Training method and device of classification model, state recognition method and device

Abstract

The embodiments of the present disclosure disclose a classification model training method and device, a state recognition method and device. The classification model training method comprises: obtaining training sample data, wherein the training sample data comprises graph features, non-graph features and state information of a subject; determining a first prediction result of the subject through the non-graph features, and comparing the first prediction result with the state information to determine a first error; determining a second prediction result of the subject according to at least the graph features, and comparing the second prediction result with the state information to determine a second error; determining a first label of the subject based on the relative sizes of the first error and the second error, wherein the first label indicates whether to use or not to use the graph features for state recognition, and is used for training a first classification model, so that the first classification model obtained by training can recognize whether the subject needs to use the graph features for state recognition, and is beneficial to targeted state recognition of different subjects, so as to improve the recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of information technology, and in particular, to a classification model training method and device, and a state recognition method and device. BACKGROUND

[0002] In many scenarios, some states of an object are not obvious, and the states of the object need to be accurately recognized so as to make decisions accordingly. For example, in the scenarios of contract review and enterprise investment, the risk of an enterprise needs to be reviewed, and if the risk is regarded as a state of the enterprise, the state of the enterprise needs to be accurately recognized to avoid property loss. The existing state recognition method usually recognizes the state based on the attribute features of the object itself, and has the problem of poor accuracy. SUMMARY

[0003] To solve the problems in the related art, the embodiments of the present disclosure provide a classification model training method and device, and a state recognition method and device.

[0004] In a first aspect, a classification model training method is provided in the embodiments of the present disclosure.

[0005] Specifically, the classification model training method comprises:

[0006] obtaining training sample data, the training sample data comprising graph features, non-graph features and state information of a subject;

[0007] determining a first prediction result of the subject by the non-graph features, and comparing the first prediction result with the state information to determine a first error;

[0008] determining a second prediction result of the subject according to at least the graph features, and comparing the second prediction result with the state information to determine a second error;

[0009] determining a first label of the subject based on the relative size of the first error and the second error, the first label indicating whether to use or not to use the graph features for state recognition, and being used for training a first classification model.

[0010] With reference to the first aspect, in an implementation form of the first aspect, the method further comprises:

[0011] determining a training sample subset of the training sample data in which the graph features are used for state recognition based on the first label;

[0012] for the subject in the training sample subset, determining a second label of the subject based on the comparison of the first prediction result and the second prediction result, the second label indicating that the effect direction of the graph features is a positive direction or a negative direction, and being used for training a second classification model.

[0013] In a second aspect, the embodiments of the present disclosure provide a state recognition method.

[0014] Specifically, the state recognition method comprises:

[0015] obtaining a graph feature and a non-graph feature of a subject;

[0016] determining a first classification result using or not using the graph feature based on a first classification model;

[0017] in a case where the first classification result indicates that the graph feature is used for state recognition, determining the state of the subject according to at least the graph feature of the subject.

[0018] With reference to the second aspect, in a first implementation of the second aspect, the method further comprises:

[0019] in a case where the first classification result indicates that the graph feature is not used for state recognition, determining the state of the subject according to the non-graph feature of the subject.

[0020] With reference to the second aspect, in a second implementation of the second aspect, in the case where the first classification result indicates that the graph feature is used for state recognition, the state of the subject is determined according to at least the graph feature of the subject, which comprises:

[0021] processing the graph feature and the non-graph feature based on a second classification model to determine a second classification result indicating whether the graph feature effect is in a positive direction or a negative direction;

[0022] processing the graph feature and the non-graph feature based on a state recognition model corresponding to the second classification result to determine the state of the subject.

[0023] With reference to the second implementation of the second aspect, in a third implementation of the second aspect, the state recognition model comprises a first state recognition model for processing the graph feature effect in the positive direction and a second state recognition model for processing the graph feature effect in the negative direction, and the processing of the graph feature and the non-graph feature based on the state recognition model corresponding to the second classification result to determine the state of the subject comprises:

[0024] in a case where the second classification result indicates that the graph feature effect is in the positive direction, determining the state of the subject based on the first state recognition model;

[0025] in a case where the second classification result indicates that the graph feature effect is in the negative direction, determining the state of the subject based on the second state recognition model.

[0026] With reference to the third implementation of the second aspect, in a fourth implementation of the second aspect, the method further includes:

[0027] in a case where the first classification result indicates that the graph feature is not used for state recognition, processing non-graph features of the subject using a third state recognition model to determine a state of the subject,

[0028] wherein, in a case where the subject is a plurality, the method further includes:

[0029] aggregating outputs of the first state recognition model, the second state recognition model and the third state recognition model to obtain state recognition results of the plurality of subjects.

[0030] In a third aspect, the embodiments of the present disclosure provide a business risk identification method.

[0031] Specifically, the business risk identification method includes:

[0032] obtaining graph features and non-graph features of a to-be-identified business;

[0033] determining a first classification result of using or not using the graph features based on a first classification model;

[0034] in a case where the first classification result indicates that the graph feature is used for risk identification, identifying a risk of the business according to at least the graph features of the business.

[0035] In a fourth aspect, the embodiments of the present disclosure provide a training device of a classification model.

[0036] Specifically, the training device of the classification model includes:

[0037] a first obtaining module configured to obtain training sample data, the training sample data including graph features, non-graph features and state information of a subject;

[0038] a first determining module configured to determine a first prediction result of the subject through the non-graph features, and compare the first prediction result with the state information to determine a first error;

[0039] a second determining module configured to determine a second prediction result of the subject according to at least the graph features, and compare the second prediction result with the state information to determine a second error;

[0040] a third determining module configured to determine a first label of the subject based on a relative size of the first error and the second error, the first label indicating that the graph feature is used or not used for state recognition, and being used for training a first classification model.

[0041] With reference to the fourth aspect, in an implementation form of the fourth aspect, the method further includes:

[0042] The fourth determining module is configured to determine, based on the first label, a training sample subset in the training sample data for which state recognition is performed using the graph feature;

[0043] The fifth determining module is configured to determine, for a subject in the training sample subset, a second label of the subject based on a comparison of the first prediction result and the second prediction result, the second label indicating that the graph feature effect direction is a positive direction or a negative direction, and being used for training a second classification model.

[0044] With reference to the fifth aspect, in an implementation form of the fifth aspect, the apparatus further includes:

[0045] Specifically, the state recognition apparatus includes:

[0046] The second obtaining module is configured to obtain a graph feature and a non-graph feature of a subject;

[0047] The sixth determining module is configured to determine, based on the first classification model, a first classification result of whether to use the graph feature.

[0048] The seventh determining module is configured to determine, in a case where the first classification result indicates that the graph feature is used for state recognition, a state of the subject based at least on the graph feature of the subject.

[0049] With reference to the fifth aspect, in a first implementation form of the fifth aspect, the apparatus further includes:

[0050] The eighth determining module is configured to determine, in a case where the first classification result indicates that the graph feature is not used for state recognition, the state of the subject based on the non-graph feature of the subject.

[0051] With reference to the fifth aspect, in a second implementation form of the fifth aspect, the determining, in the case where the first classification result indicates that the graph feature is used for state recognition, the state of the subject based at least on the graph feature of the subject includes:

[0052] processing the graph feature and the non-graph feature based on a second classification model to determine a second classification result of whether the graph feature effect direction is a positive direction or a negative direction;

[0053] processing the graph feature and the non-graph feature based on a state recognition model corresponding to the second classification result to determine the state of the subject.

[0054] With reference to the second implementation manner of the fifth aspect, in a third implementation manner of the fifth aspect, the state recognition model includes a first state recognition model for processing the graph feature effect in a positive direction and a second state recognition model for processing the graph feature effect in a negative direction, and the state of the subject is determined by processing the graph feature and the non-graph feature based on the state recognition model corresponding to the second classification result, including:

[0055] In a case where the second classification result indicates that the graph feature effect direction is a positive direction, the state of the subject is determined based on the first state recognition model.

[0056] In a case where the second classification result indicates that the graph feature effect direction is a negative direction, the state of the subject is determined based on the second state recognition model.

[0057] With reference to the third implementation manner of the fifth aspect, in a fourth implementation manner of the fifth aspect, the apparatus further includes:

[0058] An eighth determining module configured to, in a case where the first classification result indicates that the graph feature is not used for state recognition, determine the state of the subject by processing the non-graph feature of the subject based on a third state recognition model.

[0059] An aggregating module configured to aggregate outputs of the first state recognition model, the second state recognition model and the third state recognition model to obtain state recognition results of a plurality of subjects.

[0060] In a sixth aspect, an electronic device is provided, including a memory and a processor, where the memory is configured to store one or more computer instructions, and the one or more computer instructions are executed by the processor to implement the method in the first aspect to the third aspect and various implementation manners thereof.

[0061] In a seventh aspect, a computer readable storage medium is provided, and the computer readable storage medium stores computer instructions, and the computer instructions are executed by a processor to implement the method in the first aspect to the third aspect and various implementation manners thereof.

[0062] According to the technical scheme provided by the embodiment of the present disclosure, the training sample data is obtained, the training sample data includes the graph feature, the non-graph feature and the state information of the subject; the first prediction result of the subject is determined according to the non-graph feature, and compared with the state information to determine the first error; the second prediction result of the subject is determined at least according to the graph feature, and compared with the state information to determine the second error; the first label of the subject is determined based on the relative size of the first error and the second error, the first label indicates whether to use the graph feature for state identification, and is used for training the first classification model, so that the first classification model obtained by training can identify whether the subject needs to use the graph feature for state identification, which is beneficial to the targeted state identification of different subjects, so as to improve the identification accuracy.

[0063] According to the technical scheme provided by the embodiment of the present disclosure, the training sample data in which the graph feature is used for state identification is determined based on the first label; for the subject in the training sample subset, the second label of the subject is determined based on the comparison of the first prediction result and the second prediction result, the second label indicates that the effect direction of the graph feature is a positive direction or a negative direction, and is used for training the second classification model, so that the second classification model obtained by training can identify the effect direction of the graph feature on the subject, which is beneficial to the targeted state identification of different subjects, so as to improve the identification accuracy.

[0064] According to the technical scheme provided by the embodiment of the present disclosure, the graph feature and the non-graph feature of the subject are obtained; the first classification result of using or not using the graph feature is determined based on the first classification model; in the case that the first classification result indicates that the graph feature is used for state identification, the state of the subject is determined at least according to the graph feature of the subject, so that the targeted state identification of different subjects can improve the accuracy of state identification.

[0065] According to the technical scheme provided by the embodiment of the present disclosure, in the case that the first classification result indicates that the graph feature is not used for state identification, the state of the subject is determined according to the non-graph feature of the subject, so that different state identification means is adopted according to different first classification results, which can improve the accuracy of state identification.

[0066] According to the technical scheme provided by the embodiment of the present disclosure, the graph feature and the non-graph feature are processed based on the second classification model to determine the second classification result that the effect direction of the graph feature is a positive direction or a negative direction; the graph feature and the non-graph feature are processed based on the state identification model corresponding to the second classification result to identify and determine the state of the subject, so that different state identification models are adopted according to the second classification result, which can further improve the accuracy of state identification.

[0067] According to the technical scheme provided by the embodiment of the present disclosure, by using the first state recognition model for processing the graph feature effect in the positive direction and the second state recognition model for processing the graph feature effect in the negative direction, the state of the subject is determined by processing the graph feature and the non-graph feature based on the state recognition model corresponding to the second classification result, that is, in the case that the second classification result indicates that the graph feature effect direction is positive, the state of the subject is determined based on the first state recognition model, and in the case that the second classification result indicates that the graph feature effect direction is negative, the state of the subject is determined based on the second state recognition model, thereby achieving targeted state recognition for different subjects, and the accuracy of state recognition can be improved.

[0068] According to the technical scheme provided by the embodiment of the present disclosure, in the case that the first classification result indicates that the state recognition is not performed using the graph feature, the non-graph feature of the subject is processed using the third state recognition model to determine the state of the subject, and in the case that the subject is multiple, the method further includes: aggregating the outputs of the first state recognition model, the second state recognition model and the third state recognition model to obtain the state recognition results of the multiple subjects, thereby achieving targeted state recognition for different subjects, and the accuracy of state recognition can be improved.

[0069] According to the technical scheme provided by the embodiment of the present disclosure, by obtaining the graph feature and the non-graph feature of the enterprise to be recognized, determining the first classification result of using or not using the graph feature based on the first classification model, and in the case that the first classification result indicates that the risk of the enterprise is recognized using the graph feature, the risk of the enterprise is recognized at least according to the graph feature of the enterprise, thereby achieving targeted state recognition for different subjects, and the accuracy of state recognition can be improved.

[0070] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0071] Other features, objects and advantages of the present disclosure will become more apparent from the following detailed description of the non-limiting embodiments combined with the attached drawings. In the drawings:

[0072] Figures 1A-1D A schematic diagram of a subject relationship graph according to an embodiment of the present disclosure is shown;

[0073] Figure 2 A flowchart of a state recognition method according to an embodiment of the present disclosure is shown;

[0074] Figure 3 A flowchart of a state recognition method according to another embodiment of the present disclosure is shown;

[0075] Figure 4 FIG. 1 shows a flowchart of a method for training a classification model according to an embodiment of the present disclosure;

[0076] Figure 5 FIG. 1 shows a flowchart of a method for training a classification model according to an embodiment of the present disclosure;

[0077] Figure 6 FIG. 1 shows a flowchart of a method for training a classification model according to an embodiment of the present disclosure;

[0078] Figure 7 FIG. 1 shows a flowchart of a method for training a classification model according to an embodiment of the present disclosure;

[0079] Figure 8 FIG. 1 shows a flowchart of a method for training a classification model according to an embodiment of the present disclosure;

[0080] Figure 9 FIG. 1 shows a flowchart of a method for training a classification model according to an embodiment of the present disclosure;

[0081] Figure 10 FIG. 1 shows a flowchart of a method for training a classification model according to an embodiment of the present disclosure;

[0082] Figure 11 FIG. 1 shows a flowchart of a method for training a classification model according to an embodiment of the present disclosure; DETAILED DESCRIPTION

[0083] Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so as to be easily implemented by those skilled in the art. Also, portions irrelevant to the description of the exemplary embodiments are omitted in the accompanying drawings for the sake of clarity.

[0084] In the present disclosure, it should be understood that terms such as "include" or "have" are intended to indicate that there are features, numbers, steps, actions, components, parts or combinations thereof disclosed in the specification, and do not exclude the possibility of adding one or more other features, numbers, steps, actions, components, parts or combinations thereof.

[0085] It should also be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict. The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0086] As described above, the existing state recognition method generally performs state recognition based on the attribute features of the subjects themselves, without considering the association relationship between the subjects, and has the problem of poor accuracy.

[0087] However, if the association relationship between the subjects is added as an input to the training of the prediction model, due to the very complex influence of the association relationship on the to-be-identified subject and the absence of a certain rule, the training is often difficult to converge, or the problem of overfitting occurs, and the effect is not good.

[0088] The embodiments of the present disclosure provide a state recognition method and a training method of a classification model. The training method of the classification model comprises: obtaining training sample data, the training sample data comprising graph features, non-graph features and state information of a subject; determining a first prediction result of the subject through the non-graph features, and comparing the first prediction result with the state information to determine a first error; determining a second prediction result of the subject according to at least the graph features, and comparing the second prediction result with the state information to determine a second error; determining a first label of the subject based on the relative sizes of the first error and the second error, the first label indicating whether to use or not to use the graph features for state recognition, and being used for training a first classification model, so that the first classification model trained can recognize whether the subject needs to use the graph features for state recognition, which is beneficial to targeted state recognition of different subjects, so as to improve the recognition accuracy.

[0089] Figures 1A-1D A schematic diagram of a subject relationship graph according to an embodiment of the present disclosure is shown.

[0090] As shown in the figure, Figures 1A-1D In each of the subject relationship graphs,

[0091] As Figures 1A-1D shown, if the labels of the neighbor nodes are used to infer the label of the center node A, it is difficult to obtain a good effect. For example, if the data distribution in the relationship graph is as Figure 1A shown, the features learned by the model should conform to "if there are more samples with label 1 in the neighbor nodes, the probability that the label of the A node is 1 is large". If the data distribution in the relationship graph is as Figure 1B shown, the features learned by the model should conform to "if there are more samples with label 0 in the neighbor nodes, the probability that the label of the A node is 1 is large". If the data distribution in the relationship graph is as Figure 1C shown, the features learned by the model should conform to "if there are more samples with label 1 in the neighbor nodes, the probability that the label of the A node is 0 is large". If the data distribution in the relationship graph is as Figure 1DAs shown, the features learned by the model should conform to the graph feature of the label distribution of the neighbor nodes not helping to predict the label of the center node A.

[0092] The distribution in the actual data often has many distributions between Figures 1A-1D If the same set of graph feature models is used to learn the distribution rules of all samples, for example, the data is mixed for training, then the final result is likely to be that the overall distribution is as shown in Figures 1A-1D The label distribution of the neighbor nodes is relatively average around the center node in terms of the true label of the center node, so the graph feature is not helpful. Figure 1D

[0093] Embodiments of the present disclosure propose a graph model classification scheme based on self-supervised pre-classification to solve the graph model effect on the data set with different label distributions in the graph. Through pre-classification of samples, better recognition effect can be achieved on different types of sample categories.

[0094] Figure 2 A flowchart of a state recognition method according to an embodiment of the present disclosure is shown.

[0095] As shown in Figure 2 , the method comprises operations S210-S230:

[0096] In operation S210, the graph features and non-graph features of the subject are obtained;

[0097] In operation S220, based on the first classification model, a first classification result is determined using or not using the graph features;

[0098] In operation S230, in the case where the first classification result indicates that the graph features are used for state recognition, the state of the subject is determined at least according to the graph features of the subject.

[0099] ​According to an embodiment of the present disclosure, the graph feature comprises a feature determined based on an association of the subject to be identified in the subject relationship graph. A graph convolutional network (GCN) or a GraphSage and the like are end-to-end models directly trained using neighbor node attributes, and the graph feature is used when the models are used to identify the state. Alternatively, the attributes of the neighbor nodes can be calculated offline first, aggregated to the current node as additional graph features, and then trained using a general classifier such as a gradient boosting decision tree (GBDT), a random forest, an extreme gradient boosting (XGBoost), and the like, and the model obtained using this method is also used to identify the state. Other commonly used prediction models are prediction models that do not use graph features, and these models are used to identify the state without using the graph feature.

[0100] According to an embodiment of the present disclosure, the graph feature comprises a feature determined based on a state of a neighboring subject of the subject to be identified in the subject relationship graph.

[0101] According to the technical solution provided by the embodiment of the present disclosure, the graph feature determined based on the state of the neighboring subject of the subject to be identified in the subject relationship graph is used, so that the information of the neighboring subject can be effectively used, and the accuracy of the state identification of the subject to be identified is improved.

[0102] According to an embodiment of the present disclosure, the non-graph feature is a concept opposite to the graph feature, and refers to a feature of the subject to be identified itself. For example, taking an enterprise subject as an example, the basic business information (such as registered capital, number of people, establishment time…) of the enterprise, the financial indicators (net assets, net profit, revenue…), historical risk information (the number of times of being sued last year, whether there has been a loss of trust…) and the like can be used as non-graph features of the enterprise. In the embodiment of the present disclosure, whether the graph feature is used or not, the non-graph feature can be used for state identification.

[0103] According to the technical solution provided by the embodiment of the present disclosure, the graph feature and the non-graph feature of the subject are obtained; based on the first classification model, a first classification result of using or not using the graph feature is determined; in the case that the first classification result indicates that the graph feature is used for state identification, the state of the subject is determined at least according to the graph feature of the subject, so that targeted state identification of different subjects is realized, and the graph feature is used to identify the state of the subject to be identified only when the first classification result indicates that the graph feature of the subject to be identified is needed, so that the accuracy of state identification is improved.

[0104] Figure 3A flowchart of a state recognition method according to another embodiment of the present disclosure is shown.

[0105] like Figure 3 As shown, the method is in Figure 2 Based on the illustrated embodiment, operation S310 is also included.

[0106] In operation S310, if the first classification result indicates that state recognition is not performed using graph features, the state of the subject is determined based on the non-graph features of the subject.

[0107] According to embodiments of this disclosure, for a subject to be identified that does not require the use of graph features, during the identification state, the non-graph features can be used without employing the graph features of the subject to be identified.

[0108] According to the technical solution provided in this disclosure, by determining the state of the subject based on the non-graph features of the subject when the first classification result indicates that state recognition is not performed using graph features, different state recognition methods can be adopted for different first classification results, thereby improving the accuracy of state recognition.

[0109] The following is combined Figure 4 The training method for the first classification model is illustrated by example.

[0110] Figure 4 A flowchart illustrating a training method for a classification model according to an embodiment of the present disclosure is shown.

[0111] like Figure 4 As shown, the method may further include operations S410 to S440:

[0112] In operation S410, training sample data is obtained, which includes the graph features, non-graph features and state information of the subject.

[0113] In operation S420, a first prediction result of the subject is determined by the non-graph features and compared with the state information to determine a first error;

[0114] In operation S430, at least based on the graph features, a second prediction result of the subject is determined, and compared with the state information to determine a second error;

[0115] In operation S440, a first label for the subject is determined based on the relative magnitudes of the first error and the second error. The first label indicates whether or not graph features are used for state recognition and is used to train a first classification model.

[0116] According to the embodiments of the present disclosure, the obtained training sample data is preprocessed, and then the training sample data is divided into a training set and a test set. The classifier is trained on the training set and evaluated on the test set. The classifier can be a binary classifier.

[0117] According to the embodiments of the present disclosure, a third state recognition model that does not use graph features can be constructed based on the non-graph features. The non-graph features in the training sample data can be used to determine a first prediction result by using the third state recognition model.

[0118] According to the embodiments of the present disclosure, a fourth state recognition model that uses graph features can be constructed based on the non-graph features and the graph features. The non-graph features and the graph features in the training sample data can be used to determine a second prediction result by using the fourth state recognition model.

[0119] For example, the current training sample data has an original label y train . Using a classifier such as a gradient boosting tree, a random forest, an extreme gradient boosting, etc., a risk recognition model that only uses non-graph features can be constructed as a third state recognition model, and a prediction result y pred1 is generated. Another risk recognition model that uses both non-graph features and graph features can be constructed as a fourth state recognition model, and a prediction result y pred2 is generated. Then, the absolute error |δy1| of y pred1 and ytrain, and the absolute error |δy2| of y pred2 and ytrain can be calculated. If |δy2| is less than |δy1|, it indicates that the sample using graph features is more effective than not using graph features, and vice versa. The sign(|δy2|-|δy1|) is used as a first label indicating whether graph features are needed or not in the embodiments of the present disclosure, and a classifier is trained, i.e., a first classification model. After the training is completed, the first classification model is used to predict the training sample data, and a dataset Train_1 that does not use graph features and a dataset Train_2 that needs to use graph features are obtained. Different state recognition methods or models are used for state recognition for the dataset Train_1 that does not use graph features and the dataset Train_2 that uses graph features, which can effectively improve the accuracy of recognition.

[0120] According to the technical solution provided in this disclosure, training sample data is obtained, including graph features, non-graph features, and state information of the subject; a first prediction result of the subject is determined by the non-graph features and compared with the state information to determine a first error; a second prediction result of the subject is determined at least based on the graph features and compared with the state information to determine a second error; a first label of the subject is determined based on the relative magnitude of the first error and the second error, the first label indicating whether graph features are used for state recognition and used to train a first classification model. Thus, the trained first classification model can identify whether the subject needs to use graph features for state recognition, which is beneficial for achieving targeted state recognition for different subjects, thereby improving recognition accuracy.

[0121] For data that requires graph features, it's helpful to explain some attributes of neighboring nodes to determine the current node's state. For example... Figures 1A-1C The neighbor node attributes of the three graphs clearly affect the central node A, but the differences are... Figure 1A The attributes of neighboring nodes and the state of the central node A are positively correlated, while Figure 1B and Figure 1C The attributes of neighboring nodes and the state of the central node A are negatively correlated. For Figure 1A In this case, the learned pattern is: "The larger the state value of a neighboring node, the more likely the graph features will affect the y-value of node A." pred2 "Pull up in a larger direction." And for... Figure 1B and Figure 1C In other words, the pattern they learned is: "The larger the state value of a neighboring node, the more likely the graph features will affect the y-value of node A." pred2 "Pull up in the direction of smaller values" or "The smaller the state value of the neighboring node, the more the graph features will pull the y-value of node A." pred2 "Pull up in a larger direction", that is, "the graph features pull the y-axis of node A". pred2 "Pull up in the opposite direction of the state value change of the neighboring node".

[0122] In this regard, embodiments of the present disclosure may optionally employ a further classification method to process the data, thereby improving prediction accuracy.

[0123] According to an embodiment of this disclosure, when the first classification result indicates that state recognition is performed using graph features, determining the state of the subject at least based on the graph features of the subject includes:

[0124] Based on the second classification model, the graph features and the non-graph features are processed to determine whether the graph feature effect direction is positive or negative;

[0125] identify a state of the subject based on the second classification result.

[0126] According to the embodiments of the present disclosure, a second classification model can be selected to identify the direction of the graph feature effect, for example, a positive direction or a negative direction. For the subject with different directions of the graph feature effect, different state recognition models are used for processing to identify the state of the subject.

[0127] According to the embodiments of the present disclosure, the second classification result of the direction of the graph feature effect is determined by processing the graph feature and the non-graph feature based on the second classification model. The state recognition model corresponding to the second classification result is used to process the graph feature and the non-graph feature to identify the state of the subject. Therefore, different state recognition models are used according to the second classification result, which can further improve the accuracy of state recognition.

[0128] According to the embodiments of the present disclosure, the state recognition model includes a first state recognition model for processing the graph feature effect in the positive direction and a second state recognition model for processing the graph feature effect in the negative direction. The state recognition model corresponding to the second classification result is used to process the graph feature and the non-graph feature to identify the state of the subject, which includes:

[0129] In the case where the second classification result indicates that the direction of the graph feature effect is positive, the state of the subject is determined based on the first state recognition model.

[0130] In the case where the second classification result indicates that the direction of the graph feature effect is negative, the state of the subject is determined based on the second state recognition model.

[0131] According to the embodiments of the present disclosure, the first state recognition model and the second state recognition model can be trained respectively for data with different second classification results. That is, the fourth state recognition model in the foregoing embodiments can be implemented as two independent state recognition models, which are respectively used to process the subject to be identified with the direction of the graph feature effect being positive or negative, and identify the state thereof.

[0132] According to the embodiments of the present disclosure, in the case where the second classification result indicates that the direction of the graph feature effect is positive, the state of the subject is determined based on the first state recognition model for processing the graph feature effect in the positive direction. In the case where the second classification result indicates that the direction of the graph feature effect is negative, the state of the subject is determined based on the second state recognition model for processing the graph feature effect in the negative direction. Therefore, the state recognition is targeted for different subjects, which can improve the accuracy of state recognition.

[0133] The following will be described in combination withFigure 5 The training method of the second classification model is exemplarily described.

[0134] Figure 5 A flowchart of a training method of a classification model according to another embodiment of the present disclosure is shown.

[0135] As Figure 5 shown, the method can further include operations S510 and S520 based on the first label: Figure 4

[0136] In operation S510, a training sample subset using graph features for state recognition in the training sample data is determined based on the first label.

[0137] In operation S520, for the subjects in the training sample subset, the second label of the subjects is determined based on the comparison of the first prediction result and the second prediction result, the second label indicating that the graph feature effect direction is a positive direction or a negative direction, and being used for training a second classification model.

[0138] Similar to the first classification model, the original label of the current training data is y train , a risk recognition model is constructed using GBDT, random forest, XGBoost, etc. classifiers, and non-graph features as a third state recognition model, and the prediction result is y pred1 , another risk recognition model is constructed using non-graph features and graph features as a fourth state recognition model, and the prediction result is y pred2 .

[0139] According to an embodiment of the present disclosure, the second label can be determined by non-graph features and graph features, for example, the prediction result obtained by using only non-graph features and the prediction result obtained by using non-graph features and graph features can be compared to determine that the graph feature effect direction is a positive direction or a negative direction.

[0140] According to an embodiment of the present disclosure, after y pred1 and y pred2 are determined, the relative size of the two can be judged. For example, as shown in the following table:

[0141] Table 1

[0142] Sample No. [[ y train ]]> [[ y pred1 ]]> [[ y pred2 ]]> y pred2 y pred1 ?]]> 1 0.5 0.6 0.7 Yes 2 0.5 0.6 0.8 Yes 3 0.5 0.6 0.55 No 4 0.5 0.6 0.52 No 5 0.5 0.6 0.4 No 6 0.5 0.6 0.3 No 7 0.5 0.4 0.3 No 8 0.5 0.4 0.2 No 9 0.5 0.4 0.45 Yes 10 0.5 0.4 0.47 Yes 11 0.5 0.4 0.6 Yes 12 0.5 0.4 0.7 Yes

[0143] In an embodiment of the present disclosure, sign(y pred2 >y pred1 ​) For the label, the training samples can be divided into two groups {1, 2, 9, 10, 11, 12} and {3, 4, 5, 6, 7, 8}, the first group is the sample whose graph feature effect direction is positive direction after using the graph feature, that is, the predicted value is larger than that without using the graph feature; the second group is the sample whose graph feature effect direction is negative direction after using the graph feature, the predicted value is smaller than that without using the graph feature. sign (y pred2 >y pred1 ) is the second label. By training the classification model based on the sample data with the labeled second label, a second classification model for generating a second classification result can be obtained.

[0144] According to the technical scheme provided by the embodiments of the present disclosure, the training sample subset using the graph feature for state recognition in the training sample data is determined based on the first label; for the subject in the training sample subset, the second label of the subject is determined based on the comparison of the first prediction result and the second prediction result, the second label indicates that the graph feature effect direction is positive direction or negative direction, and is used for training a second classification model, so that the second classification model obtained by training can identify the effect influence direction of the graph feature on the subject, which is beneficial to the state recognition of different subjects in a targeted manner, so as to improve the recognition accuracy.

[0145] According to the embodiments of the present disclosure, the method further comprises:

[0146] In the case that the first classification result indicates that the state recognition is not performed using the graph feature, the non-graph feature of the subject is processed using a third state recognition model to determine the state of the subject,

[0147] Wherein, in the case that the subject is multiple, the method further comprises:

[0148] The outputs of the first state recognition model, the second state recognition model and the third state recognition model are summarized to obtain the state recognition result of the multiple subjects.

[0149] Figure 6 A flowchart of a state recognition method according to another embodiment of the present disclosure is shown.

[0150] As Figure 6As shown, for each sample, the pre-trained first classification model first determines whether graph features need to be used. If graph features are not needed, the third state recognition model is used directly for state recognition. If graph features are needed, the second classification model is used to determine the positive or negative correlation of the graph features for the current sample. If the graph feature effect is determined to be positive, the first state recognition model is used for state recognition; if the graph feature effect is determined to be negative, the second state recognition model is used for state recognition. After all subjects to be identified have been predicted, the states are summarized, or in other words, the prediction results are merged and evaluated. The above process is similar for the test set or the data used in actual production. In the testing phase, after the state summary, an indicator evaluation can also be performed.

[0151] As mentioned above, if the distribution of graph features has inconsistent effects on the current node, using a single graph model often has little effect compared to not using any graph model at all. This is because training with multiple samples of inconsistent distributions can cause the effects of graph features to cancel each other out. In this embodiment, a classifier is used beforehand to classify samples with the same graph distribution, and then a suitable model is trained for each class of samples. In this way, graph features will have a more significant effect.

[0152] To address this, this embodiment of the disclosure designs two cascaded self-supervised classifiers. On the training set, the first classification model distinguishes whether the training set samples require the use of a graph model, while the second classification model further distinguishes whether the current graph features have a positive or negative impact on the prediction result for those training set samples that do require the use of a graph model. Finally, corresponding models are trained on these three types of training set samples respectively for use on the test set. This fully leverages the capabilities of the graph model.

[0153] According to the technical solution provided in the embodiments of this disclosure, when the first classification result indicates that no graph features are used for state recognition, a third state recognition model is used to process the non-graph features of the subject to determine the state of the subject. In the case of multiple subjects, the method further includes: summarizing the outputs of the first state recognition model, the second state recognition model and the third state recognition model to obtain the state recognition results of multiple subjects, thereby enabling targeted state recognition for different subjects and improving the accuracy of state recognition.

[0154] Figure 7 A flowchart illustrating an enterprise risk identification method according to an embodiment of this disclosure is shown.

[0155] like Figure 7 As shown, the method includes operations S710 to S730:

[0156] During operation of S710, obtain the graph features and non-graph features of the enterprise to be identified;

[0157] In operation S720, a first classification result using or not using the graph feature is determined based on the first classification model.

[0158] In operation S730, in a case where the first classification result indicates that the graph feature is used for risk identification, a risk of the enterprise is identified at least according to the graph feature of the enterprise.

[0159] According to the technical scheme provided by the embodiment of the present disclosure, the graph feature and the non-graph feature of the to-be-identified enterprise are obtained, the first classification result using or not using the graph feature is determined based on the first classification model, and in a case where the first classification result indicates that the graph feature is used for risk identification, the risk of the enterprise is identified at least according to the graph feature of the enterprise, thereby achieving targeted state identification for different subjects and improving the accuracy of state identification.

[0160] For example, before an enterprise intends to sign a contract with other subjects, the company lawyer of the enterprise can apply the method of the embodiment of the present disclosure to evaluate the risk state of the opposite subject as a basis for decision-making.

[0161] Those skilled in the art can understand that the above scheme in the embodiment of the present disclosure can not only be used to determine the risk of the to-be-identified enterprise, but also be used to identify the state or risk of an organization or personnel within or outside the enterprise, and be used to identify the state or risk of a biological or non-biological object in nature, and be used to identify the state or risk of a to-be-identified subject in the form of data.

[0162] For example, the subject of the embodiment of the present disclosure can be a merchant in an e-commerce platform, and the standardized state index of the merchant is judged according to the method of the embodiment of the present disclosure, and the standardized state index is used to manage the merchant or to handle complaints against the merchant. For another example, a lawyer can determine the standardized state index of a subject by the method of the embodiment of the present disclosure, which can be used as an evaluation of the subject and can be used to simulate the viewpoint of the court. For another example, the procuratorate can apply the method of the embodiment of the present disclosure to make a prediction, which can be used to form a plan. In addition, in the processing of various civil, administrative management, or civil and administrative disputes, the people's court, the arbitration committee, the government and the working departments of the government can all apply the standardized state index generated by the method of the embodiment of the present disclosure as a reference. The method of the embodiment of the present disclosure can also be used to check whether the focus of the dispute has changed.

[0163] Figure 8 A block diagram of a training device 800 of a classification model according to an embodiment of the present disclosure is shown. The device can be realized as part or all of an electronic device by software, hardware, or a combination of both.

[0164] As Figure 8As shown, the training apparatus 800 of the classification model comprises a first obtaining module 810, a first determining module 820, a second determining module 830, and a third determining module 840.

[0165] The first obtaining module 810 is configured to obtain training sample data, wherein the training sample data comprises graph features, non-graph features, and state information of a subject.

[0166] The first determining module 820 is configured to determine a first prediction result of the subject by using the non-graph features, and compare the first prediction result with the state information to determine a first error.

[0167] The second determining module 830 is configured to determine a second prediction result of the subject according to at least the graph features, and compare the second prediction result with the state information to determine a second error.

[0168] The third determining module 840 is configured to determine a first label of the subject based on a relative size of the first error and the second error, wherein the first label indicates whether to use or not to use the graph features for state identification, and is used for training a first classification model.

[0169] According to the technical scheme provided by the embodiments of the present disclosure, the first obtaining module 810 is configured to obtain training sample data, wherein the training sample data comprises graph features, non-graph features, and state information of a subject; the first determining module 820 is configured to determine a first prediction result of the subject by using the non-graph features, and compare the first prediction result with the state information to determine a first error; the second determining module 830 is configured to determine a second prediction result of the subject according to at least the graph features, and compare the second prediction result with the state information to determine a second error; and the third determining module 840 is configured to determine a first label of the subject based on a relative size of the first error and the second error, wherein the first label indicates whether to use or not to use the graph features for state identification, and is used for training a first classification model. Thus, the first classification model trained can identify whether the subject needs to use the graph features for state identification, which is beneficial to achieving targeted state identification for different subjects, so as to improve the identification accuracy.

[0170] According to the embodiments of the present disclosure, the training apparatus 800 of the classification model can further comprise a fourth determining module 850 and a fifth determining module 860.

[0171] The fourth determining module 850 is configured to determine a training sample subset of the training sample data for which the graph features are used for state identification based on the first label.

[0172] The fifth determining module 860 is configured to determine a second label for the subject in the training sample subset based on a comparison of the first prediction result and the second prediction result. The second label indicates whether the direction of the graph feature effect is positive or negative and is used to train the second classification model.

[0173] According to the technical solution provided in this embodiment, the fourth determining module 850 is configured to determine a subset of training samples for state recognition using graph features in the training sample data based on the first label; the fifth determining module 860 is configured to determine a second label for the subject in the subset of training samples based on a comparison of the first prediction result and the second prediction result. The second label indicates whether the direction of the graph feature effect is positive or negative, and is used to train a second classification model. Thus, the trained second classification model can identify the direction of the graph feature effect on the subject, which is beneficial for achieving targeted state recognition for different subjects and improving the recognition accuracy.

[0174] Figure 9 A block diagram of a state recognition device 900 according to an embodiment of the present disclosure is shown. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both.

[0175] like Figure 9 As shown, the state recognition device 900 includes a second acquisition module 910, a sixth determination module 920, and a seventh determination module 930.

[0176] The second acquisition module 910 is configured to acquire the graph features and non-graph features of the subject;

[0177] The sixth determining module 920 is configured to determine a first classification result based on the first classification model, whether or not the graph features are used;

[0178] The seventh determining module 930 is configured to determine the state of the subject based at least on the graph features of the subject if the first classification result indicates that state recognition is performed using graph features.

[0179] According to the technical solution provided in the embodiments of this disclosure, the second obtaining module 910 is configured to obtain graph features and non-graph features of the subject; the sixth determining module 920 is configured to determine a first classification result based on a first classification model, using or not using the graph features; and the seventh determining module 930 is configured to determine the state of the subject at least based on the graph features of the subject if the first classification result indicates that graph features are used for state recognition, thereby enabling targeted state recognition for different subjects and improving the accuracy of state recognition.

[0180] According to the embodiment of the present disclosure, the state recognition apparatus 900 can further include an eighth determination module 940 configured to determine the state of the subject according to non-graphic features of the subject in a case where the first classification result indicates that the state recognition is not to be performed using the graphic features.

[0181] According to the embodiment of the present disclosure, the eighth determination module 940 is configured to determine the state of the subject according to non-graphic features of the subject in a case where the first classification result indicates that the state recognition is not to be performed using the graphic features, so that different state recognition means are adopted for different first classification results, and the accuracy of state recognition can be improved.

[0182] According to the embodiment of the present disclosure, in a case where the first classification result indicates that the state recognition is to be performed using the graphic features, the state of the subject is determined according to at least the graphic features of the subject, including:

[0183] processing the graphic features and the non-graphic features based on a second classification model to determine a second classification result indicating that the direction of the effect of the graphic features is a positive direction or a negative direction;

[0184] processing the graphic features and the non-graphic features based on a state recognition model corresponding to the second classification result to determine the state of the subject.

[0185] According to the embodiment of the present disclosure, the second classification model is used to process the graphic features and the non-graphic features to determine a second classification result indicating that the direction of the effect of the graphic features is a positive direction or a negative direction, and the graphic features and the non-graphic features are processed based on a state recognition model corresponding to the second classification result to determine the state of the subject, so that different state recognition models are adopted according to the second classification result, and the accuracy of state recognition can be further improved.

[0186] According to the embodiment of the present disclosure, the state recognition model includes a first state recognition model for processing the graphic features with a positive direction of effect and a second state recognition model for processing the graphic features with a negative direction of effect, and the processing of the graphic features and the non-graphic features based on the state recognition model corresponding to the second classification result to determine the state of the subject includes:

[0187] in a case where the second classification result indicates that the direction of the effect of the graphic features is a positive direction, the state of the subject is determined based on the first state recognition model;

[0188] in a case where the second classification result indicates that the direction of the effect of the graphic features is a negative direction, the state of the subject is determined based on the second state recognition model.

[0189] According to the technical scheme provided by the embodiment of the present disclosure, in the case that the second classification result indicates that the direction of the picture feature effect is positive, the state of the subject is determined based on the first state recognition model for processing the picture feature effect in the positive direction; in the case that the second classification result indicates that the direction of the picture feature effect is negative, the state of the subject is determined based on the second state recognition model for processing the picture feature effect in the negative direction, thereby achieving targeted state recognition for different subjects, and the accuracy of state recognition can be improved.

[0190] According to the embodiment of the present disclosure, the state recognition device 900 can further include:

[0191] The eighth determination module 940 is configured to, in the case that the first classification result indicates that the picture feature is not used for state recognition, process the non-picture feature of the subject using a third state recognition model to determine the state of the subject.

[0192] The aggregation module 950 is configured to aggregate the outputs of the first state recognition model, the second state recognition model and the third state recognition model to obtain the state recognition results of multiple subjects.

[0193] According to the technical scheme provided by the embodiment of the present disclosure, the eighth determination module 940 is configured to, in the case that the first classification result indicates that the picture feature is not used for state recognition, process the non-picture feature of the subject using a third state recognition model to determine the state of the subject; and the aggregation module 950 is configured to aggregate the outputs of the first state recognition model, the second state recognition model and the third state recognition model to obtain the state recognition results of multiple subjects, thereby achieving targeted state recognition for different subjects, and the accuracy of state recognition can be improved.

[0194] The present disclosure also discloses an electronic device, Figure 10 A block diagram of an electronic device 1000 according to an embodiment of the present disclosure is shown.

[0195] As shown in the figure, Figure 10 The electronic device 1000 includes a memory 1001 and a processor 1002, wherein the memory 1001 is used to store one or more computer instructions, and the one or more computer instructions are executed by the processor 1002 to implement the following operations:

[0196] Obtain training sample data, the training sample data including picture features, non-picture features and state information of a subject;

[0197] Determine a first prediction result of the subject through the non-picture features, and compare the first prediction result with the state information to determine a first error;

[0198] determine a second prediction result of the subject according to the graph feature, and compare the second prediction result with the state information to determine a second error;

[0199] determine a first label of the subject based on relative sizes of the first error and the second error, the first label indicating whether to use or not to use the graph feature for state identification, and the first label being used for training a first classification model.

[0200] According to an embodiment of the present disclosure, the processor 1002 is further configured to perform:

[0201] determine a training sample subset in the training sample data for which state identification is performed using the graph feature based on the first label;

[0202] for the subject in the training sample subset, determine a second label of the subject based on comparison of the first prediction result and the second prediction result, the second label indicating whether the graph feature effect direction is a positive direction or a negative direction, and the second label being used for training a second classification model.

[0203] According to an embodiment of the present disclosure, the one or more computer instructions are executed by the processor 1002 to implement the following operations:

[0204] obtain a graph feature and a non-graph feature of a subject;

[0205] determine a first classification result of using or not using the graph feature based on the first classification model;

[0206] in a case where the first classification result indicates that state identification is performed using the graph feature, determine a state of the subject according to at least the graph feature of the subject.

[0207] According to an embodiment of the present disclosure, the processor 1002 is further configured to perform:

[0208] in a case where the first classification result indicates that state identification is not performed using the graph feature, determine the state of the subject according to the non-graph feature of the subject.

[0209] According to an embodiment of the present disclosure, the in the case where the first classification result indicates that state identification is performed using the graph feature, determining the state of the subject according to at least the graph feature of the subject, comprises:

[0210] processing the graph feature and the non-graph feature based on the second classification model to determine a second classification result of whether the graph feature effect direction is a positive direction or a negative direction;

[0211] processing the graph feature and the non-graph feature based on a state identification model corresponding to the second classification result to identify and determine the state of the subject.

[0212] According to an embodiment of the present disclosure, the state recognition model comprises a first state recognition model for processing the graph feature effect as a positive direction and a second state recognition model for processing the graph feature effect as a negative direction, and the state of the subject is determined by processing the graph feature and the non-graph feature based on the state recognition model corresponding to the second classification result, comprising:

[0213] In a case where the second classification result indicates that the graph feature effect direction is a positive direction, the state of the subject is determined based on the first state recognition model.

[0214] In a case where the second classification result indicates that the graph feature effect direction is a negative direction, the state of the subject is determined based on the second state recognition model.

[0215] According to an embodiment of the present disclosure, in a case where the first classification result indicates that the graph feature is not used for state recognition, the state of the subject is determined according to the non-graph feature of the subject, comprising:

[0216] In a case where the first classification result indicates that the graph feature is not used for state recognition, the non-graph feature of the subject is processed using a third state recognition model to determine the state of the subject.

[0217] In a case where the subject is multiple, the processor 1002 is further configured to perform:

[0218] The outputs of the first state recognition model, the second state recognition model and the third state recognition model are summarized to obtain state recognition results of multiple subjects.

[0219] According to an embodiment of the present disclosure, the one or more computer instructions are executed by the processor 1002 to implement the following operations:

[0220] Obtaining the graph feature and the non-graph feature of the enterprise to be recognized;

[0221] Based on the first classification model, a first classification result of using or not using the graph feature is determined.

[0222] In a case where the first classification result indicates that the graph feature is used for risk recognition, the risk of the enterprise is recognized at least according to the graph feature of the enterprise.

[0223] Figure 11 A structural schematic diagram of a computer system suitable for implementing the method and device of the present disclosure is shown.

[0224] As Figure 11As shown, the computer system 1100 includes a processing unit 1101 which can execute the various processing functions of the embodiments described above according to a program stored in a read-only memory (ROM) 1102 or loaded into a random access memory (RAM) 1103 from a storage section 1108. Various programs and data required by the system 1100 for operation are also stored in the RAM 1103. The processing unit 1101, the ROM 1102, and the RAM 1103 are connected to each other by a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0225] Connected to the I / O interface 1105 are an input section 1106 including a keyboard, a mouse, etc.; an output section 1107 including a display device such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1108 including a hard disk, etc.; and a communication section 1109 including a network interface card such as a LAN card, a modem, etc. The communication section 1109 performs communication processing via a network such as the Internet. A drive 1110 is also connected to the I / O interface 1105 as required. A removable recording medium 1111 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 1110 as required, so that a computer program read therefrom is installed into the storage section 1108 as required. The processing unit 1101 can be implemented as a CPU, a GPU, a TPU, a FPGA, a NPU, etc.

[0226] In particular, the methods described above can be implemented as a computer software program according to embodiments of the present disclosure. For example, embodiments of the present disclosure include a computer program product comprising a computer program tangibly embodied on a non-transitory computer readable medium, the computer program containing program code for executing the methods described above. In such embodiments, the computer program can be downloaded and installed from a network via the communication section 1109 and / or installed from the removable recording medium 1111.

[0227] The flow and block diagrams in the drawings represent possible architectural, functional, and operational architectures of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block can represent a module, a segment, or a portion of code that comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or operations, or combinations of special purpose hardware and computer instructions.

[0228] The units or modules described in the embodiments of the present disclosure can be implemented by software or by programmable hardware. The described units or modules can also be arranged in a processor, and the names of the units or modules do not constitute a limitation on the units or modules themselves in some cases.

[0229] As another aspect, the present disclosure also provides a computer readable storage medium, which can be the computer readable storage medium included in the electronic device or the computer system in the above embodiments, or can exist separately from the device and not be assembled into the device. The computer readable storage medium stores one or more programs for execution by one or more processors to perform the methods described in the present disclosure.

[0230] The above description is merely preferred embodiments of the present disclosure and a description of principles of applied technologies. It should be understood by those skilled in the art that the scope of the present disclosure is not limited to the technical solutions formed by the specific combinations of the above technical features, and also includes other technical solutions formed by the combinations of the above technical features or equivalent features without departing from the inventive concept. For example, the above technical features can be replaced with the technical features disclosed in the present disclosure (but not limited to) having similar functions to form technical solutions.

Claims

1. A method for training a classification model, comprising: obtaining training sample data, the training sample data comprising graph features, non-graph features and state information of a subject, wherein the graph features are used to represent features determined based on states of adjacent subjects of the subject in a subject relationship graph, and the non-graph features are used to represent features of the subject itself; determining a first prediction result of the subject based on the non-graph features, and comparing the first prediction result with the state information to determine a first error; determining a second prediction result of the subject based on at least the graph features, and comparing the second prediction result with the state information to determine a second error; determining a first label of the subject based on a relative size of the first error and the second error, the first label indicating whether or not to use the graph features for state recognition, and being used to train a first classification model, wherein the first classification model is used to distinguish whether or not the graph features need to be used for training set samples adopted by the first classification model; the method further comprises: in response to the second error being less than the first error, determining that the first label is to use the graph features for state recognition; and in response to the second error being greater than or equal to the first error, determining that the first label is not to use the graph features for state recognition.

2. The method of claim 1, wherein, the method further comprises: determining a training sample subset of the training sample data in which the graph features are used for state recognition based on the first label; for the subject in the training sample subset, determining a second label of the subject based on a comparison of the first prediction result and the second prediction result, the second label indicating whether or not the graph features have a positive or negative effect, and being used to train a second classification model. 3.A method for state recognition, comprising: obtaining graph features and non-graph features of a subject, wherein the graph features are used to represent features determined based on states of adjacent subjects of the subject in a subject relationship graph, and the non-graph features are used to represent features of the subject itself; determining a first classification result of using or not using the graph features based on a first classification model, wherein the first classification model is used to distinguish whether or not the graph features need to be used for training set samples adopted by the first classification model; in a case where the first classification result indicates that the graph features are used for state recognition, determining a state of the subject based on at least the graph features. 4.The method of claim 3, further comprising: in a case where the first classification result indicates that the graph features are not used for state recognition, determining the state of the subject based on the non-graph features.

5. The method of claim 3, wherein, the determining the state of the subject based on at least the graph features in the case where the first classification result indicates that the graph features are used for state recognition, comprises: processing the graph features and the non-graph features based on a second classification model to determine a second classification result of whether or not the graph features have a positive or negative effect; and processing the graph features and the non-graph features based on a state recognition model corresponding to the second classification result to determine the state of the subject.

6. The method of claim 5, wherein, The state recognition model comprises a first state recognition model for processing the graph feature effect as a positive direction and a second state recognition model for processing the graph feature effect as a negative direction, and the state of the subject is determined by processing the graph feature and the non-graph feature based on the state recognition model corresponding to the second classification result, comprising: In the case that the second classification result indicates that the graph feature effect direction is a positive direction, the state of the subject is determined based on the first state recognition model; In the case that the second classification result indicates that the graph feature effect direction is a negative direction, the state of the subject is determined based on the second state recognition model.

7. The method of claim 6, further comprising: In the case that the first classification result indicates that the graph feature is not used for state recognition, the non-graph feature is processed using a third state recognition model to determine the state of the subject, In the case that the subject is multiple, the method further comprises: The outputs of the first state recognition model, the second state recognition model and the third state recognition model are aggregated to obtain the state recognition results of the multiple subjects.

8. An enterprise risk recognition method, comprising: Obtaining graph features and non-graph features of an enterprise to be recognized, wherein the graph features are used to represent features determined based on the states of adjacent subjects of a subject to be recognized in a subject relationship graph, and the non-graph features are used to represent features of the subject to be recognized itself; Based on a first classification model, a first classification result of using or not using the graph features is determined, wherein the first classification model is used to distinguish whether the training set samples adopted by the first classification model need to use the graph features; In the case that the first classification result indicates that the graph features are used for risk recognition, the risk of the enterprise is recognized at least according to the graph features.

9. A training device of a classification model, comprising: A first obtaining module configured to obtain training sample data, the training sample data comprising graph features, non-graph features and state information of a subject, wherein the graph features are used to represent features determined based on the states of adjacent subjects of a subject to be recognized in a subject relationship graph, and the non-graph features are used to represent features of the subject to be recognized itself; A first determining module configured to determine a first prediction result of the subject by the non-graph features, and compare with the state information to determine a first error; A second determining module configured to determine a second prediction result of the subject at least according to the graph features, and compare with the state information to determine a second error; A third determining module configured to determine a first label of the subject based on the relative size of the first error and the second error, the first label indicating whether to use or not to use the graph features for state recognition, and used for training a first classification model, wherein the first classification model is used to distinguish whether the training set samples adopted by the first classification model need to use the graph features. The training apparatus of the classification model is further configured to perform the following steps: in response to the second error being less than the first error, determining that the first label is to be used for state recognition using the graph feature; and in response to the second error being greater than or equal to the first error, determining that the first label is not to be used for state recognition using the graph feature.

10. The apparatus of claim 9, wherein, Further comprising: A fourth determination module configured to determine, based on the first label, a training sample subset in the training sample data that is to be used for state recognition using the graph feature; A fifth determination module configured to, for a subject in the training sample subset, determine, based on a comparison of the first prediction result and the second prediction result, a second label of the subject, the second label indicating that the graph feature effect direction is a positive direction or a negative direction, and the second label being used for training a second classification model.

11. A state recognition apparatus, comprising: A second obtaining module configured to obtain a graph feature and a non-graph feature of a subject, wherein the graph feature is used to represent a feature determined based on a state of an adjacent subject of the subject in a subject relationship graph, and the non-graph feature is used to represent a feature of the subject itself; A sixth determination module configured to determine, based on a first classification model, a first classification result of whether to use or not to use the graph feature, wherein the first classification model is used to distinguish whether a training set sample adopted by the first classification model needs to use the graph feature; A seventh determination module configured to, in a case where the first classification result indicates that the graph feature is to be used for state recognition, determine a state of the subject at least according to the graph feature.

12. The apparatus of claim 11, further comprising: An eighth determination module configured to, in a case where the first classification result indicates that the graph feature is not to be used for state recognition, determine the state of the subject according to the non-graph feature.

13. The apparatus of claim 11, wherein, The case where the first classification result indicates that the graph feature is to be used for state recognition, and the state of the subject is determined at least according to the graph feature of the subject, comprises: processing the graph feature and the non-graph feature based on a second classification model to determine a second classification result indicating that the graph feature effect direction is a positive direction or a negative direction; processing the graph feature and the non-graph feature based on a state recognition model corresponding to the second classification result to determine the state of the subject.

14. The apparatus of claim 13, wherein, The state recognition model comprises a first state recognition model used to process a positive direction of graph feature effect and a second state recognition model used to process a negative direction of graph feature effect, and the processing of the graph feature and the non-graph feature based on the state recognition model corresponding to the second classification result to determine the state of the subject comprises: in a case where the second classification result indicates that the graph feature effect direction is a positive direction, determining the state of the subject based on the first state recognition model; and in a case where the second classification result indicates that the graph feature effect direction is a negative direction, determining the state of the subject based on the second state recognition model.

15. The apparatus of claim 14, further comprising: An eighth determining module is configured to, in a case where the first classification result indicates that the graph feature is not used for state recognition, process the non-graph feature using a third state recognition model to determine the state of the subject; An aggregating module is configured to aggregate outputs of the first state recognition model, the second state recognition model and the third state recognition model to obtain state recognition results of a plurality of subjects.

16. An electronic device, comprising: A computer readable storage medium storing one or more computer instructions, wherein the one or more computer instructions are executed by a processor to implement the method steps of any one of claims 1-8.

17. A readable storage medium, having stored thereon computer instructions, characterized in that, The computer instructions are executed by the processor to implement the method steps of any one of claims 1-8.

Images