present invention. Figure 3 As shown, the pre-trained knowledge graph model includes: an atomic layer, a basic concept layer, a cause concept layer, and a phenomenon layer deployed from the bottom up; assuming that the target network user A's dissatisfied service scenario label is "tariff" dissatisfaction, the random forest algorithm is first used to filter out target-related feature labels related to "tariff" dissatisfaction, wherein the target-related feature labels may include: out-of-package traffic last month, out-of-package voice last month; assuming that based on the target network user's business behavior data, the target network user's out-of-package traffic last month was 500M, and the out-of-package voice last month was 500 minutes; correspondingly, the above "500M" and "500 minutes" are the user feature data of the target network user; the above target association feature labels are combined The correspondence between the identification information of the feature tag and the user feature data is used as the input of the atomic layer, that is, "last month's out-of-package traffic - 500M" and "last month's out-of-package voice - 500 minutes" are used as the input of the atomic layer, and the atomic layer is used to input the two pairs of correspondences of "last month's out-of-package traffic (identification information of the target-associated feature tag) - 500M (user feature data)" and "last month's out-of-package voice (identification information of the target-associated feature tag) - 500 minutes (user feature data)" into the basic concept layer; among them, the identification information of the target-associated feature tag is the feature column name of the target-associated feature tag, and the above feature column name consists of English fields and has the same meaning as the target-associated feature tag.

[0037] S1032, using the basic concept layer in the knowledge graph model, based on the user feature data corresponding to the target-related feature label related to the dissatisfied service scenario label of the target network user, determine the target feature classification label corresponding to the target-related feature label, and input the identification information of the target feature classification label into the cause concept layer in the knowledge graph model; specifically, the target-related feature label and the target feature classification label are in a one-to-one mapping relationship. Figure 3As shown in , the target feature classification label corresponding to the target-associated feature label "out-of-package traffic last month" is "more out-of-package traffic last month" or "less out-of-package traffic last month", where the concepts of "more" or "less" are obtained by simply mapping the numerical values of the user feature data corresponding to the target-associated feature label. For example, if it is pre-defined that the feature classification label corresponding to out-of-package traffic greater than 300M is more out-of-package traffic, and the feature classification label corresponding to out-of-package voice greater than 300 minutes is more out-of-package voice, then the feature classification label corresponding to "out-of-package traffic 500M last month" is "more out-of-package traffic last month", and the feature classification label corresponding to "out-of-package voice 500 minutes last month" is "more out-of-package voice last month", and then "more out-of-package traffic last month" and "more out-of-package voice last month" are input as target feature classification labels into the cause concept layer, so that the target cause concept label can be obtained based on the target feature classification label mapping in the cause concept layer.

[0038] S1033, using the above-mentioned cause concept layer to determine the target cause concept label and the corresponding association weight corresponding to the target feature classification label, and sorting a preset number of target cause concept labels with the highest association weights in descending order, and determining them as the dissatisfaction cause label information of the target network user under the dissatisfaction business scenario label; wherein the corresponding association weight is the edge weight of the connecting edge between the target feature classification label and the target cause concept label. Specifically, with respect to the process of determining the association weight between the target feature classification label and the target cause concept label, in the above-mentioned step S1033, the target cause concept label and the corresponding association weight corresponding to the target feature classification label are determined using the above-mentioned cause concept layer, specifically including: using the cause concept layer based on the mapping relationship between the preset business scenario label and the cause concept label, determining the target cause concept label having a mapping relationship with the target feature classification label under the unsatisfactory business scenario label; specifically, based on the mapping relationship between the preset business scenario label and the cause concept label, the target cause concept label corresponding to the unsatisfactory business scenario label targeted by the target network user can be determined, and in the above-mentioned step S1032, the target feature classification label under the unsatisfactory business scenario label targeted by the target network user can be determined. Therefore, the target cause concept label corresponding to the unsatisfactory business scenario label targeted by the target network user is the target cause concept label having a mapping relationship with the target feature classification label under the unsatisfactory business scenario label; for each target cause concept label, the edge weight of the connecting edge between the target cause concept label and any target feature classification label is determined as the association weight between the target cause concept label and the target feature classification label.

[0039] Specifically, the cause concept layer includes cause concept labels corresponding to all unsatisfied business scenario labels. When predicting the cause of dissatisfaction for a target network user, the target feature classification label of the target network user is input into the cause concept layer in the knowledge graph model. The cause concept layer is used to determine the association weight between the target feature classification label and the target cause concept label corresponding to the target feature classification label, and a preset number of target cause concept labels with the highest association weights are sorted in descending order as the unsatisfied cause label information of the target network user under the unsatisfied business scenario label; wherein the target cause concept label is determined by the unsatisfied business scenario label in the phenomenon layer, and the mapping relationship between the unsatisfied business scenario label and the cause concept label can be determined based on historical business-related data, and the cause concept label corresponding to the target unsatisfied business scenario label is determined as the target cause concept label, that is, when predicting the cause of dissatisfaction for a target network user, first determine the unsatisfied business scenario of the target network user (that is, the target unsatisfied business scenario label), and then, based on the unsatisfied business scenario label, determine the target cause concept label. The service scenario filters out the associated feature labels related to the unsatisfactory service scenario, maps the associated feature labels to the target feature classification labels, and determines the target cause concept labels corresponding to the unsatisfactory service scenario. At this time, the target cause concept labels with a mapping relationship with the target feature classification labels in the unsatisfactory service scenario can be determined. Further, the connection relationship (i.e., the connection edge) between each target cause concept label and any target feature classification label is determined, and the edge weight of the connection edge is determined as the association weight between the target cause concept label and the target feature classification label. The edge weight is determined based on sample data during the construction of the knowledge graph model. The edge weight represents the degree of association between the target cause concept label and the target feature classification label. The larger the weight value of the edge weight, the greater the degree of association, indicating that the target cause concept label is more likely to be the cause of dissatisfaction of the target network user. Therefore, the association weights are sorted in descending order, and a preset number of target cause concept labels with the highest sorting order are determined as the cause of dissatisfaction of the target network user under the unsatisfactory service scenario label.

[0040] For example, in the above Figure 3 As shown in , it is assumed that the dissatisfied service scenario of a target network user is dissatisfied with the “price”, and the target cause concept labels corresponding to the “price” dissatisfaction in the cause concept layer include: “voice over package”, “traffic over package”, “overlay fee ratio”, and “out-of-package ratio”; among them, the target cause concept label can also be refined into multiple secondary cause concepts. Therefore, the first-level cause concept can be determined based on the edge weight first, and then the target secondary cause concept can be selected from all the secondary cause concepts under the first-level cause concept as the dissatisfaction cause label information. For example, in the above Figure 3As shown in , "voice overset" can be called the target cause concept label, "voice overset 1 time" and "voice overset multiple times" can be called secondary cause concept labels, wherein the secondary cause concept label is to increase the readability of the business personnel and refine the target cause concept label. In the embodiment of the present invention, the target cause concept label can also be called the primary cause concept label; in the above Figure 3 As shown in , it is assumed that the target feature classification labels of user A in the basic concept layer are: "more traffic outside the package last month", "more voice outside the package last month"; among them, it is assumed that the association weight between the target cause concept label "voice exceeds the package" and the target feature classification label "more traffic outside the package last month" is 0.1, and the association weight between the target cause concept label "more voice outside the package last month" is 0.8; the association weight between the target cause concept label "traffic exceeds the package" and the target feature classification label "more traffic outside the package last month" is 0.8, and the association weight between the target cause concept label "more voice outside the package last month" is 0.01; the association weight between the target cause concept label "overlay fee ratio" and the target feature classification label "more traffic outside the package last month" is 0. The correlation weight between the target cause concept label "proportion of out-of-package" and the target feature classification label "more traffic outside the package last month" is 0.2, and the correlation weight between the target cause concept label "more voice outside the package last month" is 0.06; arranging the above correlation weights in descending order, it can be concluded that the correlation weights between the two target cause concept labels "voice out of package" and "traffic out of package" and the target feature classification label are larger, while the other correlation weights are smaller or even negligible. Therefore, the two target cause concept labels "traffic out of package" and "traffic out of package" are determined as the dissatisfaction reason label information of user A under the label of "tariff" dissatisfaction.

[0041] Furthermore, in order to improve the user service experience, the predicted target user's dissatisfaction reason label information can be used to guide the optimization of the service for which the user is dissatisfied. Based on this, in the above step S103, the target user's dissatisfaction reason is predicted using a pre-trained knowledge graph model and based on the user feature data corresponding to the above target-related feature label. After obtaining the target user's dissatisfaction reason label information under the unsatisfied service scenario label, the following is also included:

[0042] According to the determined dissatisfaction reason label information, the service information to be recommended for the target network user is determined, and the service information to be recommended is sent to the client corresponding to the target network user; wherein, the service information to be recommended can be used to make up for the target network user's dissatisfaction with the business service, thereby improving the user's service experience, that is, applying the predicted dissatisfaction reasons to a larger analysis and operation scenario, providing a basis for business personnel to optimize the business services that users are dissatisfied with, such as Figure 4 As shown, based on the cause concept labels in the cause concept layer, we can provide users with services such as "product recommendation" and "activity recommendation", or apply the cause concept labels to larger operational scenarios such as "customer retention", and can also be extended to the company's macro data analysis, such as "revenue growth" and other scenarios to predict the company's benefits from a macro perspective; for example, in the above Figure 4 As shown, the target cause concept label corresponding to the feature classification label "number of complaints about cost questioning" in the basic concept layer is "cost questioning complaints". Among them, the categories of the feature classification label "number of complaints about cost questioning" are mapped to "days" and "costs", making the hierarchy of the basic concept layer clearer; and the obtained target cause concept label "cost questioning complaints" is applied to the "customer retention" operation scenario to further enhance the user's service experience.

[0043] Furthermore, before predicting the reasons for satisfaction of target users, it is first necessary to design a schema for the architecture of the knowledge graph model based on the prediction of satisfaction reasons. Schema design is the conceptual model and logical basis of the knowledge graph. It uses the rule definition of the relationship between entities in the architecture to constrain the data input into the knowledge graph. Therefore, schema design for internal operation and maintenance scenarios is the basis for knowledge graph construction.

[0044] Specifically, the above-mentioned knowledge graph model is designed into a four-layer architecture. Through this four-layer architecture, the knowledge graph can be easily applied to other scenarios. As long as there is data corresponding to the scenario, new phenomena and cause concepts can be integrated, thereby reducing the workload of expanding the scenario. In addition, the above-mentioned knowledge graph maps the phenomenon layer, atomic layer, and hierarchical theory for the operation scenario into the basic concept layer and the cause concept layer; among them, the entities in the knowledge graph are explanatory semantic concepts of maintenance, recommendation, and service strategies for the operation scenario, so as to improve the accuracy of operation recommendations and provide understandable and explainable operation strategies for "customer needs"; based on this, Figure 5a As shown, in step S101, before determining the dissatisfied service scenario label targeted by the target network user among the multiple service scenario labels included in the phenomenon layer in the knowledge graph model for cause prediction, the following is further included:

[0045] S501, multiple business scenario tags related to the target business service are used as the first entity to construct the phenomenon layer; specifically, the construction of the knowledge graph first depends on the satisfaction survey plan for network users. The knowledge graph is used as an application to predict the reasons for the dissatisfaction of target network users. The survey results of the satisfaction survey plan shown in the following table (1) are the basis for the links between the entities of the knowledge graph; the satisfaction survey score is 0-10 points, where 0-6 points are dissatisfied and 7-10 points are satisfied.

[0046]

[0047] Table (1)

[0048] Based on the satisfaction survey results in Table (1) above, multiple unsatisfactory business scenario labels corresponding to the target business service can be determined, and the multiple business scenario labels can be used as the first entity to construct the phenomenon layer; for example, the score of voice call satisfaction in Table (1) above is 3 points, and the score of mobile phone Internet satisfaction is 2 points, which means that the customer satisfaction with both voice calls and mobile phone Internet is unsatisfactory. Furthermore, it can be determined that in the business scenario of "price", the user experience is unsatisfactory, and based on this business scenario, the various entities in the knowledge graph related to the business scenario are screened out as the basis for linking between entities in the knowledge graph model.

[0049] S502, using a machine learning algorithm to filter out associated feature tags related to the multiple business scenario tags, and using the filtered multiple associated feature tags as the second entity to construct an atomic layer;

[0050] In the construction of the atomic layer in the knowledge graph model, each entity node is a sample user who has joined the network. With the sample user as the center, the relationship between the user and each business element in the knowledge graph is constructed. In addition, more than 100 satisfaction-related tags in the big data tag library are included as entities to link the target network users to form the form of "customer number A-unsatisfied tag A-user number B-unsatisfied tag B"; among them, the data source of the atomic layer can be regarded as the semantic network structure reorganization of the big data tag library. Based on the "link prediction + collaborative filtering" method, the knowledge graph can be constructed and reasoned based solely on the atomic layer. Based on the unified analysis of the satisfaction big data tags, customized entity extraction is performed on each tag, such as "number of voice over-packages in the past three months", "N-1 month out-of-package traffic charges", "N-1 month whether unsatisfied Furthermore, entity concepts can be extracted through artificial intelligence algorithms. For example, the Bi-LSTM+Attention algorithm uses a bidirectional LSTM model to encode input data to obtain a representation of the entire input data information. The addition of an attention mechanism and an RNN-like model during decoding can better capture the required data information and the relationships between the data. The Text-CNN algorithm uses a CNN to encode the input model, and a dilated-CNN can be used to obtain a longer representation of the data information. The RNN-like model is still used during decoding to generate the required data description. The Self-Attention algorithm uses a self-attention mechanism to more effectively obtain a representation of the input data vector information. The Bert algorithm uses a large-scale pre-trained network to extract entity concepts. Specifically, based on the above-mentioned business scenario labels, a machine learning algorithm is used to screen out associated feature labels that correspond to each business scenario label. The machine learning algorithm can be an algorithm for finding feature importance, such as the Random Forest Classifier algorithm, to filter out associated feature labels that are irrelevant to all business scenario labels and use associated feature labels related to the business scenario labels as the second entity to construct the atomic layer. For example, when the business scenario label is "tariff dissatisfaction", the associated feature label related to "tariff dissatisfaction" can be as shown in the following table (2), where the identification information of the associated feature label is the feature column name of the user feature data.

[0051]

[0052]

[0053] Table (2)

[0054] Furthermore, after filtering out all the associated feature tags based on the dissatisfied business scenario tags, the user feature data of the sample users can be input into the corresponding associated feature tags. For example, as shown in Table (3) below, after filtering out the associated feature tags related to the dissatisfied “price” business scenario shown in Table (2) above, the user feature data of the sample users dissatisfied with the “price” can be input into the corresponding associated feature tags.

[0055]

[0056] Table (3)

[0057] S503: The feature classification labels corresponding to the associated feature labels are used as the third entity to construct a basic concept layer. Specifically, there is a one-to-one mapping between the associated feature labels and the feature classification labels. For example, to construct the cause concept layer of the "price" satisfaction prediction knowledge graph, it is necessary to import the dissatisfaction reasons defined based on the evaluation results of users' dissatisfaction with their tariffs and use them as cause concept labels to explain the entities in the atomic layer. Each sample user predicted to be dissatisfied has a corresponding dissatisfaction cause concept label.

[0058] S504, using the cause concept tag corresponding to the business scenario tag as the fourth entity to construct a cause concept layer;

[0059] S505, using a machine learning method to iteratively train and update the model parameters of the preset logistic regression model based on the pre-selected model training sample set, to obtain updated model parameters, until the loss function corresponding to the logistic regression model converges;

[0060] S506, determining the edge weight of the connection edge between the feature classification label and the cause concept label based on the updated model parameters when the loss function converges;

[0061] S507: Based on the edge weights between the feature classification labels and the cause concept labels, and the bottom-up deployment of the atomic layer, basic concept layer, cause concept layer, and phenomenon layer, a knowledge graph model for cause prediction is constructed. Specifically, Figure 5b A schematic diagram of the specific structure of the knowledge graph model constructed in the method for predicting causes of dissatisfaction based on knowledge graphs provided in an embodiment of the present invention. Figure 5bAs shown in , the knowledge graph model is divided into atomic layer, basic concept layer, cause concept layer, and phenomenon layer from bottom to top; among them, "business scenario label 1 to business scenario label k" is the first entity in the phenomenon layer; "associated feature label 1 to associated feature label n" is the second entity in the atomic layer; "feature classification label 1 to feature classification label z" is the third entity in the basic concept layer; "cause concept label 1 to cause concept label m" is the fourth entity in the cause concept layer. Among them, by iteratively training and updating the model parameters of the preset logistic regression model, the edge weight of the connection edge between the feature classification label in the basic concept layer and the cause concept label in the cause concept layer can be obtained. Specifically, the model training sample set includes multiple model training samples, wherein each model training sample includes: the correspondence between the feature classification label and the cause concept label, that is, the correspondence is used to characterize the multiple feature classification labels corresponding to at least one cause concept label that causes the target network user to be dissatisfied with a certain service scenario. For example, as shown in Table (4) below, the target network user with the mobile phone number "18222057113" is dissatisfied with the "home broadband" service scenario. For the cause concept label "abnormal disconnection for many consecutive days", the corresponding feature classification labels are "many home broadband flash disconnections in N-1 month" and "large wireless black spot coverage area in N-2 month". Therefore Based on real sample data, the edge weight of the connection edge between the feature classification label and the cause concept label can be obtained by iteratively training the logistic regression model. The edge weight can represent the degree of association between the feature classification label and the cause concept label, that is, the larger the edge weight, the greater the probability that the cause of user dissatisfaction under the feature classification label includes the cause concept label; wherein, the entities of the cause concept layer include all potential causes of dissatisfaction in the unsatisfactory business scenario. For example, the method for obtaining the cause of dissatisfaction with "feature" is clustering + graph neural network + decision tree, and the sample set is constructed based on the self-test data and quarterly evaluation data of the operating company. The imported pre-selected model training sample set includes user number, associated feature label corresponding to user feature data, feature classification label corresponding to associated feature label, and cause concept label corresponding to feature classification label as shown in the following table (4):

[0062]

[0063]

[0064] Table (4)

[0065] Furthermore, the model parameters of the preset logistic regression model are iteratively trained and updated based on the pre-selected model training sample set using a machine learning method to obtain updated model parameters, specifically including:

[0066] For each model training sample in the model training sample set, the model training sample is binary-converted to obtain the corresponding one-hot encoding of the model training sample, and the one-hot encoding is used as the input of the preset logistic regression model; based on the one-hot encoding corresponding to each model training sample, the loss function corresponding to the logistic regression model is updated; the loss function is minimized to obtain the updated model parameters. In specific implementations, the gradient descent method can be used to minimize the loss function to obtain the updated model parameters.

[0067] Specifically, since the feature classification labels corresponding to the basic concept layer are text, the feature classification labels need to be converted into one-hot encoding (one hot encoding). One-hot encoding converts the feature classification labels into binary encoding and then uses them as features for model training. Compared with directly using the feature classification labels as features, one-hot encoding can avoid the error of assuming that the higher the category value in the feature classification labels, the better. For example, suppose the basic concept layer has four feature classification labels, namely "high traffic last month", "low traffic last month", "male gender", and "female gender". If the feature classification labels corresponding to a target user are "high traffic last month" and "male gender", then the one-hot encoding of the target user is [1,0,1,0], where the 1 corresponding to the first bit in the one-hot encoding represents that the target user has the feature classification label "high traffic last month", and the 0 corresponding to the second bit represents that the target user does not have the feature classification label "low traffic last month". Similarly, the feature classification labels corresponding to the target user can be converted into one-hot encoding for model training of the knowledge graph model. In actual prediction, target users often have multiple reasons for dissatisfaction, corresponding to multiple reason concept labels. Therefore, this is a multi-label classification task. The number of category labels for multi-label classification is uncertain. Some samples may have only one category label, and some samples may have multiple category labels. For example, some target users may have only one reason for dissatisfaction, and some target users may have multiple reasons for dissatisfaction. Therefore, commonly used classifiers are not applicable. It is necessary to adopt a multi-classifier method, treat each label in the multi-label as a single label, implement a binary classification algorithm for each label, and then integrate them to obtain the final result. For example, assuming that there are only three reason concept labels A, B, and C in the reason concept layer, and the reason labels corresponding to a sample user are A and C. First, it is necessary to convert them into one-hot encoding [1,0,1], and then train three binary classification models for the reason concept labels A, B, and C respectively, and then integrate them to output multi-label results.

[0068] In this knowledge graph, the logistic regression model is used as the binary classification model. Other classification models, such as support vector machines and tree models, are not very explanatory. Even though the tree model can obtain feature importance, it cannot be used as edge weights. Therefore, the tree model must still be used for prediction, and prediction cannot be made directly based on the graph structure. That is, each cause concept label corresponds to a logistic regression model. The model parameters of the logistic regression model corresponding to each cause concept label are iteratively trained to obtain a trained logistic regression model.

[0069] Because the logistic regression model is relatively clear and edge weights can be obtained from the model, the logistic regression model is highly interpretable. The model parameters naturally have statistical meaning and can be used as edge weights. Later, when making predictions, predictions can be made directly based on the graph structure. Therefore, the logistic regression classifier was ultimately selected. The essence of logistic regression is to assume that the data follows a logistic distribution and then use maximum likelihood to estimate the parameters. The logistic distribution is a continuous probability distribution with a distribution function of:

[0070]

[0071] Among them, g(z) represents the probability of predicting a sample as a positive example, and the hypothetical function form of logistic regression is as follows:

[0072]

[0073] Among them, h θ (x) represents the probability that the target cause concept label used for prediction is a positive example, x represents the unique hot coding combination corresponding to the feature classification label of the sample user and the cause concept label, θ represents the model parameter to be obtained, and the unique hot coding is input into the hypothesis function corresponding to the logistic regression model to obtain the predicted probability of whether the cause concept label is a cause of dissatisfaction. The logistic regression model can be a linear logistic regression model or a nonlinear logistic regression model. Taking the linear logistic regression model as an example, Y n =θ T =θ 0n X0+θ 1n X1+...+θ zn X z .

[0074] Among them, z represents the number of feature classification labels contained in a sample user, n represents the number of cause concept labels contained in a sample user, X0 to X z Represents all the feature classification labels corresponding to a sample user, θ 0n to θ zn represents the edge weight between the nth target reason concept label and each feature classification label of a sample user, for example, θ 1nis the edge weight between the feature classification label represented by X1 and the nth cause concept label.

[0075] Furthermore, based on the prediction probability corresponding to each model training sample and the maximum likelihood method, the loss function of the logistic regression function can be written as:

[0076]

[0077] Among them, y (i) is the true label, h θ (x (i) ) is the predicted label; the loss function J(θ) represents the degree of proximity between the true label and the predicted label. Specifically, when the model parameter θ of the loss function is minimized, the true label is closest to the predicted label. m represents the number of sample users in the model training set. Then, using gradient descent, we can find the model parameters that minimize the loss function. These model parameters are the optimal solution for the logistic regression function.

[0078] Furthermore, the above-mentioned loss function is minimized to obtain the updated model parameters, which specifically includes:

[0079] If the first gradient of the loss function of the previous round of model training is consistent with the sign of the second gradient of the loss function of the current round of model training, the model parameters of the logistic regression model are updated based on the first parameter update formula to obtain the updated model parameters of the current round; if the first gradient of the loss function of the previous round of model training is inconsistent with the sign of the second gradient of the loss function of the current round of model training, the model parameters of the logistic regression model are updated based on the second parameter update formula to obtain the updated model parameters of the current round; wherein, the first parameter update formula is The second parameter update formula is: θ: represents the model parameters after the current round of update, θ represents the model parameters after the previous round of update, Represents the second gradient of the loss function of this round of model training, and α represents the learning rate of the loss function. Among them, the first gradient of the loss function of the previous round of model training is the gradient value obtained by substituting the model parameters updated in the previous round into the loss function and deriving the loss function; the second gradient of the loss function of this round of model training is the gradient value obtained by substituting the model parameters updated in the previous round into the loss function and deriving the loss function; in specific implementation, since the loss function is a continuous convex function, there will only be one global optimal point and no local optimal point, so when selecting an optimizer, optimizers based on momentum or adaptive learning rate do not need to be used, because the main purpose of those optimizers is to prevent the loss function from falling into the local optimum. A simple small batch gradient descent is sufficient. This solution adds adjustments to the traditional small batch gradient descent and proposes a new optimizer, which can speed up the algorithm to solve the parameters. The specific approach is as follows: Small batch gradient descent:

[0080] At each step of the algorithm, we randomly extract a small batch of samples from the training set with m samples. The number of samples is m. We first calculate the loss function of m samples and then solve the gradient. The gradient formula is as follows:

[0081]

[0082] Then set the learning rate α and continuously update it according to the following first parameter update formula until the point of the minimum loss function is reached.

[0083]

[0084] However, when the amount of sample user data is large, this approximation process may take too long, and generally a smaller learning rate is used, which will make the process longer. Therefore, the present invention adopts the following approximation means:

[0085] First, the method of calculating the loss function and the gradient remains unchanged, but at each step, the gradient calculated in the previous step is saved. At this time, the gradient calculated in the previous step and the gradient calculated this time have the following two situations:

[0086] 1) When the sign of the first gradient of the loss function of the previous round of model training is consistent with the sign of the second gradient of the loss function of the current round of model training, the model parameters are updated according to the above first parameter update formula steps;

[0087] 2) When the signs of the first gradient of the loss function of the previous round of model training are inconsistent with the second gradient of the loss function of the current round of model training, the model parameters are updated according to the following second parameter update formula;

[0088]

[0089] The principle is that when the signs of the first gradient of the loss function from the previous model training cycle and the second gradient of the loss function from the current model training cycle differ, the lowest point in the loss function is somewhere in the middle. Jumping directly to the middle in this situation may result in being too close to the lowest point in the loss function, and this often leads to oscillation during gradient descent. Using the above-mentioned second parameter calculation formula for updating can alleviate this oscillation. Using this optimizer, a higher learning rate can be used initially to more quickly reach the optimal model parameter point. The learning rate can then be gradually reduced until the lowest point is reached. This results in faster convergence than simple mini-batch gradient descent with a fixed learning rate. After model training, the model parameters are extracted, representing the relationships and edge weights from the basic concept layer to the causal concept layer, for use in inference during prediction. Furthermore, the parameters of logistic regression follow a normal distribution, making it unlikely that one parameter is too large or too small. Regularization can also be added to logistic regression to prevent overfitting.

[0090] After the model training is completed, the reasons for the dissatisfaction of the target users can be inferred. The output of the knowledge graph model during inference can be shown in the following table (5), where the confidence score is the probability of outputting the reasons for the dissatisfaction of the target users, and a CSV file of the top 2 million customer groups with the highest confidence score is output. The big data micro-marketing platform is used to maintain the stock operation. The output columns include the number of the target users, whether they are dissatisfied, reason concept 1, reason concept 2, and confidence score. The example is as follows:

[0091]

[0092]

[0093] Table (5)

[0094] Furthermore, in step S503, the feature classification label corresponding to the associated feature label is used as the third entity to construct a basic concept layer, which specifically includes:

[0095] If the user feature data corresponding to the associated feature tag is discrete data, the classification reference point of the feature classification tag corresponding to the associated feature tag is determined based on at least two categories of the user feature data; specifically, the discrete data can be the gender information, occupation information and other classified information of the target network user. For example, if the associated feature tag is a gender tag, the user feature data can be divided into two categories: "male" or "female". Further, "male" and "female" can be used as feature classification labels to construct a basic concept layer; for example, if the associated feature tag is an occupation label, when the occupation is "nurse", the user feature data can be divided into two categories: "nurse" or "non-nurse". Further, "nurse" and "non-nurse" can be used as feature classification labels to construct a basic concept layer.

[0096] For discrete data, there is no need to save it because discrete data itself represents a mapping. With the above mapping relationship, when using the knowledge graph model for prediction, the corresponding feature classification label can be inferred from the user feature data corresponding to the associated feature label from the bottom up.

[0097] Furthermore, if the user feature data corresponding to the associated feature label is continuous data, a pre-selected quantile determination sample set is obtained; and, based on the quantile determination sample set, the maximum eigenvalue of the user feature data at the head of the first preset proportion and the minimum eigenvalue of the user feature data at the tail of the second preset proportion are determined in the order of the eigenvalues of the user feature data from small to large; and the maximum eigenvalue is determined as the first quantile and the minimum eigenvalue is determined as the second quantile; wherein, the first quantile is smaller than the second quantile, if the eigenvalue corresponding to the user feature data is smaller than the first quantile, the feature classification label is determined to be a first label for characterizing any one of low, small, and few; if the eigenvalue corresponding to the user feature data is greater than the second quantile, the feature classification label is determined to be a second label for characterizing any one of high, large, and many; specifically, for a certain associated feature label whose user feature data is continuous data, the sample historical data corresponding to the associated feature label is sorted from small to large, and the quantiles of the above historical data are obtained by calculation, for example, for "last month's traffic "This associated feature label is sorted from small to large by the sample historical data of the associated feature label, and through calculation, 90% is used as a first preset ratio to determine the maximum eigenvalue of the user feature data at the head of the first preset ratio; and 10% is used as a second preset ratio to determine the minimum eigenvalue of the user feature data at the tail of the second preset ratio for reasoning during prediction. When the input user feature data is greater than the 90% percentile, that is, greater than the maximum eigenvalue, the feature data is mapped to any one of high, large, and many; when the input user feature data is less than the 10% percentile, the feature data is mapped to any one of low, small, and few. Among them, continuous data must be converted into binary data before it can be converted into links between atomic-level entities and basic concept-level entities. If continuous data within 10-90% is also involved in the mapping, the links between atomic-level entities and basic concept-level entities will be too dense, resulting in excessive computational complexity during prediction and reasoning. Therefore, only the cases of less than 10% and greater than 90% are considered to ensure that the links are relatively sparse, so as to achieve greater confidence discrimination during prediction and reasoning.

[0098] The feature classification labels corresponding to the associated feature labels and the quantile reference points corresponding to the feature classification labels are used as the third entity to construct the basic concept layer. Specifically, by setting the reference points, the discrete data and continuous data corresponding to the above-mentioned associated feature labels are mapped to obtain feature classification labels, and the above-mentioned feature classification labels are used as the third entity to construct the basic concept layer. Regarding the construction process of the knowledge graph model, the atomic layer, basic concept layer, cause concept layer and phenomenon layer are deployed from bottom to top, and the optimal model parameters are obtained based on the iterative training of the model parameters of the logistic regression model. The model parameters are used as the edge weights of the connecting edges between the feature classification labels and the cause concept labels, and then a knowledge graph model for cause prediction is constructed. In this way, the knowledge graph model can be used to predict the causes of dissatisfaction and output the causes of user dissatisfaction, thereby improving the accuracy and efficiency of the prediction of the causes of user dissatisfaction, thereby providing guidance for subsequent service optimization.

[0099] An embodiment of the present invention provides a method for predicting causes of dissatisfaction based on a knowledge graph, the method comprising: determining a dissatisfaction cause label for a target network user among a plurality of business scenario labels contained in a phenomenon layer in a knowledge graph model used for cause prediction; and determining user characteristic data corresponding to a target associated characteristic label related to the dissatisfaction cause label for the target network user based on the business behavior data of the target network user and a plurality of alternative associated characteristic labels; predicting the cause of dissatisfaction of the target network user using a knowledge graph model and based on the user characteristic data, obtaining dissatisfaction cause label information of the target network user under the dissatisfaction cause label; wherein the dissatisfaction cause label information is determined based on the edge weight between the cause concept label contained in the cause concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model. In an embodiment of the present invention, first, among the multiple business scenario labels contained in the phenomenon layer in the knowledge graph model used for cause prediction, the unsatisfied business scenario label targeted by the target network user is determined; and based on the business behavior data of the target network user and multiple alternative associated feature labels, the user feature data corresponding to the target associated feature label related to the unsatisfied business scenario label is determined, and the user feature data is input into the above-mentioned knowledge graph model, and then based on the edge weights between the cause concept labels contained in the cause concept layer in the knowledge graph model and the feature classification labels contained in the basic concept layer in the knowledge graph model, the unsatisfied cause label information of the target network user under the unsatisfied business scenario label is obtained, so that more accurate services can be provided to the target network user based on the unsatisfied cause label information, thereby improving the service usage experience of the target network user.

[0100] Corresponding to the dissatisfaction cause prediction method based on knowledge graph provided in the above embodiment, based on the same technical concept, the embodiment of the present invention also provides a dissatisfaction cause prediction device based on knowledge graph. Figure 6 A schematic diagram of the module composition of a dissatisfaction cause prediction device based on a knowledge graph provided in an embodiment of the present invention, wherein the dissatisfaction cause prediction device based on a knowledge graph is used to perform Figures 1 to 5b The dissatisfaction cause prediction method based on knowledge graph is shown in Figure 6 As shown, the dissatisfaction cause prediction device based on the knowledge graph includes:

[0101] The scenario label determination module 602 is configured to determine the dissatisfied business scenario label targeted by the target network user from among the multiple business scenario labels included in the phenomenon layer of the knowledge graph model used for cause prediction;

[0102] A user feature determination module 604 is configured to determine user feature data corresponding to the target associated feature tag related to the unsatisfactory service scenario tag based on the service behavior data of the target network user and a plurality of candidate associated feature tags;

[0103] The dissatisfaction reason prediction module 606 is used to use the knowledge graph model and based on the user feature data to predict the dissatisfaction reason of the target network user, and obtain the dissatisfaction reason label information of the target network user under the unsatisfactory business scenario label; wherein, the dissatisfaction reason label information is determined based on the edge weight between the cause concept label contained in the cause concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model.

[0104] Alternatively, as Figure 7 As shown, in one embodiment of the present invention, the dissatisfaction cause prediction device based on the knowledge graph further includes:

[0105] A phenomenon layer construction module 702 is configured to construct a phenomenon layer using a plurality of business scenario tags related to a target business service as a first entity;

[0106] An atomic layer construction module 704 is configured to use a machine learning algorithm to filter associated feature tags related to the multiple business scenario tags, and use the filtered associated feature tags as second entities to construct an atomic layer;

[0107] A basic concept layer construction module 706 is configured to construct a basic concept layer using the feature classification label corresponding to the associated feature label as a third entity;

[0108] A cause concept layer construction module 708 is configured to construct a cause concept layer using the cause concept tag corresponding to the business scenario tag as a fourth entity;

[0109] A model parameter updating module 710 is configured to iteratively train and update the model parameters of a preset logistic regression model based on a pre-selected model training sample set using a machine learning method, to obtain updated model parameters, until the loss function corresponding to the logistic regression model converges;

[0110] An edge weight determination module 712 is configured to determine an edge weight of a connection edge between the feature classification label and the cause concept label based on the updated model parameters when the loss function converges;

[0111] The knowledge graph model construction module 714 is used to construct a knowledge graph model for cause prediction based on the edge weights and the atomic layer, the basic concept layer, the cause concept layer, and the phenomenon layer deployed from bottom to top.

[0112] The device for predicting causes of user dissatisfaction based on knowledge graph in an embodiment of the present invention first determines the unsatisfied business scenario label targeted by the target network user among the multiple business scenario labels contained in the phenomenon layer in the knowledge graph model used for cause prediction; and based on the business behavior data of the target network user and multiple alternative associated feature labels, determines the user feature data corresponding to the target associated feature label related to the unsatisfied business scenario label, inputs the user feature data into the above-mentioned knowledge graph model, and then obtains the unsatisfied cause label information of the target network user under the unsatisfied business scenario label based on the edge weight between the cause concept label contained in the cause concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model, so as to provide the target network user with more accurate services based on the unsatisfied cause label information and improve the service usage experience of the target network user.

[0113] The device for predicting causes of user dissatisfaction based on knowledge graph provided in an embodiment of the present invention can implement each process in the embodiment corresponding to the above-mentioned method for predicting causes of user dissatisfaction based on knowledge graph. To avoid repetition, they will not be described here.

[0114] It should be noted that the device for predicting causes of user dissatisfaction based on knowledge graph provided in an embodiment of the present invention and the method for predicting causes of user dissatisfaction based on knowledge graph provided in an embodiment of the present invention are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned method for predicting causes of user dissatisfaction based on knowledge graph, and the repetitive parts will not be repeated.

[0115] Corresponding to the method for predicting the causes of user dissatisfaction based on the knowledge graph provided in the above embodiment, based on the same technical concept, an embodiment of the present invention further provides a computer device, which is used to execute the above method for predicting the causes of user dissatisfaction based on the knowledge graph. Figure 8 A schematic diagram of the structure of a computer device for implementing various embodiments of the present invention is shown in FIG. Figure 8 As shown. The computer device may have relatively large differences due to different configurations or performances, and may include one or more processors 801 and memory 802, and the memory 802 may store one or more storage applications or data. Among them, the memory 802 can be a temporary storage or a persistent storage. The application stored in the memory 802 may include one or more modules (not shown in the figure), and each module may include a series of computer executable instructions in the computer device. Furthermore, the processor 801 can be configured to communicate with the memory 802 to execute a series of computer executable instructions in the memory 802 on the computer device. The computer device may also include one or more power supplies 803, one or more wired or wireless network interfaces 804, one or more input and output interfaces 805, and one or more keyboards 806.

[0116] Specifically in this embodiment, the computer device includes a processor, a communication interface, a memory, and a communication bus; wherein the processor, the communication interface, and the memory communicate with each other via the bus; the memory is used to store computer programs; and the processor is used to execute the programs stored in the memory to implement the following method steps:

[0117] Determine the dissatisfied business scenario label targeted by the target network user from among multiple business scenario labels included in the phenomenon layer of the knowledge graph model used for cause prediction;

[0118] Determining user feature data corresponding to a target associated feature tag related to the unsatisfactory service scenario tag based on the service behavior data of the target network user and a plurality of candidate associated feature tags;

[0119] Utilizing the knowledge graph model and based on the user feature data, the reasons for dissatisfaction of the target network user are predicted to obtain the label information of the reasons for dissatisfaction of the target network user under the label of the unsatisfied business scenario; wherein, the label information of the reasons for dissatisfaction is determined based on the edge weight between the reason concept label contained in the reason concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model.

[0120] The computer device in the embodiment of the present invention first determines the unsatisfied business scenario label targeted by the target network user among the multiple business scenario labels contained in the phenomenon layer in the knowledge graph model used for cause prediction; and based on the business behavior data of the target network user and multiple alternative associated feature labels, determines the user feature data corresponding to the target associated feature label related to the unsatisfied business scenario label, inputs the user feature data into the above-mentioned knowledge graph model, and then obtains the unsatisfied cause label information of the target network user under the unsatisfied business scenario label based on the edge weight between the cause concept label contained in the cause concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model, so as to provide the target network user with more accurate services based on the unsatisfied cause label information and improve the service usage experience of the target network user.

[0121] The computer device provided in the embodiment of the present invention can implement each process in the embodiment corresponding to the above-mentioned method for predicting causes of dissatisfaction based on knowledge graph. To avoid repetition, it will not be described here.

[0122] It should be noted that the computer device provided by the embodiment of the present invention and the method for predicting causes of dissatisfaction based on knowledge graph provided by the embodiment of the present invention are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned method for predicting causes of dissatisfaction based on knowledge graph, and the repetitive parts will not be repeated.

[0123] Corresponding to the dissatisfaction cause prediction method based on knowledge graph provided in the above embodiment, based on the same technical concept, an embodiment of the present invention further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the following method steps:

[0124] Determine the dissatisfied business scenario label targeted by the target network user from among multiple business scenario labels included in the phenomenon layer of the knowledge graph model used for cause prediction;

[0125] Determining user feature data corresponding to a target associated feature tag related to the unsatisfactory service scenario tag based on the service behavior data of the target network user and a plurality of candidate associated feature tags;

[0126] Utilizing the knowledge graph model and based on the user feature data, the reasons for dissatisfaction of the target network user are predicted to obtain the label information of the reasons for dissatisfaction of the target network user under the label of the unsatisfied business scenario; wherein, the label information of the reasons for dissatisfaction is determined based on the edge weight between the reason concept label contained in the reason concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model.

[0127] The computer-readable storage medium in the embodiment of the present invention first determines the unsatisfied business scenario label targeted by the target network user among the multiple business scenario labels contained in the phenomenon layer in the knowledge graph model used for cause prediction; and based on the business behavior data of the target network user and multiple alternative associated feature labels, determines the user feature data corresponding to the target associated feature label related to the unsatisfied business scenario label, inputs the user feature data into the above-mentioned knowledge graph model, and then obtains the unsatisfied cause label information of the target network user under the unsatisfied business scenario label based on the edge weight between the cause concept label contained in the cause concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model, so as to provide the target network user with more accurate services based on the unsatisfied cause label information and improve the service usage experience of the target network user.

[0128] The computer-readable storage medium provided in an embodiment of the present invention can implement each process in the embodiment corresponding to the above-mentioned method for predicting causes of dissatisfaction based on knowledge graph. To avoid repetition, they will not be described here.

[0129] It should be noted that the computer-readable storage medium provided by the embodiment of the present invention and the method for predicting causes of dissatisfaction based on the knowledge graph provided by the embodiment of the present invention are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned method for predicting causes of dissatisfaction based on the knowledge graph, and the repetitive parts will not be repeated.

[0130] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0132] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0134] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0135] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0136] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0137] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0138] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] The above are merely embodiments of the present invention and are not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A method for predicting causes of dissatisfaction based on knowledge graph, characterized in that: The method comprises: Determine the dissatisfied business scenario label targeted by the target network user from among multiple business scenario labels included in the phenomenon layer of the knowledge graph model used for cause prediction; Determining user feature data corresponding to a target associated feature tag related to the unsatisfactory service scenario tag based on the service behavior data of the target network user and a plurality of candidate associated feature tags; Utilizing the knowledge graph model and based on the user feature data, the reasons for dissatisfaction of the target network user are predicted to obtain the label information of the reasons for dissatisfaction of the target network user under the label of the unsatisfied business scenario; wherein, the label information of the reasons for dissatisfaction is determined based on the edge weight between the reason concept label contained in the reason concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model.

2. The method according to claim 1, characterized in that The method of using the knowledge graph model and based on the user feature data to predict the dissatisfaction reasons of the target network user to obtain the dissatisfaction reason label information of the target network user under the dissatisfaction service scenario label includes: The corresponding relationship between the identification information of the target-related feature tag and the user feature data is used as the input of the atomic layer in the knowledge graph model, and the corresponding relationship is input into the basic concept layer in the knowledge graph model using the atomic layer; Determine a target feature classification label corresponding to the target association feature label using the basic concept layer based on the user feature data, and input identification information of the target feature classification label into the cause concept layer in the knowledge graph model; The cause concept layer is used to determine the target cause concept label and the corresponding association weight corresponding to the target feature classification label, and a preset number of the target cause concept labels with the highest association weights are sorted in descending order to determine the dissatisfaction cause label information of the target network user under the dissatisfaction business scenario label.

3. The method according to claim 2, characterized in that The determining of the target cause concept label and the corresponding association weight corresponding to the target feature classification label by using the cause concept layer includes: Determine a target cause concept label having a mapping relationship with the target feature classification label under the unsatisfactory business scenario label by utilizing the cause concept layer based on a mapping relationship between preset business scenario labels and cause concept labels; For each target cause concept label, the edge weight of the connecting edge between the target cause concept label and any target feature classification label is determined as the association weight between the target cause concept label and the target feature classification label.

4. The method according to claim 1, wherein Among the multiple business scenario labels included in the phenomenon layer of the knowledge graph model used for cause prediction, before determining the dissatisfied business scenario label targeted by the target network user, the following is also included: The phenomenon layer is constructed by taking multiple business scenario tags related to the target business service as the first entity; Using a machine learning algorithm to filter associated feature tags related to the multiple business scenario tags, and using the filtered associated feature tags as a second entity to construct an atomic layer; The feature classification label corresponding to the associated feature label is used as the third entity to construct a basic concept layer; The cause concept label corresponding to the business scenario label is used as the fourth entity to construct a cause concept layer; Using a machine learning method to iteratively train and update the model parameters of a preset logistic regression model based on a pre-selected model training sample set, to obtain updated model parameters, until the loss function corresponding to the logistic regression model converges; Determining an edge weight of a connection edge between the feature classification label and the cause concept label based on the updated model parameters when the loss function converges; Based on the edge weights and the atomic layer, the basic concept layer, the cause concept layer, and the phenomenon layer deployed from bottom to top, a knowledge graph model for cause prediction is constructed.

5. The method according to claim 4, characterized in that The method utilizes a machine learning method to iteratively train and update the model parameters of a preset logistic regression model based on a pre-selected model training sample set to obtain updated model parameters, including: For each model training sample in the model training sample set, perform binary conversion on the model training sample to obtain a one-hot encoding corresponding to the model training sample, and use the one-hot encoding as input to a preset logistic regression model; Based on the one-hot encoding corresponding to each of the model training samples, updating the loss function corresponding to the logistic regression model; The loss function is minimized to obtain updated model parameters.

6. The method according to claim 5, characterized in that The minimizing the loss function to obtain updated model parameters includes: If the first gradient of the loss function of the previous round of model training is consistent with the second gradient of the loss function of the current round of model training, the model parameters of the logistic regression model are updated based on the first parameter update formula to obtain the updated model parameters of the current round; If the first gradient of the loss function of the previous round of model training is inconsistent with the second gradient of the loss function of the current round of model training, the model parameters of the logistic regression model are updated based on the second parameter update formula to obtain the updated model parameters of the current round; Among them, the first parameter update formula is: , the second parameter update formula is , Represents the model parameters after this round of update, Represents the model parameters after the last round of update, Represents the second gradient of the loss function of this round of model training, Represents the loss function learning rate.

7. The method according to claim 4, characterized in that The step of constructing a basic concept layer by using the feature classification label corresponding to the associated feature label as a third entity includes: If the user feature data corresponding to the associated feature label is discrete data, determining a quantile reference point of the feature classification label corresponding to the associated feature label based on at least two categories of the user feature data; If the user feature data corresponding to the associated feature tag is continuous data, obtaining a pre-selected quantile determination sample set; and, Determine a sample set based on the quantiles, and determine the maximum eigenvalue of the user feature data at the head of a first preset proportion and the minimum eigenvalue of the user feature data at the tail of a second preset proportion in the order of the eigenvalues of the user feature data from small to large; and The maximum eigenvalue is determined as a first quantile and the minimum eigenvalue is determined as a second quantile; wherein the first quantile is smaller than the second quantile, if the eigenvalue corresponding to the user feature data is smaller than the first quantile, then the feature classification label is determined to be a first label for representing any one of low, small, and few; if the eigenvalue corresponding to the user feature data is greater than the second quantile, then the feature classification label is determined to be a second label for representing any one of high, large, and many; The feature classification label corresponding to the associated feature label and the quantile reference point corresponding to the feature classification label are used as a third entity to construct a basic concept layer.

8. A device for predicting causes of dissatisfaction based on knowledge graph, characterized in that: include: A scenario label determination module is used to determine the dissatisfied business scenario label targeted by the target network user from among the multiple business scenario labels included in the phenomenon layer in the knowledge graph model used for cause prediction; A user feature determination module is configured to determine user feature data corresponding to a target associated feature tag related to the unsatisfactory service scenario tag based on the service behavior data of the target network user and a plurality of candidate associated feature tags; A dissatisfaction reason prediction module is used to use the knowledge graph model and based on the user feature data to predict the dissatisfaction reason of the target network user, and obtain the dissatisfaction reason label information of the target network user under the unsatisfactory business scenario label; wherein, the dissatisfaction reason label information is determined based on the edge weight between the cause concept label contained in the cause concept layer in the knowledge graph model and the feature classification label contained in the basic concept layer in the knowledge graph model.

9. A computer device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus; wherein the processor, the communication interface and the memory communicate with each other through the bus; the memory is used to store computer programs; the processor is used to execute the programs stored in the memory to implement the method for predicting causes of dissatisfaction based on knowledge graphs as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for predicting causes of dissatisfaction based on a knowledge graph as described in any one of claims 1 to 7.

Images