A method, device, equipment and medium for determining abnormal samples

Abstract

The embodiments of the present invention provide a method, apparatus, device and medium for determining abnormal samples. The method determines the first feature corresponding to the first coefficient of variation with the largest value, the first training sample of the first model corresponding to the maximum feature weight of the first feature, and the second training sample of the second model corresponding to the minimum feature weight of the first feature. Based on the first training sample and the second training sample, a first target training sample and a second target training sample are obtained, and model training is performed. The second coefficient of variation of the first feature is calculated according to at least one model and the first target model, and the third coefficient of variation of the first feature is calculated according to at least one model model and the second target model. If the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, it is determined that there is no abnormal sample, thereby realizing the judgment of whether there is an abnormal sample in the sample data set.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, device, equipment and medium for determining an abnormal sample. Background Art

[0002] With the development of science and technology, models are being applied to more and more technical fields. Specifically, when people need to implement a certain function, they will prepare a large number of sample data sets in advance, and then select multiple groups of sample data from the sample data sets. Based on each selected group of sample data, each corresponding trained model is obtained. In order to select the model with the best performance from the multiple trained models and put it into production, the model will be tested using black box technology after the model training is completed. A large amount of data will be input into the trained model. For the prediction results output by the model, the model with the most accurate prediction results will be selected as the optimal model.

[0003] However, in the sample data set used for model training, there may be abnormal samples with data anomalies. If possible, the abnormal samples with data anomalies may cause inaccurate predictions of the trained model and affect the selection of the optimal model. However, in the prior art, there is no method for determining whether there are abnormal samples in the sample data set. Summary of the invention

[0004] The embodiments of the present invention provide a method, apparatus, device and medium for determining abnormal samples, which realize the determination of abnormal samples in a sample data set.

[0005] An embodiment of the present invention provides a method for determining an abnormal sample, the method comprising:

[0006] For at least one trained model, calculate the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model, and identify the first feature corresponding to the first coefficient of variation with the largest value;

[0007] Identify a first model corresponding to a maximum feature weight of the first feature, and a second model corresponding to a minimum feature weight;

[0008] Identify a first sub-training sample that is different from a first training sample for training the first model and a second training sample for training the second model, and a second sub-training sample that is the same as the second training sample in the first training sample, divide the first sub-training sample into two parts, and combine them with the second sub-training sample respectively to obtain a first target training sample and a second target training sample, and perform model training based on the first target training sample and the second target training sample respectively to obtain a first target model and a second target model;

[0009] Based on the at least one model and the first target model, the second coefficient of variation of the first feature is calculated, and based on the at least one model and the second target model, the third coefficient of variation of the first feature is calculated. If the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, it is determined that there are no abnormal samples.

[0010] The embodiment of the present invention further provides an abnormal sample determination device, the device comprising:

[0011] A calculation module, for calculating the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model for at least one trained model, and identifying a first feature corresponding to a first coefficient of variation with a maximum value;

[0012] an identification module, used to identify a first model corresponding to a maximum feature weight of the first feature, and a second model corresponding to a minimum feature weight;

[0013] An abnormal sample determination module is used to identify different first sub-training samples in a first training sample for training the first model and a second training sample for training the second model, and a second sub-training sample in the first training sample that is the same as the second training sample, divide the first sub-training sample into two parts, and combine them with the second sub-training sample respectively to obtain a first target training sample and a second target training sample, and perform model training based on the first target training sample and the second target training sample respectively to obtain a first target model and a second target model; calculate the second coefficient of variation of the first feature according to the at least one model and the first target model, and calculate the third coefficient of variation of the first feature according to the at least one model and the second target model; if the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, it is determined that there is no abnormal sample.

[0014] An embodiment of the present invention further provides an electronic device, which includes at least a processor and a memory, and the processor is used to implement the steps of any of the above-mentioned abnormal sample determination methods when executing a computer program stored in the memory.

[0015] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of any of the above-mentioned abnormal sample determination methods.

[0016] In an embodiment of the present invention, a first feature corresponding to a first coefficient of variation with a maximum value, a first training sample of a first model corresponding to a maximum feature weight of the first feature, and a second training sample of a second model corresponding to a minimum feature weight of the first feature are determined; based on the first training sample and the second training sample, a first target training sample and a second target training sample are obtained, and model training is performed; a second coefficient of variation of the first feature is calculated according to at least one model and the first target model; and a third coefficient of variation of the first feature is calculated according to at least one model and the second target model; if both the second coefficient of variation and the third coefficient of variation are smaller than the first coefficient of variation, it is determined that no abnormal sample exists, thereby realizing the judgment of whether there are abnormal samples in the sample data set. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0018] Figure 1 A schematic diagram of an abnormal sample determination process provided by an embodiment of the present invention;

[0019] Figure 2 A schematic diagram of the structure of an abnormal sample determination device provided by an embodiment of the present invention;

[0020] Figure 3 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] In order to determine whether there are abnormal samples in a sample data set for model training, the present invention provides an abnormal sample determination method, device, equipment and medium.

[0023] Embodiment 1:

[0024] Figure 1 A schematic diagram of an abnormal sample determination process provided by an embodiment of the present invention, the process includes the following steps:

[0025] S101: For at least one trained model, calculate the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model, and identify the first feature corresponding to the first coefficient of variation with the largest value.

[0026] An abnormal sample determination method provided by an embodiment of the present invention is applied to an electronic device, which may be a PC, a server, or other device.

[0027] In the embodiment of the present invention, when a sample data set is used to train each model, the training samples used by each model are generally part of the sample data set. Even if a sample data set is used to train multiple models with the same structure, the training results of each model are different. However, each trained model has the same features, but each feature has a different feature weight in each model, and each model includes at least two features.

[0028] In an embodiment of the present invention, for a sample data set, in order to determine whether there are abnormal samples in the sample data set, it is necessary to make a judgment based on the model where the maximum feature weight corresponding to the feature with the largest degree of discreteness in the model trained on the sample data set is located.

[0029] Specifically, for at least one model trained based on a sample data set, the coefficient of variation of each feature is calculated according to the feature weight corresponding to each feature in each model, and the first feature corresponding to the first coefficient of variation with the largest value is identified, so that the first feature can be used to subsequently determine whether there are abnormal samples in the sample data set.

[0030] For each feature, when determining the degree of discreteness of the feature according to the feature weight of the feature in each model, any one of the average value, standard deviation and coefficient of variation of the feature weight of the feature can be used to determine the degree of discreteness of the feature. However, if the difference between the maximum feature weight and the minimum feature weight of the feature is large, the accuracy and authenticity of the degree of discreteness of the feature determined by the average value or standard deviation will be reduced. In order to reduce the impact of the large difference between the maximum feature weight and the minimum feature weight of the feature on the calculation of the degree of discreteness of the feature, in an embodiment of the present invention, when determining the degree of discreteness of each feature, the degree of discreteness of the feature is determined by the coefficient of variation of the feature.

[0031] S102: Identify a first model corresponding to the maximum feature weight of the first feature and a second model corresponding to the minimum feature weight.

[0032] In order to further narrow the scope of abnormal samples, in an embodiment of the present invention, after determining the first feature corresponding to the first coefficient of variation with the largest coefficient of variation value, the first model corresponding to the maximum feature weight in the first feature and the second model corresponding to the minimum feature weight are determined. At this time, if there are abnormal samples in the sample data set, the abnormal samples are related to the training samples of the first model and the second model.

[0033] Specifically, in an embodiment of the present invention, after determining the first feature corresponding to the first coefficient of variation with the largest coefficient of variation value, the feature weight of the first feature in each model is counted, and the counted feature weights are compared to determine the first model corresponding to the maximum feature weight and the second model corresponding to the minimum feature weight.

[0034] S103: Identify first sub-training samples that are different in a first training sample for training the first model and a second training sample for training the second model, and a second sub-training sample in the first training sample that is the same as the second training sample, divide the first sub-training sample into two parts, and combine them with the second sub-training sample respectively to obtain a first target training sample and a second target training sample, and perform model training based on the first target training sample and the second target training sample respectively to obtain a first target model and a second target model.

[0035] In the embodiment of the present invention, after determining the first model corresponding to the maximum feature weight in the first feature and the second model corresponding to the minimum feature weight, the range of abnormal samples is determined to be the training samples of the first model and the second model. In order to determine the abnormal samples, the first training sample of the first model and the second training sample of the second model are determined, and the first sub-training sample of the first training sample that is different from the second training sample and the second sub-training sample of the first training sample that is the same as the second training sample are identified.

[0036] After obtaining the first sub-training sample and the second sub-training sample, in order to further select abnormal samples, a new target training sample is constructed based on the first sub-training sample and the second sub-training sample, and model training is performed based on the newly constructed target training sample, and the first sub-training sample and the second sub-training sample are analyzed according to the results of the model training.

[0037] Specifically, in the embodiment of the present invention, since the first model is the model corresponding to the maximum feature weight of the first feature, and the second model is the model corresponding to the minimum feature weight of the first feature, the difference between the first model and the second model is the greatest for the first feature. The second sub-training sample is a training sample included in both the first training sample corresponding to the first model and the second training sample corresponding to the second model. Therefore, it is considered that there are no abnormal samples in the second sub-training sample. If there are abnormal samples in the sample data set, the abnormal samples exist in the first sub-training sample. In order to determine the abnormal samples, the first sub-training sample is divided into two parts, and is combined with the second sub-training sample respectively to obtain the first target training sample and the second target training sample, and then model training is performed based on the first target training sample and the second target training sample respectively to obtain the first target model and the second target model.

[0038] Among them, in the embodiment of the present invention, when the first sub-training sample is divided into two parts, the division can be arbitrarily, but in order to facilitate the subsequent judgment of abnormal samples, if the number of samples contained in the first sub-training sample is an even number, the first sub-training sample is divided into two parts with the same number of samples contained, and if the number of samples contained in the first sub-training sample is an odd number, the first sub-training sample is divided into two parts with a difference of one in the number of samples contained.

[0039] S104: Calculate the second coefficient of variation of the first feature based on the at least one model and the first target model, and calculate the third coefficient of variation of the first feature based on the at least one model and the second target model. If the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, determine that there are no abnormal samples.

[0040] In an embodiment of the present invention, after the first target model and the second target model are trained, abnormal samples can be determined based on the first target model and the second target model. Specifically, the second coefficient of variation of the first feature is calculated based on at least one previously trained model and the first target model, and the third coefficient of variation of the first feature is calculated based on at least one previously trained model and the second target model, and the abnormal samples are determined based on the numerical relationship between the second coefficient of variation and the first coefficient of variation, and the numerical relationship between the third coefficient of variation and the first coefficient of variation. Specifically, if the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, it is determined that there are no abnormal samples. If at least one of the second coefficient of variation and the third coefficient of variation is not smaller than the first coefficient of variation, the range of abnormal samples will continue to be narrowed.

[0041] In an embodiment of the present invention, a first feature corresponding to a first coefficient of variation with a maximum value, a first training sample of a first model corresponding to a maximum feature weight of the first feature, and a second training sample of a second model corresponding to a minimum feature weight of the first feature are determined; based on the first training sample and the second training sample, a first target training sample and a second target training sample are obtained, and model training is performed; a second coefficient of variation of the first feature is calculated according to at least one model and the first target model; and a third coefficient of variation of the first feature is calculated according to at least one model and the second target model; if both the second coefficient of variation and the third coefficient of variation are smaller than the first coefficient of variation, it is determined that no abnormal sample exists, thereby realizing the judgment of whether there are abnormal samples in the sample data set.

[0042] Embodiment 2:

[0043] In order to calculate the coefficient of variation of each feature in the trained models with the same structure, based on the above embodiment, in the embodiment of the present invention, the coefficient of variation of each feature is calculated according to the feature weight corresponding to each feature in each model, including:

[0044] Determine the average value and standard deviation of the feature weight of each feature according to the feature weight corresponding to each feature in each model;

[0045] The ratio of the standard deviation of the feature weight of each feature to the mean value is determined as the coefficient of variation of the feature.

[0046] In an embodiment of the present invention, for each feature, the feature weight of the feature in each model may be the same or different. When calculating the coefficient of variation of the feature, it is necessary to use the mean value and standard deviation of the feature weight corresponding to the feature, and then determine the ratio of the standard deviation to the mean value as the coefficient of variation of the feature.

[0047] Specifically, the average value of the feature weight of each feature can be calculated using the following formula:

[0048]

[0049] Among them, Q j Represents the feature T j The feature weight, PQ j Denotes the feature weight Q j The average value of X represents the feature weight Q j The number of models, Q ji represents the feature weight of the ith model.

[0050] Specifically, the standard deviation of the feature weight of each feature can be calculated using the following formula:

[0051]

[0052] Among them, σ j Represents the feature T j The standard deviation of the feature weights, PQ j Denotes the feature weight Q j The average value, Q ji represents the feature weight of the i-th model, and X represents the feature weight Q j , that is, the number of models.

[0053] The coefficient of variation of each feature can be calculated using the following formula:

[0054]

[0055] Among them, CV j Represents the feature T j The coefficient of variation, σ j Represents the feature T j The standard deviation of the feature weights, PQ j Denotes the feature weight Q j The average value of .

[0056] Embodiment 3:

[0057] In order to narrow the range of abnormal samples, based on the above embodiments, in an embodiment of the present invention, if at least one of the second coefficient of variation and the third coefficient of variation is not less than the first coefficient of variation, the method further includes:

[0058] If the second coefficient of variation is not less than the first coefficient of variation, then identify the largest fourth coefficient of variation among the coefficients of variation corresponding to other features except the first feature and the second feature corresponding to the fourth coefficient of variation; identify the third model corresponding to the maximum feature weight of the second feature and the fourth model corresponding to the minimum feature weight; identify the third training sample for training the third model and the third sub-training sample that is different from the fourth training sample for training the fourth model, and the fourth sub-training sample in the third training sample that is the same as the fourth training sample, divide the third sub-training sample into two parts, and combine them with the fourth sub-training sample respectively to obtain a third target training sample and a fourth target training sample, and perform model training based on the third target training sample and the fourth target training sample respectively to obtain a third target model and a fourth target model; calculate the fifth coefficient of variation of the second feature according to the at least one model and the third target model, and calculate the sixth coefficient of variation of the second feature according to the at least one model and the fourth target model, if the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, determine that there is no abnormal sample; or

[0059] If the third coefficient of variation is not less than the first coefficient of variation, then identify the largest seventh coefficient of variation among the coefficients of variation corresponding to other features except the first feature and the third feature corresponding to the seventh coefficient of variation; identify the fifth model corresponding to the maximum feature weight of the third feature and the sixth model corresponding to the minimum feature weight; identify the fifth sub-training sample that is different from the fifth training sample for training the fifth model and the sixth training sample for training the sixth model, and the sixth sub-training sample that is the same as the sixth training sample in the fifth training sample, divide the fifth sub-training sample into two parts, and combine them with the sixth sub-training sample respectively to obtain a fifth target training sample and a sixth target training sample, and perform model training based on the fifth target training sample and the sixth target training sample respectively to obtain a fifth target model and a sixth target model; calculate the eighth coefficient of variation of the third feature according to the at least one model and the fifth target model, and calculate the ninth coefficient of variation of the third feature according to the at least one model and the sixth target model, if the eighth coefficient of variation and the ninth coefficient of variation are both less than the seventh coefficient of variation, determine that there is no abnormal sample; or

[0060] If the second coefficient of variation and the third coefficient of variation are not less than the first coefficient of variation, and the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, and the eighth coefficient of variation and the ninth coefficient of variation are both less than the seventh coefficient of variation, it is determined that there is no abnormal sample.

[0061] In the embodiment of the present invention, if it is not determined that there are no abnormal samples, that is, at least one of the second coefficient of variation and the third coefficient of variation is not less than the first coefficient of variation, then the range of the coefficient of variation needs to be further narrowed. At this time, there are three possible situations, one is that the second coefficient of variation is not less than the first coefficient of variation, one is that the third coefficient of variation is not less than the first coefficient of variation, and the other is that both the second coefficient of variation and the third coefficient of variation are not less than the first coefficient of variation. For each situation, the range of abnormal samples obtained is different.

[0062] Specifically, if the second coefficient of variation is not less than the first coefficient of variation, then identify the fourth coefficient of variation with the largest among the coefficients of variation corresponding to features other than the first feature and the second feature corresponding to the fourth coefficient of variation, identify the third model corresponding to the maximum feature weight of the second feature and the fourth model corresponding to the minimum feature weight, identify the third training sample for training the third model and the third sub-training sample for training the fourth model that is different from the fourth training sample, and identify the fourth sub-training sample in the third training sample that is the same as the fourth training sample.

[0063] Since the third model is the model corresponding to the maximum feature weight of the second feature, and the fourth model is the model corresponding to the minimum feature weight of the second feature, in this case, the difference between the third model and the fourth model is the largest for the second feature. The fourth sub-training sample is a training sample included in both the third training sample corresponding to the third model and the fourth training sample corresponding to the fourth model. Therefore, it is considered that there is no abnormal sample in the fourth sub-training sample. If there is an abnormal sample in the sample data set, the abnormal sample exists in the third sub-training sample.

[0064] In order to determine abnormal samples, the third sub-training sample is divided into two parts, and the two parts are respectively combined with the fourth sub-training sample to obtain a third target training sample and a fourth target training sample, and model training is performed based on the third target training sample and the fourth target training sample to obtain a third target model and a fourth target model, respectively. According to the at least one model and the third target model, the fifth coefficient of variation of the second feature is calculated, and according to the at least one model and the fourth target model, the sixth coefficient of variation of the second feature is calculated. If both the fifth coefficient of variation and the sixth coefficient of variation are smaller than the fourth coefficient of variation, it is determined that there is no abnormal sample. If at least one of the fifth coefficient of variation and the sixth coefficient of variation is not smaller than the fourth coefficient of variation, the range of abnormal samples will continue to be narrowed.

[0065] If the third coefficient of variation is not less than the first coefficient of variation, then identify the seventh coefficient of variation that is the largest among the coefficients of variation corresponding to other features except the first feature and the third feature corresponding to the seventh coefficient of variation; identify the fifth model corresponding to the maximum feature weight of the third feature and the sixth model corresponding to the minimum feature weight; identify the fifth sub-training sample that is different from the fifth training sample for training the fifth model and the sixth training sample for training the sixth model, as well as the sixth sub-training sample in the fifth training sample that is the same as the sixth training sample.

[0066] Since the fifth model is the model corresponding to the maximum feature weight of the third feature, and the sixth model is the model corresponding to the minimum feature weight of the third feature, in this case, the fifth model and the sixth model have the greatest difference for the third feature. The sixth sub-training sample is a training sample included in both the fifth training sample corresponding to the fifth model and the sixth training sample corresponding to the sixth model. Therefore, it is considered that there is no abnormal sample in the sixth sub-training sample. If there is an abnormal sample in the sample data set, the abnormal sample exists in the fifth sub-training sample.

[0067] In order to determine abnormal samples, the fifth sub-training sample is divided into two parts, and the two parts are respectively combined with the sixth sub-training sample to obtain a fifth target training sample and a sixth target training sample, and model training is performed based on the fifth target training sample and the sixth target training sample to obtain a fifth target model and a sixth target model; according to the at least one model and the fifth target model, the eighth coefficient of variation of the third feature is calculated, and according to the at least one model and the sixth target model, the ninth coefficient of variation of the third feature is calculated. If the eighth coefficient of variation and the ninth coefficient of variation are both smaller than the seventh coefficient of variation, it is determined that there is no abnormal sample. At least one of the eighth coefficient of variation and the ninth coefficient of variation is not smaller than the seventh coefficient of variation, which will continue to narrow the range of abnormal samples.

[0068] If the second coefficient of variation and the third coefficient of variation are not less than the first coefficient of variation, then determine whether the fifth coefficient of variation and the sixth coefficient of variation are less than the fourth coefficient of variation, and whether the eighth coefficient of variation and the ninth coefficient of variation are less than the seventh coefficient of variation. If it is determined that the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, and the eighth coefficient of variation and the ninth coefficient of variation are both less than the seventh coefficient of variation, then determine that there is no abnormal sample. Otherwise, continue to narrow the range of abnormal samples.

[0069] Embodiment 4:

[0070] In order to narrow the range of abnormal samples, on the basis of the above embodiments, in an embodiment of the present invention, if at least one of the fifth coefficient of variation and the sixth coefficient of variation is not less than the fourth coefficient of variation, the method further includes:

[0071] If the fifth coefficient of variation is not less than the fourth coefficient of variation, then identify the largest tenth coefficient of variation among the coefficients of variation corresponding to other features except the second feature and the fourth feature corresponding to the tenth coefficient of variation; identify the seventh model corresponding to the maximum feature weight of the fourth feature and the eighth model corresponding to the minimum feature weight; identify the seventh sub-training sample that is different from the seventh training sample for training the seventh model and the eighth training sample for training the eighth model, and the eighth sub-training sample that is the same as the eighth training sample in the seventh training sample, divide the seventh sub-training sample into two parts, and combine them with the eighth sub-training sample respectively to obtain a seventh target training sample and an eighth target training sample, and perform model training based on the seventh target training sample and the eighth target training sample respectively to obtain a seventh target model and an eighth target model; calculate the eleventh coefficient of variation of the fourth feature according to the at least one model and the seventh target model, and calculate the twelfth coefficient of variation of the fourth feature according to the at least one model and the eighth target model, if the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, determine that there is no abnormal sample; or

[0072] If the sixth coefficient of variation is not less than the fourth coefficient of variation, then identify the largest thirteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the fifth feature corresponding to the thirteenth coefficient of variation; identify the ninth model corresponding to the maximum feature weight of the fifth feature and the tenth model corresponding to the minimum feature weight; identify the ninth sub-training sample that is different from the ninth training sample for training the ninth model and the tenth training sample for training the tenth model, and the tenth sub-training sample that is the same as the tenth training sample in the ninth training sample, divide the ninth sub-training sample into two parts, and combine them with the tenth sub-training sample respectively to obtain a ninth target training sample and a tenth target training sample, and perform model training based on the ninth target training sample and the tenth target training sample respectively to obtain a ninth target model and a tenth target model; calculate the fourteenth coefficient of variation of the fifth feature according to the at least one model and the ninth target model, and calculate the fifteenth coefficient of variation of the fifth feature according to the at least one model and the tenth target model, if the fourteenth coefficient of variation and the fifteenth coefficient of variation are both less than the thirteenth coefficient of variation, determine that there is no abnormal sample; or

[0073] If the fifth coefficient of variation and the sixth coefficient of variation are not less than the fourth coefficient of variation, and the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, and the fourteenth coefficient of variation and the fifteenth coefficient of variation are both less than the thirteenth coefficient of variation, it is determined that there is no abnormal sample.

[0074] In the embodiment of the present invention, if it is not determined that there are no abnormal samples, that is, at least one of the fifth coefficient of variation and the sixth coefficient of variation is not less than the fourth coefficient of variation, then the range of the coefficient of variation needs to be further narrowed. At this time, there are three possible situations, one is that the fifth coefficient of variation is not less than the fourth coefficient of variation, one is that the sixth coefficient of variation is not less than the fourth coefficient of variation, and the other is that both the fifth coefficient of variation and the sixth coefficient of variation are not less than the fourth coefficient of variation. For each situation, the range of abnormal samples obtained is different.

[0075] Specifically, if the fifth coefficient of variation is not less than the fourth coefficient of variation, then identify the tenth coefficient of variation that is the largest among the coefficients of variation corresponding to other features except the second feature and the fourth feature corresponding to the tenth coefficient of variation; identify the seventh model corresponding to the maximum feature weight of the fourth feature and the eighth model corresponding to the minimum feature weight; identify the seventh sub-training sample that is different from the seventh training sample for training the seventh model and the eighth training sample for training the eighth model, as well as the eighth sub-training sample in the seventh training sample that is the same as the eighth training sample.

[0076] Since the seventh model is the model corresponding to the maximum feature weight of the fourth feature, and the eighth model is the model corresponding to the minimum feature weight of the fourth feature, in this case, the seventh model and the eighth model have the greatest difference for the fourth feature. The eighth sub-training sample is a training sample included in both the seventh training sample corresponding to the seventh model and the eighth training sample corresponding to the eighth model. Therefore, it is considered that there is no abnormal sample in the eighth sub-training sample. If there is an abnormal sample in the sample data set, the abnormal sample exists in the seventh sub-training sample.

[0077] In order to determine abnormal samples, the seventh sub-training sample is divided into two parts, and the two parts are respectively combined with the eighth sub-training sample to obtain a seventh target training sample and an eighth target training sample, and model training is performed based on the seventh target training sample and the eighth target training sample to obtain a seventh target model and an eighth target model; the eleventh coefficient of variation of the fourth feature is calculated according to the at least one model and the seventh target model, and the twelfth coefficient of variation of the fourth feature is calculated according to the at least one model and the eighth target model. If the eleventh coefficient of variation and the twelfth coefficient of variation are both smaller than the tenth coefficient of variation, it is determined that there is no abnormal sample. If at least one of the eleventh coefficient of variation and the twelfth coefficient of variation is smaller than the tenth coefficient of variation, the range of abnormal samples will continue to be narrowed.

[0078] If the sixth coefficient of variation is not less than the fourth coefficient of variation, then identify the thirteenth coefficient of variation that is the largest among the coefficients of variation corresponding to the other features except the second feature and the fifth feature corresponding to the thirteenth coefficient of variation; identify the ninth model corresponding to the maximum feature weight of the fifth feature and the tenth model corresponding to the minimum feature weight; identify the ninth sub-training sample that is different from the ninth training sample for training the ninth model and the tenth training sample for training the tenth model, as well as the tenth sub-training sample in the ninth training sample that is the same as the tenth training sample.

[0079] Since the ninth model is the model corresponding to the maximum feature weight of the fifth feature, and the tenth model is the model corresponding to the minimum feature weight of the fifth feature, in this case, the ninth model and the tenth model have the greatest difference for the fifth feature. The tenth sub-training sample is a training sample included in both the ninth training sample corresponding to the ninth model and the tenth training sample corresponding to the tenth model. Therefore, it is considered that there is no abnormal sample in the tenth sub-training sample. If there is an abnormal sample in the sample data set, the abnormal sample exists in the ninth sub-training sample.

[0080] In order to determine abnormal samples, the ninth sub-training sample is divided into two parts, and the two parts are respectively combined with the tenth sub-training sample to obtain a ninth target training sample and a tenth target training sample, and model training is performed based on the ninth target training sample and the tenth target training sample to obtain a ninth target model and a tenth target model; the fourteenth coefficient of variation of the fifth feature is calculated according to the at least one model and the ninth target model, and the fifteenth coefficient of variation of the fifth feature is calculated according to the at least one model and the tenth target model. If the fourteenth coefficient of variation and the fifteenth coefficient of variation are both smaller than the thirteenth coefficient of variation, it is determined that there is no abnormal sample. If at least one of the fourteenth coefficient of variation and the fifteenth coefficient of variation is not smaller than the thirteenth coefficient of variation, the range of abnormal samples will continue to be narrowed.

[0081] If the fifth coefficient of variation and the sixth coefficient of variation are not less than the fourth coefficient of variation, then determine whether the eleventh coefficient of variation and the twelfth coefficient of variation are less than the tenth coefficient of variation, and determine whether the fourteenth coefficient of variation and the fifteenth coefficient of variation are less than the thirteenth coefficient of variation. If it is determined that the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, and the fourteenth coefficient of variation and the fifteenth coefficient of variation are both less than the thirteenth coefficient of variation, then it is determined that there is no abnormal sample. Otherwise, the range of abnormal samples will continue to be narrowed.

[0082] Embodiment 5:

[0083] In order to determine abnormal samples, based on the above embodiments, in an embodiment of the present invention, the method further includes:

[0084] If the eleventh coefficient of variation is not less than the tenth coefficient of variation, determining the other training samples in the seventh target training sample except the eighth sub-training sample as abnormal samples; or

[0085] If the twelfth coefficient of variation is not less than the tenth coefficient of variation, determining other training samples in the eighth target training sample except the eighth sub-training sample as abnormal samples; or

[0086] If the eleventh coefficient of variation and the twelfth coefficient of variation are both not less than the tenth coefficient of variation, the seventh sub-training sample is determined as an abnormal sample.

[0087] In the embodiment of the present invention, if at least one of the eleventh coefficient of variation and the twelfth coefficient of variation is not less than the tenth coefficient of variation, the abnormal sample can be determined based on the eleventh coefficient of variation and the twelfth coefficient of variation. At this time, there are three possible situations, one is that the eleventh coefficient of variation is not less than the tenth coefficient of variation, one is that the twelfth coefficient of variation is not less than the tenth coefficient of variation, and the other is that both the eleventh coefficient of variation and the twelfth coefficient of variation are not less than the tenth coefficient of variation.

[0088] Specifically, if the eleventh coefficient of variation is not less than the tenth coefficient of variation, the other training samples in the seventh target training sample except the eighth sub-training sample are determined as abnormal samples. If the twelfth coefficient of variation is not less than the tenth coefficient of variation, the other training samples in the eighth target training sample except the eighth sub-training sample are determined as abnormal samples. If both the eleventh coefficient of variation and the twelfth coefficient of variation are not less than the tenth coefficient of variation, the seventh sub-training sample is determined as an abnormal sample.

[0089] Embodiment 6:

[0090] In order to determine abnormal samples, based on the above embodiments, in an embodiment of the present invention, the method further includes:

[0091] If the fourteenth coefficient of variation is not less than the thirteenth coefficient of variation, determining the other training samples in the ninth target training sample except the tenth sub-training sample as abnormal samples; or

[0092] If the fifteenth coefficient of variation is not less than the thirteenth coefficient of variation, determining the other training samples in the tenth target training sample except the tenth sub-training sample as abnormal samples; or

[0093] If the fourteenth coefficient of variation and the fifteenth coefficient of variation are both not less than the thirteenth coefficient of variation, the ninth sub-training sample is determined as an abnormal sample.

[0094] In the embodiment of the present invention, if at least one of the fourteenth coefficient of variation and the fifteenth coefficient of variation is not less than the thirteenth coefficient of variation, then the abnormal sample can be determined based on the fourteenth coefficient of variation and the fifteenth coefficient of variation. At this time, there are three possible situations, one is that the fourteenth coefficient of variation is not less than the thirteenth coefficient of variation, one is that the fifteenth coefficient of variation is not less than the thirteenth coefficient of variation, and the other is that both the fourteenth coefficient of variation and the fifteenth coefficient of variation are not less than the thirteenth coefficient of variation.

[0095] Specifically, if the fourteenth coefficient of variation is not less than the thirteenth coefficient of variation, the other training samples in the ninth target training sample except the tenth sub-training sample are determined as abnormal samples. If the fifteenth coefficient of variation is not less than the thirteenth coefficient of variation, the other training samples in the tenth target training sample except the tenth sub-training sample are determined as abnormal samples. If both the fourteenth coefficient of variation and the fifteenth coefficient of variation are not less than the thirteenth coefficient of variation, the ninth sub-training sample is determined as an abnormal sample.

[0096] Embodiment 7:

[0097] In order to narrow the range of abnormal samples, based on the above embodiments, in an embodiment of the present invention, if at least one of the eleventh coefficient of variation and the twelfth coefficient of variation is not less than the tenth coefficient of variation, the method further includes:

[0098] If the eighth coefficient of variation is not less than the seventh coefficient of variation, then identify the largest sixteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the sixth feature corresponding to the sixteenth coefficient of variation; identify the eleventh model corresponding to the maximum feature weight of the sixth feature and the twelfth model corresponding to the minimum feature weight; identify the eleventh sub-training sample that is different from the eleventh training sample for training the eleventh model and the twelfth training sample for training the twelfth model, and the twelfth sub-training sample that is the same as the twelfth training sample in the eleventh training sample, divide the eleventh sub-training sample into two parts, and combine them with the twelfth sub-training sample respectively to obtain the eleventh target training sample and the twelfth target training sample, and perform model training based on the eleventh target training sample and the twelfth target training sample respectively to obtain the eleventh target model and the twelfth target model; calculate the seventeenth coefficient of variation of the sixth feature according to the at least one model and the eleventh target model, and calculate the eighteenth coefficient of variation of the sixth feature according to the at least one model and the twelfth target model, if the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, determine that there is no abnormal sample; or

[0099] If the ninth coefficient of variation is not less than the seventh coefficient of variation, then identify the largest nineteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the seventh feature corresponding to the nineteenth coefficient of variation; identify the thirteenth model corresponding to the maximum feature weight of the seventh feature and the fourteenth model corresponding to the minimum feature weight; identify the thirteenth sub-training sample that is different from the thirteenth training sample for training the thirteenth model and the fourteenth training sample for training the fourteenth model, and the fourteenth sub-training sample that is the same as the fourteenth training sample in the thirteenth training sample, divide the thirteenth sub-training sample into two parts, and combine them with the fourteenth sub-training sample respectively to obtain a thirteenth target training sample and a fourteenth target training sample, and perform model training based on the thirteenth target training sample and the fourteenth target training sample respectively to obtain a thirteenth target model and a fourteenth target model; calculate the twentieth coefficient of variation of the sixth feature according to the at least one model and the thirteenth target model, and calculate the twenty-first coefficient of variation of the sixth feature according to the at least one model and the fourteenth target model, if the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, then determine that there is no abnormal sample; or

[0100] If the eighth coefficient of variation and the ninth coefficient of variation are not less than the seventh coefficient of variation, and the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, and the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, it is determined that there are no abnormal samples.

[0101] In the embodiment of the present invention, if it is not determined that there are no abnormal samples, that is, at least one of the eighth coefficient of variation and the ninth coefficient of variation is not less than the seventh coefficient of variation, then the range of the coefficient of variation needs to be further narrowed. At this time, there are three possible situations, one is that the eighth coefficient of variation is not less than the seventh coefficient of variation, one is that the ninth coefficient of variation is not less than the seventh coefficient of variation, and the other is that both the eighth coefficient of variation and the ninth coefficient of variation are not less than the seventh coefficient of variation. For each situation, the range of abnormal samples obtained is different.

[0102] Specifically, if the eighth coefficient of variation is not less than the seventh coefficient of variation, then identify the sixteenth coefficient of variation that is the largest among the coefficients of variation corresponding to other features except the second feature and the sixth feature corresponding to the sixteenth coefficient of variation; identify the eleventh model corresponding to the maximum feature weight of the sixth feature and the twelfth model corresponding to the minimum feature weight; identify the eleventh sub-training sample that is different from the eleventh training sample for training the eleventh model and the twelfth training sample for training the twelfth model, as well as the twelfth sub-training sample in the eleventh training sample that is the same as the twelfth training sample.

[0103] Since the eleventh model is the model corresponding to the maximum feature weight of the sixth feature, and the twelfth model is the model corresponding to the minimum feature weight of the sixth feature, in this case, the difference between the eleventh model and the twelfth model is the greatest for the sixth feature. The twelfth sub-training sample is a training sample included in both the eleventh training sample corresponding to the eleventh model and the twelfth training sample corresponding to the twelfth model. Therefore, it is considered that there is no abnormal sample in the twelfth sub-training sample. If there is an abnormal sample in the sample data set, the abnormal sample exists in the eleventh sub-training sample.

[0104] In order to determine abnormal samples, the eleventh sub-training sample is divided into two parts, and the two parts are respectively combined with the twelfth sub-training sample to obtain an eleventh target training sample and a twelfth target training sample, and model training is performed based on the eleventh target training sample and the twelfth target training sample to obtain an eleventh target model and a twelfth target model; the seventeenth coefficient of variation of the sixth feature is calculated according to the at least one model and the eleventh target model, and the eighteenth coefficient of variation of the sixth feature is calculated according to the at least one model and the twelfth target model. If the seventeenth coefficient of variation and the eighteenth coefficient of variation are both smaller than the sixteenth coefficient of variation, it is determined that there is no abnormal sample. If at least one of the seventeenth coefficient of variation and the eighteenth coefficient of variation is not smaller than the sixteenth coefficient of variation, the range of abnormal samples will continue to be narrowed.

[0105] If the ninth coefficient of variation is not less than the seventh coefficient of variation, then identify the nineteenth coefficient of variation that is the largest among the coefficients of variation corresponding to other features except the second feature and the seventh feature corresponding to the nineteenth coefficient of variation; identify the thirteenth model corresponding to the maximum feature weight of the seventh feature and the fourteenth model corresponding to the minimum feature weight; identify the thirteenth sub-training sample that is different from the thirteenth training sample used to train the thirteenth model and the fourteenth training sample used to train the fourteenth model, as well as the fourteenth sub-training sample in the thirteenth training sample that is the same as the fourteenth training sample.

[0106] Since the thirteenth model is the model corresponding to the maximum feature weight of the sixth feature, and the fourteenth model is the model corresponding to the minimum feature weight of the sixth feature, in this case, the thirteenth model and the fourteenth model have the greatest difference for the sixth feature. The fourteenth sub-training sample is a training sample included in both the thirteenth training sample corresponding to the thirteenth model and the fourteenth training sample corresponding to the fourteenth model. Therefore, it is considered that there is no abnormal sample in the fourteenth sub-training sample. If there is an abnormal sample in the sample data set, the abnormal sample exists in the thirteenth sub-training sample.

[0107] In order to determine the abnormal samples, the thirteenth sub-training sample is divided into two parts, and the two parts are respectively combined with the fourteenth sub-training sample to obtain the thirteenth target training sample and the fourteenth target training sample, and model training is performed based on the thirteenth target training sample and the fourteenth target training sample to obtain the thirteenth target model and the fourteenth target model; the twentieth coefficient of variation of the sixth feature is calculated according to the at least one model and the thirteenth target model, and the twenty-first coefficient of variation of the sixth feature is calculated according to the at least one model and the fourteenth target model. If the twentieth coefficient of variation and the twenty-first coefficient of variation are both smaller than the nineteenth coefficient of variation, it is determined that there is no abnormal sample. If at least one of the twentieth coefficient of variation and the twenty-first coefficient of variation is not smaller than the nineteenth coefficient of variation, the range of abnormal samples will continue to be narrowed.

[0108] If the eighth and ninth coefficients of variation are not less than the seventh coefficient of variation, then determine whether the seventeenth and eighteenth coefficients of variation are less than the sixteenth coefficient of variation, and determine whether the twentieth and twenty-first coefficients of variation are less than the nineteenth coefficient of variation. If it is determined that the seventeenth and eighteenth coefficients of variation are both less than the sixteenth coefficient of variation, and the twentieth and twenty-first coefficients of variation are both less than the nineteenth coefficient of variation, then determine that there is no abnormal sample. Otherwise, the range of abnormal samples will continue to be narrowed.

[0109] Embodiment 8:

[0110] In order to determine abnormal samples, based on the above embodiments, in an embodiment of the present invention, the method further includes:

[0111] If the seventeenth coefficient of variation is not less than the sixteenth coefficient of variation, determining the other training samples in the eleventh target training sample except the twelfth sub-training sample as abnormal samples; or

[0112] If the eighteenth coefficient of variation is not less than the sixteenth coefficient of variation, determining other training samples in the twelfth target training sample except the twelfth sub-training sample as abnormal samples; or

[0113] If the seventeenth coefficient of variation and the eighteenth coefficient of variation are both not less than the sixteenth coefficient of variation, the eleventh sub-training sample is determined as an abnormal sample.

[0114] In the embodiment of the present invention, if at least one of the seventeenth coefficient of variation and the eighteenth coefficient of variation is not less than the sixteenth coefficient of variation, the abnormal sample can be determined based on the seventeenth coefficient of variation and the eighteenth coefficient of variation. At this time, there are three possible situations, one is that the seventeenth coefficient of variation is not less than the sixteenth coefficient of variation, one is that the eighteenth coefficient of variation is not less than the sixteenth coefficient of variation, and the other is that both the seventeenth coefficient of variation and the eighteenth coefficient of variation are not less than the sixteenth coefficient of variation.

[0115] Specifically, if the seventeenth coefficient of variation is not less than the sixteenth coefficient of variation, the other training samples in the eleventh target training sample except the twelfth sub-training sample are determined as abnormal samples. If the eighteenth coefficient of variation is not less than the sixteenth coefficient of variation, the other training samples in the twelfth target training sample except the twelfth sub-training sample are determined as abnormal samples. If both the seventeenth coefficient of variation and the eighteenth coefficient of variation are not less than the sixteenth coefficient of variation, the eleventh sub-training sample is determined as an abnormal sample.

[0116] Embodiment 9:

[0117] In order to determine abnormal samples, based on the above embodiments, in an embodiment of the present invention, the method further includes:

[0118] If the twentieth coefficient of variation is not less than the nineteenth coefficient of variation, determining the other training samples in the thirteenth target training sample except the fourteenth sub-training sample as abnormal samples; or

[0119] If the twenty-first coefficient of variation is not less than the nineteenth coefficient of variation, determining the other training samples in the fourteenth target training sample except the fourteenth sub-training sample as abnormal samples; or

[0120] If the 20th coefficient of variation and the 21st coefficient of variation are not less than the 19th coefficient of variation, the 13th sub-training sample is determined as an abnormal sample.

[0121] In the embodiment of the present invention, if at least one of the 20th coefficient of variation and the 21st coefficient of variation is not less than the 19th coefficient of variation, the abnormal sample can be determined based on the 20th coefficient of variation and the 21st coefficient of variation. At this time, there are three possible situations, one is that the 20th coefficient of variation is not less than the 19th coefficient of variation, one is that the 21st coefficient of variation is not less than the 19th coefficient of variation, and the other is that both the 20th coefficient of variation and the 21st coefficient of variation are not less than the 19th coefficient of variation.

[0122] Specifically, if the twentieth coefficient of variation is not less than the nineteenth coefficient of variation, the other training samples in the thirteenth target training sample except the fourteenth sub-training sample are determined as abnormal samples. If the twenty-first coefficient of variation is not less than the nineteenth coefficient of variation, the other training samples in the fourteenth target training sample except the fourteenth sub-training sample are determined as abnormal samples. If both the twentieth coefficient of variation and the twenty-first coefficient of variation are not less than the nineteenth coefficient of variation, the thirteenth sub-training sample is determined as an abnormal sample.

[0123] Embodiment 10:

[0124] In order to determine the optimal model among at least one model trained based on a sample data set, based on the above embodiments, in an embodiment of the present invention, the method further includes:

[0125] Based on the feature weights of each feature in each model, calculate the average difference in the feature weights of the features of each model;

[0126] The minimum mean difference with the smallest value among the mean differences is identified, and the model corresponding to the minimum mean difference is determined as the optimal model.

[0127] In an embodiment of the present invention, if it is determined that there are no abnormal samples in the sample data set, the optimal model can be selected from at least one model trained in the sample data set. Specifically, for each model, the average difference of the feature weights of all features in the model can be calculated, and the model corresponding to the minimum average difference is the optimal model.

[0128] Specifically, the average difference of the feature weights of each model can be calculated using the following formula:

[0129]

[0130] Among them, Q ji represents the feature weight of feature Tj in the i-th model, PQ j Denotes the feature weight Q j The average value of n represents the feature T j The number of MD ji represents the mean difference in feature weights of the ith model.

[0131] Embodiment 11:

[0132] Figure 2 A schematic diagram of the structure of an abnormal sample determination device provided by an embodiment of the present invention is shown in FIG. Figure 2 As shown, the device comprises:

[0133] A calculation module 201 is used to calculate the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model for at least one trained model, and identify a first feature corresponding to a first coefficient of variation with a maximum value;

[0134] An identification module 202, configured to identify a first model corresponding to a maximum feature weight of the first feature, and a second model corresponding to a minimum feature weight;

[0135] The abnormal sample determination module 203 is used to identify different first sub-training samples in the first training sample for training the first model and the second training sample for training the second model, and the second sub-training sample in the first training sample that is the same as the second training sample, divide the first sub-training sample into two parts, and combine them with the second sub-training sample respectively to obtain a first target training sample and a second target training sample, and perform model training based on the first target training sample and the second target training sample respectively to obtain a first target model and a second target model; calculate the second coefficient of variation of the first feature according to the at least one model and the first target model, and calculate the third coefficient of variation of the first feature according to the at least one model and the second target model; if the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, it is determined that there is no abnormal sample.

[0136] In a possible implementation, the calculation module 201 is specifically used to determine the mean value and standard deviation of the feature weight of each feature according to the feature weight corresponding to each feature in each model; and determine the ratio of the standard deviation of the feature weight of each feature to the mean value as the coefficient of variation of the feature.

[0137] In a possible implementation, the abnormal sample determination module 203 is further used to, if the second coefficient of variation is not less than the first coefficient of variation, identify the fourth coefficient of variation with the largest value among the coefficients of variation corresponding to other features except the first feature and the second feature corresponding to the fourth coefficient of variation; identify the third model corresponding to the maximum feature weight of the second feature and the fourth model corresponding to the minimum feature weight; identify the third training sample for training the third model and the fourth training sample for training the fourth model that are different from the third sub-training sample, and the fourth sub-training sample in the third training sample that is the same as the fourth training sample, and divide the third sub-training sample into two parts. and respectively combine with the fourth sub-training sample to obtain a third target training sample and a fourth target training sample, and respectively perform model training based on the third target training sample and the fourth target training sample to obtain a third target model and a fourth target model; calculate the fifth coefficient of variation of the second feature according to the at least one model and the third target model, and calculate the sixth coefficient of variation of the second feature according to the at least one model and the fourth target model, if the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, it is determined that there is no abnormal sample; or if the third coefficient of variation is not less than the first coefficient of variation, identify other features except the first feature The seventh coefficient of variation with the largest among the coefficients of variation corresponding to the features and the third feature corresponding to the seventh coefficient of variation; identifying the fifth model corresponding to the maximum feature weight of the third feature and the sixth model corresponding to the minimum feature weight; identifying the fifth sub-training samples that are different from the fifth training sample for training the fifth model and the sixth training sample for training the sixth model, and the sixth sub-training sample that is the same as the sixth training sample in the fifth training sample, dividing the fifth sub-training sample into two parts, and combining them with the sixth sub-training sample respectively to obtain the fifth target training sample and the sixth target training sample, and performing Model training is performed to obtain a fifth target model and a sixth target model; an eighth coefficient of variation of the third feature is calculated based on the at least one model and the fifth target model, and a ninth coefficient of variation of the third feature is calculated based on the at least one model and the sixth target model; if the eighth coefficient of variation and the ninth coefficient of variation are both smaller than the seventh coefficient of variation, it is determined that no abnormal sample exists; or if the second coefficient of variation and the third coefficient of variation are both not smaller than the first coefficient of variation, and the fifth coefficient of variation and the sixth coefficient of variation are both smaller than the fourth coefficient of variation, and the eighth coefficient of variation and the ninth coefficient of variation are both smaller than the seventh coefficient of variation, it is determined that no abnormal sample exists.

[0138] In a possible implementation, the abnormal sample determination module 203 is further used to, if the fifth coefficient of variation is not less than the fourth coefficient of variation, identify the tenth coefficient of variation that is the largest among the coefficients of variation corresponding to other features except the second feature and the fourth feature corresponding to the tenth coefficient of variation; identify the seventh model corresponding to the maximum feature weight of the fourth feature and the eighth model corresponding to the minimum feature weight; identify the seventh sub-training sample that is different from the seventh training sample for training the seventh model and the eighth training sample for training the eighth model, as well as the eighth sub-training sample in the seventh training sample that is the same as the eighth training sample, divide the seventh sub-training sample into two parts, and respectively identify the seventh sub-training sample with the seventh training sample. The eighth sub-training sample is combined to obtain a seventh target training sample and an eighth target training sample, and model training is performed based on the seventh target training sample and the eighth target training sample respectively to obtain a seventh target model and an eighth target model; the eleventh coefficient of variation of the fourth feature is calculated according to the at least one model and the seventh target model, and the twelfth coefficient of variation of the fourth feature is calculated according to the at least one model and the eighth target model, if the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, it is determined that there is no abnormal sample; or if the sixth coefficient of variation is not less than the fourth coefficient of variation, then the corresponding features other than the second feature are identified The thirteenth largest coefficient of variation among the coefficients of variation and the fifth feature corresponding to the thirteenth coefficient of variation; identifying the ninth model corresponding to the maximum feature weight of the fifth feature and the tenth model corresponding to the minimum feature weight; identifying the ninth sub-training sample that is different from the ninth training sample for training the ninth model and the tenth training sample for training the tenth model, and the tenth sub-training sample that is the same as the tenth training sample in the ninth training sample, dividing the ninth sub-training sample into two parts, and combining them with the tenth sub-training sample respectively to obtain a ninth target training sample and a tenth target training sample, and performing model training based on the ninth target training sample and the tenth target training sample respectively, to obtain to the ninth target model and the tenth target model; according to the at least one model and the ninth target model, the fourteenth coefficient of variation of the fifth feature is calculated, and according to the at least one model and the tenth target model, the fifteenth coefficient of variation of the fifth feature is calculated; if the fourteenth coefficient of variation and the fifteenth coefficient of variation are both smaller than the thirteenth coefficient of variation, it is determined that there are no abnormal samples; or if the fifth coefficient of variation and the sixth coefficient of variation are both not smaller than the fourth coefficient of variation, and the eleventh coefficient of variation and the twelfth coefficient of variation are both smaller than the tenth coefficient of variation, and the fourteenth coefficient of variation and the fifteenth coefficient of variation are both smaller than the thirteenth coefficient of variation, it is determined that there are no abnormal samples.

[0139] In a possible implementation, the abnormal sample determination module 203 is further used to determine the other training samples in the seventh target training sample except the eighth sub-training sample as abnormal samples if the eleventh coefficient of variation is not less than the tenth coefficient of variation; or to determine the other training samples in the eighth target training sample except the eighth sub-training sample as abnormal samples if the twelfth coefficient of variation is not less than the tenth coefficient of variation; or to determine the seventh sub-training sample as an abnormal sample if both the eleventh coefficient of variation and the twelfth coefficient of variation are not less than the tenth coefficient of variation.

[0140] In a possible implementation, the abnormal sample determination module 203 is further used to determine the other training samples in the ninth target training sample except the tenth sub-training sample as abnormal samples if the fourteenth coefficient of variation is not less than the thirteenth coefficient of variation; or to determine the other training samples in the tenth target training sample except the tenth sub-training sample as abnormal samples if the fifteenth coefficient of variation is not less than the thirteenth coefficient of variation; or to determine the ninth sub-training sample as an abnormal sample if both the fourteenth coefficient of variation and the fifteenth coefficient of variation are not less than the thirteenth coefficient of variation.

[0141] In a possible implementation, the abnormal sample determination module 203 is further used to, if the eighth coefficient of variation is not less than the seventh coefficient of variation, identify the sixteenth coefficient of variation that is the largest among the coefficients of variation corresponding to other features except the second feature and the sixth feature corresponding to the sixteenth coefficient of variation; identify the eleventh model corresponding to the maximum feature weight of the sixth feature and the twelfth model corresponding to the minimum feature weight; identify the eleventh sub-training sample that is different from the eleventh training sample for training the eleventh model and the twelfth training sample for training the twelfth model, and the eleventh sub-training sample that is the same as the twelfth training sample in the eleventh training sample. The method comprises the following steps: dividing the eleventh sub-training sample into two parts, and combining the two parts with the twelfth sub-training sample to obtain an eleventh target training sample and a twelfth target training sample, and performing model training based on the eleventh target training sample and the twelfth target training sample to obtain an eleventh target model and a twelfth target model; calculating the seventeenth coefficient of variation of the sixth feature according to the at least one model and the eleventh target model, and calculating the eighteenth coefficient of variation of the sixth feature according to the at least one model and the twelfth target model; if the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, number, it is determined that there is no abnormal sample; or if the ninth coefficient of variation is not less than the seventh coefficient of variation, the nineteenth coefficient of variation, which is the largest among the coefficients of variation corresponding to the other features except the second feature, and the seventh feature corresponding to the nineteenth coefficient of variation are identified; the thirteenth model corresponding to the maximum feature weight of the seventh feature and the fourteenth model corresponding to the minimum feature weight are identified; the thirteenth sub-training sample that is different from the thirteenth training sample for training the thirteenth model and the fourteenth training sample for training the fourteenth model are identified, and the fourteenth sub-training sample of the thirteenth training sample that is the same as the fourteenth training sample is identified, and the thirteenth sub-training sample is the same as the fourteenth training sample, and the The training sample is divided into two parts, and the two parts are respectively combined with the fourteenth sub-training sample to obtain a thirteenth target training sample and a fourteenth target training sample, and model training is performed based on the thirteenth target training sample and the fourteenth target training sample to obtain a thirteenth target model and a fourteenth target model; the twentieth coefficient of variation of the sixth feature is calculated according to the at least one model and the thirteenth target model, and the twenty-first coefficient of variation of the sixth feature is calculated according to the at least one model and the fourteenth target model, and if the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, it is determined that there is no abnormal sample;Or if the eighth coefficient of variation and the ninth coefficient of variation are not less than the seventh coefficient of variation, and the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, and the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, it is determined that there is no abnormal sample. ;

[0142] In a possible implementation, the abnormal sample determination module 203 is further configured to determine, if the seventeenth coefficient of variation is not less than the sixteenth coefficient of variation, the other training samples in the eleventh target training sample except the twelfth sub-training sample as abnormal samples; or if the eighteenth coefficient of variation is not less than the sixteenth coefficient of variation, determine the other training samples in the twelfth target training sample except the twelfth sub-training sample as abnormal samples; or if both the seventeenth coefficient of variation and the eighteenth coefficient of variation are not less than the sixteenth coefficient of variation, determine the eleventh sub-training sample as an abnormal sample.

[0143] In a possible implementation, the abnormal sample determination module 203 is further configured to determine, if the twentieth coefficient of variation is not less than the nineteenth coefficient of variation, the other training samples in the thirteenth target training sample except the fourteenth sub-training sample as abnormal samples; or to determine, if the twenty-first coefficient of variation is not less than the nineteenth coefficient of variation, the other training samples in the fourteenth target training sample except the fourteenth sub-training sample as abnormal samples; or to determine, if both the twentieth coefficient of variation and the twenty-first coefficient of variation are not less than the nineteenth coefficient of variation, the thirteenth sub-training sample as an abnormal sample.

[0144] In a possible implementation, the calculation module 201 is also used to calculate the average difference of the feature weights of the features of each model based on the feature weights of each feature in each model; identify the minimum average difference with the smallest value in the average differences, and determine the model corresponding to the minimum average difference as the optimal model.

[0145] Embodiment 12:

[0146] Figure 3 FIG. 1 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Based on the above embodiments, an embodiment of the present invention further provides an electronic device, such as Figure 3 As shown, it includes: a processor 301, a communication interface 302, a memory 303 and a communication bus 304, wherein the processor 301, the communication interface 302, and the memory 303 communicate with each other through the communication bus 304;

[0147] The memory 303 stores a computer program. When the program is executed by the processor 301, the processor 301 performs the following steps:

[0148] For at least one trained model, calculate the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model, and identify the first feature corresponding to the first coefficient of variation with the largest value;

[0149] Identify a first model corresponding to a maximum feature weight of the first feature, and a second model corresponding to a minimum feature weight;

[0150] Identify a first sub-training sample that is different from a first training sample for training the first model and a second training sample for training the second model, and a second sub-training sample that is the same as the second training sample in the first training sample, divide the first sub-training sample into two parts, and combine them with the second sub-training sample respectively to obtain a first target training sample and a second target training sample, and perform model training based on the first target training sample and the second target training sample respectively to obtain a first target model and a second target model;

[0151] Based on the at least one model and the first target model, the second coefficient of variation of the first feature is calculated, and based on the at least one model and the second target model, the third coefficient of variation of the first feature is calculated. If the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, it is determined that there are no abnormal samples.

[0152] In a possible implementation, calculating the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model includes:

[0153] Determine the average value and standard deviation of the feature weight of each feature according to the feature weight corresponding to each feature in each model;

[0154] The ratio of the standard deviation of the feature weight of each feature to the mean value is determined as the coefficient of variation of the feature.

[0155] In a possible implementation manner, if at least one of the second coefficient of variation and the third coefficient of variation is not less than the first coefficient of variation, the method further includes:

[0156] If the second coefficient of variation is not less than the first coefficient of variation, then identify the largest fourth coefficient of variation among the coefficients of variation corresponding to other features except the first feature and the second feature corresponding to the fourth coefficient of variation; identify the third model corresponding to the maximum feature weight of the second feature and the fourth model corresponding to the minimum feature weight; identify the third training sample for training the third model and the third sub-training sample that is different from the fourth training sample for training the fourth model, and the fourth sub-training sample in the third training sample that is the same as the fourth training sample, divide the third sub-training sample into two parts, and combine them with the fourth sub-training sample respectively to obtain a third target training sample and a fourth target training sample, and perform model training based on the third target training sample and the fourth target training sample respectively to obtain a third target model and a fourth target model; calculate the fifth coefficient of variation of the second feature according to the at least one model and the third target model, and calculate the sixth coefficient of variation of the second feature according to the at least one model and the fourth target model, if the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, determine that there is no abnormal sample; or

[0157] If the third coefficient of variation is not less than the first coefficient of variation, then identify the largest seventh coefficient of variation among the coefficients of variation corresponding to other features except the first feature and the third feature corresponding to the seventh coefficient of variation; identify the fifth model corresponding to the maximum feature weight of the third feature and the sixth model corresponding to the minimum feature weight; identify the fifth sub-training sample that is different from the fifth training sample for training the fifth model and the sixth training sample for training the sixth model, and the sixth sub-training sample that is the same as the sixth training sample in the fifth training sample, divide the fifth sub-training sample into two parts, and combine them with the sixth sub-training sample respectively to obtain a fifth target training sample and a sixth target training sample, and perform model training based on the fifth target training sample and the sixth target training sample respectively to obtain a fifth target model and a sixth target model; calculate the eighth coefficient of variation of the third feature according to the at least one model and the fifth target model, and calculate the ninth coefficient of variation of the third feature according to the at least one model and the sixth target model, if the eighth coefficient of variation and the ninth coefficient of variation are both less than the seventh coefficient of variation, determine that there is no abnormal sample; or

[0158] If the second coefficient of variation and the third coefficient of variation are not less than the first coefficient of variation, and the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, and the eighth coefficient of variation and the ninth coefficient of variation are both less than the seventh coefficient of variation, it is determined that there is no abnormal sample.

[0159] In a possible implementation manner, if at least one of the fifth coefficient of variation and the sixth coefficient of variation is not less than the fourth coefficient of variation, the method further includes:

[0160] If the fifth coefficient of variation is not less than the fourth coefficient of variation, then identify the largest tenth coefficient of variation among the coefficients of variation corresponding to other features except the second feature and the fourth feature corresponding to the tenth coefficient of variation; identify the seventh model corresponding to the maximum feature weight of the fourth feature and the eighth model corresponding to the minimum feature weight; identify the seventh sub-training sample that is different from the seventh training sample for training the seventh model and the eighth training sample for training the eighth model, and the eighth sub-training sample that is the same as the eighth training sample in the seventh training sample, divide the seventh sub-training sample into two parts, and combine them with the eighth sub-training sample respectively to obtain a seventh target training sample and an eighth target training sample, and perform model training based on the seventh target training sample and the eighth target training sample respectively to obtain a seventh target model and an eighth target model; calculate the eleventh coefficient of variation of the fourth feature according to the at least one model and the seventh target model, and calculate the twelfth coefficient of variation of the fourth feature according to the at least one model and the eighth target model, if the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, determine that there is no abnormal sample; or

[0161] If the sixth coefficient of variation is not less than the fourth coefficient of variation, then identify the largest thirteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the fifth feature corresponding to the thirteenth coefficient of variation; identify the ninth model corresponding to the maximum feature weight of the fifth feature and the tenth model corresponding to the minimum feature weight; identify the ninth sub-training sample that is different from the ninth training sample for training the ninth model and the tenth training sample for training the tenth model, and the tenth sub-training sample that is the same as the tenth training sample in the ninth training sample, divide the ninth sub-training sample into two parts, and combine them with the tenth sub-training sample respectively to obtain a ninth target training sample and a tenth target training sample, and perform model training based on the ninth target training sample and the tenth target training sample respectively to obtain a ninth target model and a tenth target model; calculate the fourteenth coefficient of variation of the fifth feature according to the at least one model and the ninth target model, and calculate the fifteenth coefficient of variation of the fifth feature according to the at least one model and the tenth target model, if the fourteenth coefficient of variation and the fifteenth coefficient of variation are both less than the thirteenth coefficient of variation, determine that there is no abnormal sample; or

[0162] If the fifth coefficient of variation and the sixth coefficient of variation are not less than the fourth coefficient of variation, and the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, and the fourteenth coefficient of variation and the fifteenth coefficient of variation are both less than the thirteenth coefficient of variation, it is determined that there is no abnormal sample.

[0163] In a possible implementation, the method further includes:

[0164] If the eleventh coefficient of variation is not less than the tenth coefficient of variation, determining the other training samples in the seventh target training sample except the eighth sub-training sample as abnormal samples; or

[0165] If the twelfth coefficient of variation is not less than the tenth coefficient of variation, determining other training samples in the eighth target training sample except the eighth sub-training sample as abnormal samples; or

[0166] If the eleventh coefficient of variation and the twelfth coefficient of variation are both not less than the tenth coefficient of variation, the seventh sub-training sample is determined as an abnormal sample.

[0167] In a possible implementation, the method further includes:

[0168] If the fourteenth coefficient of variation is not less than the thirteenth coefficient of variation, determining the other training samples in the ninth target training sample except the tenth sub-training sample as abnormal samples; or

[0169] If the fifteenth coefficient of variation is not less than the thirteenth coefficient of variation, determining the other training samples in the tenth target training sample except the tenth sub-training sample as abnormal samples; or

[0170] If the fourteenth coefficient of variation and the fifteenth coefficient of variation are both not less than the thirteenth coefficient of variation, the ninth sub-training sample is determined as an abnormal sample.

[0171] In a possible implementation manner, if at least one of the eleventh coefficient of variation and the twelfth coefficient of variation is not less than the tenth coefficient of variation, the method further includes:

[0172] If the eighth coefficient of variation is not less than the seventh coefficient of variation, then identify the largest sixteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the sixth feature corresponding to the sixteenth coefficient of variation; identify the eleventh model corresponding to the maximum feature weight of the sixth feature and the twelfth model corresponding to the minimum feature weight; identify the eleventh sub-training sample that is different from the eleventh training sample for training the eleventh model and the twelfth training sample for training the twelfth model, and the twelfth sub-training sample that is the same as the twelfth training sample in the eleventh training sample, divide the eleventh sub-training sample into two parts, and combine them with the twelfth sub-training sample respectively to obtain the eleventh target training sample and the twelfth target training sample, and perform model training based on the eleventh target training sample and the twelfth target training sample respectively to obtain the eleventh target model and the twelfth target model; calculate the seventeenth coefficient of variation of the sixth feature according to the at least one model and the eleventh target model, and calculate the eighteenth coefficient of variation of the sixth feature according to the at least one model and the twelfth target model, if the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, determine that there is no abnormal sample; or

[0173] If the ninth coefficient of variation is not less than the seventh coefficient of variation, then identify the largest nineteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the seventh feature corresponding to the nineteenth coefficient of variation; identify the thirteenth model corresponding to the maximum feature weight of the seventh feature and the fourteenth model corresponding to the minimum feature weight; identify the thirteenth sub-training sample that is different from the thirteenth training sample for training the thirteenth model and the fourteenth training sample for training the fourteenth model, and the fourteenth sub-training sample that is the same as the fourteenth training sample in the thirteenth training sample, divide the thirteenth sub-training sample into two parts, and combine them with the fourteenth sub-training sample respectively to obtain a thirteenth target training sample and a fourteenth target training sample, and perform model training based on the thirteenth target training sample and the fourteenth target training sample respectively to obtain a thirteenth target model and a fourteenth target model; calculate the twentieth coefficient of variation of the sixth feature according to the at least one model and the thirteenth target model, and calculate the twenty-first coefficient of variation of the sixth feature according to the at least one model and the fourteenth target model, if the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, then determine that there is no abnormal sample; or

[0174] If the eighth coefficient of variation and the ninth coefficient of variation are not less than the seventh coefficient of variation, and the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, and the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, it is determined that there are no abnormal samples.

[0175] In a possible implementation, the method further includes:

[0176] If the seventeenth coefficient of variation is not less than the sixteenth coefficient of variation, determining the other training samples in the eleventh target training sample except the twelfth sub-training sample as abnormal samples; or

[0177] If the eighteenth coefficient of variation is not less than the sixteenth coefficient of variation, determining other training samples in the twelfth target training sample except the twelfth sub-training sample as abnormal samples; or

[0178] If the seventeenth coefficient of variation and the eighteenth coefficient of variation are both not less than the sixteenth coefficient of variation, the eleventh sub-training sample is determined as an abnormal sample.

[0179] In a possible implementation, the method further includes:

[0180] If the twentieth coefficient of variation is not less than the nineteenth coefficient of variation, determining the other training samples in the thirteenth target training sample except the fourteenth sub-training sample as abnormal samples; or

[0181] If the twenty-first coefficient of variation is not less than the nineteenth coefficient of variation, determining the other training samples in the fourteenth target training sample except the fourteenth sub-training sample as abnormal samples; or

[0182] If the 20th coefficient of variation and the 21st coefficient of variation are not less than the 19th coefficient of variation, the 13th sub-training sample is determined as an abnormal sample.

[0183] In a possible implementation, the method further includes:

[0184] Based on the feature weights of each feature in each model, calculate the average difference in the feature weights of the features of each model;

[0185] A minimum mean difference with the smallest value among the mean differences is identified, and a model corresponding to the minimum mean difference is determined as an optimal model.

[0186] Since the principle of solving the problem by the above electronic device is similar to the abnormal sample determination method, the implementation of the above electronic device can refer to the above embodiment, and the repeated parts will not be repeated.

[0187] The communication bus mentioned in the above electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The communication interface 302 is used for communication between the above electronic device and other devices. The memory may include a random access memory (RAM) or a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor. The above processor may be a general-purpose processor, including a central processing unit, a network processor (NP), etc.; it may also be a digital signal processing processor (DSP), an application-specific integrated circuit, a field programmable gate array or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.

[0188] Embodiment 13:

[0189] On the basis of the above embodiments, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program executable by a processor, and when the program runs on the processor, the processor implements the following steps when executing:

[0190] For at least one trained model, calculate the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model, and identify the first feature corresponding to the first coefficient of variation with the largest value;

[0191] Identify a first model corresponding to a maximum feature weight of the first feature, and a second model corresponding to a minimum feature weight;

[0192] Identify a first sub-training sample that is different from a first training sample for training the first model and a second training sample for training the second model, and a second sub-training sample that is the same as the second training sample in the first training sample, divide the first sub-training sample into two parts, and combine them with the second sub-training sample respectively to obtain a first target training sample and a second target training sample, and perform model training based on the first target training sample and the second target training sample respectively to obtain a first target model and a second target model;

[0193] Based on the at least one model and the first target model, the second coefficient of variation of the first feature is calculated, and based on the at least one model and the second target model, the third coefficient of variation of the first feature is calculated. If the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, it is determined that there are no abnormal samples.

[0194] In a possible implementation, calculating the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model includes:

[0195] Determine the average value and standard deviation of the feature weight of each feature according to the feature weight corresponding to each feature in each model;

[0196] The ratio of the standard deviation of the feature weight of each feature to the mean value is determined as the coefficient of variation of the feature.

[0197] In a possible implementation manner, if at least one of the second coefficient of variation and the third coefficient of variation is not less than the first coefficient of variation, the method further includes:

[0198] If the second coefficient of variation is not less than the first coefficient of variation, then identify the largest fourth coefficient of variation among the coefficients of variation corresponding to other features except the first feature and the second feature corresponding to the fourth coefficient of variation; identify the third model corresponding to the maximum feature weight of the second feature and the fourth model corresponding to the minimum feature weight; identify the third training sample for training the third model and the third sub-training sample that is different from the fourth training sample for training the fourth model, and the fourth sub-training sample in the third training sample that is the same as the fourth training sample, divide the third sub-training sample into two parts, and combine them with the fourth sub-training sample respectively to obtain a third target training sample and a fourth target training sample, and perform model training based on the third target training sample and the fourth target training sample respectively to obtain a third target model and a fourth target model; calculate the fifth coefficient of variation of the second feature according to the at least one model and the third target model, and calculate the sixth coefficient of variation of the second feature according to the at least one model and the fourth target model, if the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, determine that there is no abnormal sample; or

[0199] If the third coefficient of variation is not less than the first coefficient of variation, then identify the largest seventh coefficient of variation among the coefficients of variation corresponding to other features except the first feature and the third feature corresponding to the seventh coefficient of variation; identify the fifth model corresponding to the maximum feature weight of the third feature and the sixth model corresponding to the minimum feature weight; identify the fifth sub-training sample that is different from the fifth training sample for training the fifth model and the sixth training sample for training the sixth model, and the sixth sub-training sample that is the same as the sixth training sample in the fifth training sample, divide the fifth sub-training sample into two parts, and combine them with the sixth sub-training sample respectively to obtain a fifth target training sample and a sixth target training sample, and perform model training based on the fifth target training sample and the sixth target training sample respectively to obtain a fifth target model and a sixth target model; calculate the eighth coefficient of variation of the third feature according to the at least one model and the fifth target model, and calculate the ninth coefficient of variation of the third feature according to the at least one model and the sixth target model, if the eighth coefficient of variation and the ninth coefficient of variation are both less than the seventh coefficient of variation, determine that there is no abnormal sample; or

[0200] If the second coefficient of variation and the third coefficient of variation are not less than the first coefficient of variation, and the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, and the eighth coefficient of variation and the ninth coefficient of variation are both less than the seventh coefficient of variation, it is determined that there is no abnormal sample.

[0201] In a possible implementation manner, if at least one of the fifth coefficient of variation and the sixth coefficient of variation is not less than the fourth coefficient of variation, the method further includes:

[0202] If the fifth coefficient of variation is not less than the fourth coefficient of variation, then identify the largest tenth coefficient of variation among the coefficients of variation corresponding to other features except the second feature and the fourth feature corresponding to the tenth coefficient of variation; identify the seventh model corresponding to the maximum feature weight of the fourth feature and the eighth model corresponding to the minimum feature weight; identify the seventh sub-training sample that is different from the seventh training sample for training the seventh model and the eighth training sample for training the eighth model, and the eighth sub-training sample that is the same as the eighth training sample in the seventh training sample, divide the seventh sub-training sample into two parts, and combine them with the eighth sub-training sample respectively to obtain a seventh target training sample and an eighth target training sample, and perform model training based on the seventh target training sample and the eighth target training sample respectively to obtain a seventh target model and an eighth target model; calculate the eleventh coefficient of variation of the fourth feature according to the at least one model and the seventh target model, and calculate the twelfth coefficient of variation of the fourth feature according to the at least one model and the eighth target model, if the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, determine that there is no abnormal sample; or

[0203] If the sixth coefficient of variation is not less than the fourth coefficient of variation, then identify the largest thirteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the fifth feature corresponding to the thirteenth coefficient of variation; identify the ninth model corresponding to the maximum feature weight of the fifth feature and the tenth model corresponding to the minimum feature weight; identify the ninth sub-training sample that is different from the ninth training sample for training the ninth model and the tenth training sample for training the tenth model, and the tenth sub-training sample that is the same as the tenth training sample in the ninth training sample, divide the ninth sub-training sample into two parts, and combine them with the tenth sub-training sample respectively to obtain a ninth target training sample and a tenth target training sample, and perform model training based on the ninth target training sample and the tenth target training sample respectively to obtain a ninth target model and a tenth target model; calculate the fourteenth coefficient of variation of the fifth feature according to the at least one model and the ninth target model, and calculate the fifteenth coefficient of variation of the fifth feature according to the at least one model and the tenth target model, if the fourteenth coefficient of variation and the fifteenth coefficient of variation are both less than the thirteenth coefficient of variation, determine that there is no abnormal sample; or

[0204] If the fifth coefficient of variation and the sixth coefficient of variation are not less than the fourth coefficient of variation, and the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, and the fourteenth coefficient of variation and the fifteenth coefficient of variation are both less than the thirteenth coefficient of variation, it is determined that there is no abnormal sample.

[0205] In a possible implementation, the method further includes:

[0206] If the eleventh coefficient of variation is not less than the tenth coefficient of variation, determining the other training samples in the seventh target training sample except the eighth sub-training sample as abnormal samples; or

[0207] If the twelfth coefficient of variation is not less than the tenth coefficient of variation, determining other training samples in the eighth target training sample except the eighth sub-training sample as abnormal samples; or

[0208] If the eleventh coefficient of variation and the twelfth coefficient of variation are both not less than the tenth coefficient of variation, the seventh sub-training sample is determined as an abnormal sample.

[0209] In a possible implementation, the method further includes:

[0210] If the fourteenth coefficient of variation is not less than the thirteenth coefficient of variation, determining the other training samples in the ninth target training sample except the tenth sub-training sample as abnormal samples; or

[0211] If the fifteenth coefficient of variation is not less than the thirteenth coefficient of variation, determining the other training samples in the tenth target training sample except the tenth sub-training sample as abnormal samples; or

[0212] If the fourteenth coefficient of variation and the fifteenth coefficient of variation are both not less than the thirteenth coefficient of variation, the ninth sub-training sample is determined as an abnormal sample.

[0213] In a possible implementation manner, if at least one of the eleventh coefficient of variation and the twelfth coefficient of variation is not less than the tenth coefficient of variation, the method further includes:

[0214] If the eighth coefficient of variation is not less than the seventh coefficient of variation, then identify the largest sixteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the sixth feature corresponding to the sixteenth coefficient of variation; identify the eleventh model corresponding to the maximum feature weight of the sixth feature and the twelfth model corresponding to the minimum feature weight; identify the eleventh sub-training sample that is different from the eleventh training sample for training the eleventh model and the twelfth training sample for training the twelfth model, and the twelfth sub-training sample that is the same as the twelfth training sample in the eleventh training sample, divide the eleventh sub-training sample into two parts, and combine them with the twelfth sub-training sample respectively to obtain the eleventh target training sample and the twelfth target training sample, and perform model training based on the eleventh target training sample and the twelfth target training sample respectively to obtain the eleventh target model and the twelfth target model; calculate the seventeenth coefficient of variation of the sixth feature according to the at least one model and the eleventh target model, and calculate the eighteenth coefficient of variation of the sixth feature according to the at least one model and the twelfth target model, if the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, determine that there is no abnormal sample; or

[0215] If the ninth coefficient of variation is not less than the seventh coefficient of variation, then identify the largest nineteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the seventh feature corresponding to the nineteenth coefficient of variation; identify the thirteenth model corresponding to the maximum feature weight of the seventh feature and the fourteenth model corresponding to the minimum feature weight; identify the thirteenth sub-training sample that is different from the thirteenth training sample for training the thirteenth model and the fourteenth training sample for training the fourteenth model, and the fourteenth sub-training sample that is the same as the fourteenth training sample in the thirteenth training sample, divide the thirteenth sub-training sample into two parts, and combine them with the fourteenth sub-training sample respectively to obtain a thirteenth target training sample and a fourteenth target training sample, and perform model training based on the thirteenth target training sample and the fourteenth target training sample respectively to obtain a thirteenth target model and a fourteenth target model; calculate the twentieth coefficient of variation of the sixth feature according to the at least one model and the thirteenth target model, and calculate the twenty-first coefficient of variation of the sixth feature according to the at least one model and the fourteenth target model, if the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, then determine that there is no abnormal sample; or

[0216] If the eighth coefficient of variation and the ninth coefficient of variation are not less than the seventh coefficient of variation, and the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, and the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, it is determined that there are no abnormal samples.

[0217] In a possible implementation, the method further includes:

[0218] If the seventeenth coefficient of variation is not less than the sixteenth coefficient of variation, determining the other training samples in the eleventh target training sample except the twelfth sub-training sample as abnormal samples; or

[0219] If the eighteenth coefficient of variation is not less than the sixteenth coefficient of variation, determining other training samples in the twelfth target training sample except the twelfth sub-training sample as abnormal samples; or

[0220] If the seventeenth coefficient of variation and the eighteenth coefficient of variation are both not less than the sixteenth coefficient of variation, the eleventh sub-training sample is determined as an abnormal sample.

[0221] In a possible implementation, the method further includes:

[0222] If the twentieth coefficient of variation is not less than the nineteenth coefficient of variation, determining the other training samples in the thirteenth target training sample except the fourteenth sub-training sample as abnormal samples; or

[0223] If the twenty-first coefficient of variation is not less than the nineteenth coefficient of variation, determining the other training samples in the fourteenth target training sample except the fourteenth sub-training sample as abnormal samples; or

[0224] If the 20th coefficient of variation and the 21st coefficient of variation are not less than the 19th coefficient of variation, the 13th sub-training sample is determined as an abnormal sample.

[0225] In a possible implementation, the method further includes:

[0226] Based on the feature weights of each feature in each model, calculate the average difference in the feature weights of the features of each model;

[0227] A minimum mean difference with the smallest value among the mean differences is identified, and a model corresponding to the minimum mean difference is determined as an optimal model.

[0228] Since the principle of solving the problem by the computer-readable medium provided above is similar to that of the abnormal sample determination method, after the processor executes the computer program in the computer-readable medium above, the steps implemented can refer to the above embodiment, and the repeated parts will not be repeated.

[0229] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete 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 disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0230] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, 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 flowchart and / or block diagram. 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.

[0231] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate 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 A function specified in one or more boxes.

[0232] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions 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.

[0233] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for determining abnormal samples, characterized in that: The method comprises: For at least one trained model, calculate the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model, and identify the first feature corresponding to the first coefficient of variation with the largest value; Identify a first model corresponding to a maximum feature weight of the first feature, and a second model corresponding to a minimum feature weight; Identify a first sub-training sample that is different from a first training sample for training the first model and a second training sample for training the second model, and a second sub-training sample that is the same as the second training sample in the first training sample, divide the first sub-training sample into two parts, and combine them with the second sub-training sample respectively to obtain a first target training sample and a second target training sample, and perform model training based on the first target training sample and the second target training sample respectively to obtain a first target model and a second target model; Based on the at least one model and the first target model, the second coefficient of variation of the first feature is calculated, and based on the at least one model and the second target model, the third coefficient of variation of the first feature is calculated. If the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, it is determined that there are no abnormal samples.

2. The method according to claim 1, characterized in that If at least one of the second coefficient of variation and the third coefficient of variation is not less than the first coefficient of variation, the method further includes: If the second coefficient of variation is not less than the first coefficient of variation, then identify the largest fourth coefficient of variation among the coefficients of variation corresponding to other features except the first feature and the second feature corresponding to the fourth coefficient of variation; identify the third model corresponding to the maximum feature weight of the second feature and the fourth model corresponding to the minimum feature weight; identify the third training sample for training the third model and the third sub-training sample that is different from the fourth training sample for training the fourth model, and the fourth sub-training sample in the third training sample that is the same as the fourth training sample, divide the third sub-training sample into two parts, and combine them with the fourth sub-training sample respectively to obtain a third target training sample and a fourth target training sample, and perform model training based on the third target training sample and the fourth target training sample respectively to obtain a third target model and a fourth target model; calculate the fifth coefficient of variation of the second feature according to the at least one model and the third target model, and calculate the sixth coefficient of variation of the second feature according to the at least one model and the fourth target model, if the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, determine that there is no abnormal sample; or If the third coefficient of variation is not less than the first coefficient of variation, then identify the largest seventh coefficient of variation among the coefficients of variation corresponding to other features except the first feature and the third feature corresponding to the seventh coefficient of variation; identify the fifth model corresponding to the maximum feature weight of the third feature and the sixth model corresponding to the minimum feature weight; identify the fifth sub-training sample that is different from the fifth training sample for training the fifth model and the sixth training sample for training the sixth model, and the sixth sub-training sample that is the same as the sixth training sample in the fifth training sample, divide the fifth sub-training sample into two parts, and combine them with the sixth sub-training sample respectively to obtain a fifth target training sample and a sixth target training sample, and perform model training based on the fifth target training sample and the sixth target training sample respectively to obtain a fifth target model and a sixth target model; calculate the eighth coefficient of variation of the third feature according to the at least one model and the fifth target model, and calculate the ninth coefficient of variation of the third feature according to the at least one model and the sixth target model, if the eighth coefficient of variation and the ninth coefficient of variation are both less than the seventh coefficient of variation, determine that there is no abnormal sample; or If the second coefficient of variation and the third coefficient of variation are not less than the first coefficient of variation, and the fifth coefficient of variation and the sixth coefficient of variation are both less than the fourth coefficient of variation, and the eighth coefficient of variation and the ninth coefficient of variation are both less than the seventh coefficient of variation, it is determined that there is no abnormal sample.

3. The method according to claim 2, characterized in that If at least one of the fifth coefficient of variation and the sixth coefficient of variation is not less than the fourth coefficient of variation, the method further includes: If the fifth coefficient of variation is not less than the fourth coefficient of variation, then identify the largest tenth coefficient of variation among the coefficients of variation corresponding to other features except the second feature and the fourth feature corresponding to the tenth coefficient of variation; identify the seventh model corresponding to the maximum feature weight of the fourth feature and the eighth model corresponding to the minimum feature weight; identify the seventh sub-training sample that is different from the seventh training sample for training the seventh model and the eighth training sample for training the eighth model, and the eighth sub-training sample that is the same as the eighth training sample in the seventh training sample, divide the seventh sub-training sample into two parts, and combine them with the eighth sub-training sample respectively to obtain a seventh target training sample and an eighth target training sample, and perform model training based on the seventh target training sample and the eighth target training sample respectively to obtain a seventh target model and an eighth target model; calculate the eleventh coefficient of variation of the fourth feature according to the at least one model and the seventh target model, and calculate the twelfth coefficient of variation of the fourth feature according to the at least one model and the eighth target model, if the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, determine that there is no abnormal sample; or If the sixth coefficient of variation is not less than the fourth coefficient of variation, then identify the largest thirteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the fifth feature corresponding to the thirteenth coefficient of variation; identify the ninth model corresponding to the maximum feature weight of the fifth feature and the tenth model corresponding to the minimum feature weight; identify the ninth sub-training sample that is different from the ninth training sample for training the ninth model and the tenth training sample for training the tenth model, and the tenth sub-training sample that is the same as the tenth training sample in the ninth training sample, divide the ninth sub-training sample into two parts, and combine them with the tenth sub-training sample respectively to obtain a ninth target training sample and a tenth target training sample, and perform model training based on the ninth target training sample and the tenth target training sample respectively to obtain a ninth target model and a tenth target model; calculate the fourteenth coefficient of variation of the fifth feature according to the at least one model and the ninth target model, and calculate the fifteenth coefficient of variation of the fifth feature according to the at least one model and the tenth target model, if the fourteenth coefficient of variation and the fifteenth coefficient of variation are both less than the thirteenth coefficient of variation, determine that there is no abnormal sample; or If the fifth coefficient of variation and the sixth coefficient of variation are not less than the fourth coefficient of variation, and the eleventh coefficient of variation and the twelfth coefficient of variation are both less than the tenth coefficient of variation, and the fourteenth coefficient of variation and the fifteenth coefficient of variation are both less than the thirteenth coefficient of variation, it is determined that there is no abnormal sample.

4. The method according to claim 3, characterized in that The method further comprises: If the eleventh coefficient of variation is not less than the tenth coefficient of variation, determining the other training samples in the seventh target training sample except the eighth sub-training sample as abnormal samples; or If the twelfth coefficient of variation is not less than the tenth coefficient of variation, determining other training samples in the eighth target training sample except the eighth sub-training sample as abnormal samples; or If the eleventh coefficient of variation and the twelfth coefficient of variation are not less than the tenth coefficient of variation, determining the seventh sub-training sample as an abnormal sample; or If the fourteenth coefficient of variation is not less than the thirteenth coefficient of variation, determining the other training samples in the ninth target training sample except the tenth sub-training sample as abnormal samples; or If the fifteenth coefficient of variation is not less than the thirteenth coefficient of variation, determining the other training samples in the tenth target training sample except the tenth sub-training sample as abnormal samples; or If the fourteenth coefficient of variation and the fifteenth coefficient of variation are both not less than the thirteenth coefficient of variation, the ninth sub-training sample is determined as an abnormal sample.

5. The method according to claim 3, characterized in that: If at least one of the eleventh coefficient of variation and the twelfth coefficient of variation is not less than the tenth coefficient of variation, the method further includes: If the eighth coefficient of variation is not less than the seventh coefficient of variation, then identify the largest sixteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the sixth feature corresponding to the sixteenth coefficient of variation; identify the eleventh model corresponding to the maximum feature weight of the sixth feature and the twelfth model corresponding to the minimum feature weight; identify the eleventh sub-training sample that is different from the eleventh training sample for training the eleventh model and the twelfth training sample for training the twelfth model, and the twelfth sub-training sample that is the same as the twelfth training sample in the eleventh training sample, divide the eleventh sub-training sample into two parts, and combine them with the twelfth sub-training sample respectively to obtain the eleventh target training sample and the twelfth target training sample, and perform model training based on the eleventh target training sample and the twelfth target training sample respectively to obtain the eleventh target model and the twelfth target model; calculate the seventeenth coefficient of variation of the sixth feature according to the at least one model and the eleventh target model, and calculate the eighteenth coefficient of variation of the sixth feature according to the at least one model and the twelfth target model, if the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, determine that there is no abnormal sample; or If the ninth coefficient of variation is not less than the seventh coefficient of variation, then identify the largest nineteenth coefficient of variation among the coefficients of variation corresponding to the other features except the second feature and the seventh feature corresponding to the nineteenth coefficient of variation; identify the thirteenth model corresponding to the maximum feature weight of the seventh feature and the fourteenth model corresponding to the minimum feature weight; identify the thirteenth sub-training sample that is different from the thirteenth training sample for training the thirteenth model and the fourteenth training sample for training the fourteenth model, and the fourteenth sub-training sample that is the same as the fourteenth training sample in the thirteenth training sample, divide the thirteenth sub-training sample into two parts, and combine them with the fourteenth sub-training sample respectively to obtain a thirteenth target training sample and a fourteenth target training sample, and perform model training based on the thirteenth target training sample and the fourteenth target training sample respectively to obtain a thirteenth target model and a fourteenth target model; calculate the twentieth coefficient of variation of the sixth feature according to the at least one model and the thirteenth target model, and calculate the twenty-first coefficient of variation of the sixth feature according to the at least one model and the fourteenth target model, if the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, then determine that there is no abnormal sample; or If the eighth coefficient of variation and the ninth coefficient of variation are not less than the seventh coefficient of variation, and the seventeenth coefficient of variation and the eighteenth coefficient of variation are both less than the sixteenth coefficient of variation, and the twentieth coefficient of variation and the twenty-first coefficient of variation are both less than the nineteenth coefficient of variation, it is determined that there are no abnormal samples.

6. The method according to claim 5, characterized in that The method further comprises: If the seventeenth coefficient of variation is not less than the sixteenth coefficient of variation, determining the other training samples in the eleventh target training sample except the twelfth sub-training sample as abnormal samples; or If the eighteenth coefficient of variation is not less than the sixteenth coefficient of variation, determining other training samples in the twelfth target training sample except the twelfth sub-training sample as abnormal samples; or If the seventeenth coefficient of variation and the eighteenth coefficient of variation are not less than the sixteenth coefficient of variation, determining the eleventh sub-training sample as an abnormal sample; or If the twentieth coefficient of variation is not less than the nineteenth coefficient of variation, determining the other training samples in the thirteenth target training sample except the fourteenth sub-training sample as abnormal samples; or If the twenty-first coefficient of variation is not less than the nineteenth coefficient of variation, determining the other training samples in the fourteenth target training sample except the fourteenth sub-training sample as abnormal samples; or If the 20th coefficient of variation and the 21st coefficient of variation are not less than the 19th coefficient of variation, the 13th sub-training sample is determined as an abnormal sample.

7. The method according to claim 1, characterized in that The method further comprises: Based on the feature weights of each feature in each model, calculate the average difference in the feature weights of the features of each model; A minimum mean difference with the smallest value among the mean differences is identified, and a model corresponding to the minimum mean difference is determined as an optimal model.

8. An abnormal sample determination device, characterized in that: The device comprises: A calculation module, for calculating the coefficient of variation of each feature according to the feature weight corresponding to each feature in each model for at least one trained model, and identifying a first feature corresponding to a first coefficient of variation with a maximum value; an identification module, used to identify a first model corresponding to a maximum feature weight of the first feature, and a second model corresponding to a minimum feature weight; An abnormal sample determination module is used to identify different first sub-training samples in a first training sample for training the first model and a second training sample for training the second model, and a second sub-training sample in the first training sample that is the same as the second training sample, divide the first sub-training sample into two parts, and combine them with the second sub-training sample respectively to obtain a first target training sample and a second target training sample, and perform model training based on the first target training sample and the second target training sample respectively to obtain a first target model and a second target model; calculate the second coefficient of variation of the first feature according to the at least one model and the first target model, and calculate the third coefficient of variation of the first feature according to the at least one model and the second target model; if the second coefficient of variation and the third coefficient of variation are both smaller than the first coefficient of variation, it is determined that there is no abnormal sample.

9. An electronic device, characterized in that: The electronic device comprises at least a processor and a memory, and the processor is used to implement the steps of the abnormal sample determination method described in any one of claims 1 to 7 when executing a computer program stored in the memory.

10. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements the steps of the abnormal sample determination method described in any one of claims 1 to 7.

Images