CN112699944A

CN112699944A - Order-returning processing model training method, processing method, device, equipment and medium

Info

Publication number: CN112699944A
Application number: CN202011636325.0A
Authority: CN
Inventors: 王欣晟; 陆堃彪; 张青清; 李航
Original assignee: China Unionpay Co Ltd
Current assignee: China Unionpay Co Ltd
Priority date: 2020-12-31
Filing date: 2020-12-31
Publication date: 2021-04-23
Anticipated expiration: 2040-12-31

Abstract

The application discloses a training method, a processing method, a device, equipment and a medium of a return processing model, and belongs to the field of data processing. The order return processing method comprises the following steps: converting the description content of the recorded order-returning event into an order-returning event vector; taking the returned order event vector as the input of a preset returned order processing model to obtain a returned order reason classification result output by the returned order processing model, wherein the returned order reason classification result is used for representing the returned order reason of the returned order event predicted by the returned order processing model, and the returned order processing model is a classification model obtained by training through a classification algorithm based on the returned order sample vector; and outputting the reason for the reason. According to the method and the device, the efficiency of the order-returning processing can be improved.

Description

Order-returning processing model training method, processing method, device, equipment and medium

Technical Field

The application belongs to the field of data processing, and particularly relates to a training method, a processing method, a device, equipment and a medium for a return processing model.

Background

The order-returning means that the order-returning initiator has doubt on the business execution result and the like, and provides rejection on the business execution result to a verification organization.

In the current stage, special operators are required to receive the suspicious service work orders for processing the order refunds, and the operators judge the situation and reason of the order refunds according to actual conditions and issue the order refunds letter. And rechecking the returned order by rechecking personnel. The order-returning process needs a plurality of processes which are carried out by the operators, the order-receiving mechanism, the rechecker, the approver and the like, the processes are complex and need to be completed manually, and the processing efficiency is very low.

Disclosure of Invention

The embodiment of the application provides a method, a device, equipment and a medium for training a return processing model, which can improve the efficiency of return processing.

In a first aspect, an embodiment of the present application provides a method for processing an invoice, including: converting the description content of the recorded order-returning event into an order-returning event vector; taking the returned order event vector as the input of a preset returned order processing model to obtain a returned order reason classification result output by the returned order processing model, wherein the returned order reason classification result is used for representing the returned order reason of the returned order event predicted by the returned order processing model, and the returned order processing model is a classification model obtained by training through a classification algorithm based on the returned order sample vector; and outputting the reason for the reason.

In a second aspect, an embodiment of the present application provides a method for training a drop-out processing model, including: converting the description content of the recorded order-returning sample event into an order-returning sample vector; setting model parameters of the classification model, and initializing the classification model according to the model parameters; selecting at least part of the order-returning sample vectors as a training set, and carrying out iterative training on the classification model to obtain an order-returning processing model, wherein the input of the order-returning processing model comprises the order-returning sample vectors, the output of the order-returning processing model comprises an order-returning reason classification result, and the order-returning reason classification result is used for representing the order-returning reasons of the order-returning sample events obtained through prediction of the order-returning processing model.

In a third aspect, an embodiment of the present application provides an order rejection processing apparatus, including: the vector conversion module is used for converting the recorded description content of the order returning event into an order returning event vector; the calculation module is used for taking the order-returning event vector as the input of a preset order-returning processing model to obtain an order-returning reason classification result output by the order-returning processing model, the order-returning reason classification result is used for representing the order-returning reason of the order-returning event predicted by the order-returning processing model, and the order-returning processing model is a classification model obtained by training through a classification algorithm based on the order-returning sample vector; and the output module outputs the reason for returning the order of the order-returning event according to the classification result of the reason for returning the order.

In a fourth aspect, an embodiment of the present application provides a training apparatus for a leave-order processing model, including: the vector conversion module is used for converting the description content of the recorded order-returning sample event into an order-returning sample vector; the initialization module is used for setting model parameters of the classification model and initializing the classification model according to the model parameters; and the training module is used for selecting at least part of the order-returning sample vectors as a training set, performing iterative training on the classification model to obtain the order-returning processing model, wherein the input of the order-returning processing model comprises the order-returning sample vectors, the output of the order-returning processing model comprises an order-returning reason classification result, and the order-returning reason classification result is used for representing the order-returning reasons of the order-returning sample events obtained by predicting through the order-returning processing model.

In a fifth aspect, an embodiment of the present application provides an order return processing apparatus, including: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the method of the first aspect.

In a sixth aspect, an embodiment of the present application provides an order-returning processing model training apparatus, including: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the rollout processing model training method of the second aspect.

In a seventh aspect, an embodiment of the present application provides a computer-readable storage medium, where computer program instructions are stored on the computer-readable storage medium, and when the computer program instructions are executed by a processor, the order-returning processing method of the first aspect or the order-returning processing model training method of the second aspect is implemented.

The application provides a training method, a processing method, a device, equipment and a medium of a receipt-returning processing model, which are obtained by training a receipt-returning sample vector in advance through a classification algorithm. The recorded order-returning event vector converted from the order-returning event is used as the input of the order-returning processing model, and the order-returning reason of the order-returning event is output through the sorting result of the order-returning reason output by the order-returning processing model, so that the order-returning processing is realized, the manual process of the order-returning processing is omitted, the flow of the order-returning processing is simplified, and the efficiency of the order-returning processing is improved.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments of the present application will be briefly described below, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.

FIG. 1 is a flow diagram of an embodiment of a drop-out processing model training method provided herein;

FIG. 2 is a flow diagram of another embodiment of a drop-out processing model training method provided herein;

FIG. 3 is a flowchart of an embodiment of a method for processing a return order provided by the present application;

FIG. 4 is a flowchart of another embodiment of a method for processing a return order provided by the present application;

FIG. 5 is a diagram illustrating an example of the contribution of sample vocabulary to the reason for returning the order according to the embodiment of the present application;

FIG. 6 is a flowchart of another embodiment of a method for processing a return order provided by the present application;

FIG. 7 is a schematic structural diagram of an embodiment of a return processing model training apparatus provided in the present application;

FIG. 8 is a schematic structural diagram of another embodiment of a return processing model training apparatus provided in the present application;

FIG. 9 is a schematic structural diagram of an exemplary embodiment of an order rejection processing apparatus;

FIG. 10 is a schematic structural diagram of another embodiment of an order rejection processing apparatus provided in the present application;

FIG. 11 is a schematic diagram illustrating an exemplary embodiment of a return processing model training apparatus provided herein;

fig. 12 is a schematic structural diagram of an embodiment of an order rejection processing apparatus provided in the present application.

Detailed Description

Features and exemplary embodiments of various aspects of the present application will be described in detail below, and in order to make objects, technical solutions and advantages of the present application more apparent, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are intended to be illustrative only and are not intended to be limiting. It will be apparent to one skilled in the art that the present application may be practiced without some of these specific details. The following description of the embodiments is merely intended to provide a better understanding of the present application by illustrating examples thereof.

The order-returning means that the order-returning initiator has doubt on the business execution result and the like, and the rejection of the business execution result proposed to the verification organization can relate to business scenes such as errors, auditing, laws and the like. In the current stage, the processing of the returned order is that the business work order is received by the operator, and the order adjusting application and the like are initiated to be rechecked by the order receiving mechanism. If the order receiving mechanism does not reply in an overdue mode, the operator needs to contact the order returning initiator to verify the service situation and restore the service scene, the applicable order returning situation and the corresponding reason are judged according to own experience and the service situation, materials required by the order returning are prepared, an order returning letter is issued, and the returned order is sent to the dispute processing rechecking personnel for rechecking. The whole order-returning process needs to involve multiple processes performed by operators, order-receiving mechanisms, recheckers, examination and approval personnel and the like, the order-returning process is complex and needs to be completed manually, and the processing efficiency is very low. And because of manual processing, the judgment subjectivity of the existing bill returning processing is strong, and the objectivity judgment is lacked.

The application provides a method, a device, equipment and a medium for training a receipt refunding processing model, which can convert the description content of a receipt refunding event into a receipt refunding event vector which can be processed by a machine, process the receipt refunding event vector by using the receipt refunding processing model obtained by the training of a receipt refunding sample event, and take the receipt refunding reason represented by the receipt refunding reason classification result output by the receipt refunding processing model as the receipt refunding reason of the receipt refunding event. The order-returning processing flow does not need manual judgment, the processing efficiency is high, and the objectivity is realized.

The following describes in detail a drop-out processing model training method, a drop-out processing model training device, a drop-out processing model training apparatus, a drop-out processing apparatus, and a computer-readable storage medium, respectively, provided by the present application.

The application provides a training method of a return processing model. Fig. 1 is a flowchart of an embodiment of a return processing model training method provided in the present application. As shown in fig. 1, the training method of the leave-out processing model may include steps S101 to S103.

In step S101, the description content of the entered destage sample event is converted into a destage sample vector.

The return sample event is an event serving as a sample for training the return processing model, and the return sample event can be selected according to the scene and the requirement, and the type and the number of the return sample events are not limited herein.

The description content of the entered sample event of the receipt may be stored sporadically in the form of a file without systematic processing. And the description of the sample event is typically stored in an unstructured or semi-structured form. For example, the description content of the sample event of the receipt is stored in a scanning file mode, a text mode or a table mode. The description content of the destaging sample event in the unstructured form or the semi-structured form is converted into a destaging sample vector, so that subsequent calculation and processing are facilitated.

The destage sample vector includes a plurality of elements, and the value of each element may represent the frequency of occurrence of the vocabulary corresponding to the element in the descriptive content of the destage sample event. The vocabulary corresponding to the elements of the returned sample vector can be determined according to the importance degree of the vocabulary in the description contents of a plurality of returned sample events. The destage sample vector can characterize the descriptive content of the destage sample event.

In step S102, model parameters of the classification model are set, and the classification model is initialized according to the model parameters.

The classification model is a model of a classification algorithm. For example, the classification model may specifically include a decision tree classification model, a bayesian classification model, a neural network model, a fuzzy classification model, a support vector machine classification model, a k-nearest neighbor model, and the like, which are not limited herein. Each class of classification model may also be classified into a finer class of models, for example, the Decision Tree classification model may include an XGBoost model, a Gradient Boosting Decision Tree (GBDT) model, and the like.

The model parameters may be determined according to the type of classification model, and are not limited herein. For example, the classification model includes an XGBoost model, and correspondingly, the model parameters may include iteration number, maximum depth, learning rate, L2 regularization term, micro-lossy function, regularization term constant parameters, and the like.

The initialized classification model may have a value of 0 and the objective function of the initialized classification model may also be 0. Through multiple iterative optimization, the classification model can be trained into a required order-returning processing model.

In step S103, at least part of the order-returning sample vectors are selected as a training set, and the classification model is iteratively trained to obtain an order-returning processing model.

Under the condition of obtaining the singleton sample vector, the singleton sample vector can be divided into two parts, wherein one part is used as a training set for model training, and the other part is used as a testing set for testing the trained model. For example, 70% of the destage sample vectors may be randomly selected as the training set and the remaining 30% of the destage sample vectors may be randomly selected as the testing set.

And performing iterative training on the classification model by using the order-returning sample vector as a training set, wherein the classification model after the iterative training is the order-returning processing model. The input to the destage processing model includes a destage sample vector. The output of the return processing model comprises the return reason classification result. The sorting result of the order-returning reasons is used for representing the order-returning reasons of the order-returning sample events predicted by the order-returning processing model.

In some examples, the reason for. The reason identification code is used for distinguishing different reason of the receipt. The reason for the reason identifier included in the reason for the reason classification result may be used as the reason for the predicted reason for the reason classification.

In other examples, the reason for returning the order may include the reason identification codes of a plurality of reasons for returning the order and the predicted value of each reason for returning the order. The prediction value can be used to characterize the matching degree of the order-returning reason and the order-returning sample event. The reason for the return order with the highest matching degree with the return order sample event is the reason for the return order of the predicted return order sample event.

The reason identification code of the reason for returning the order may specifically include chinese characters, numbers, letters, other characters, and the like, and is not limited herein. For example, there are seven reasons for a refund, respectively, "transaction unsuccessful, deducted", "transaction repeat submission clearing", "cancelled transaction", "returned (credit adjusted) transaction funds not submitted clearing", "transaction not approved", "cardholder disputed for transaction amount", "otherwise paid". The receipt-refund identification code may be the chinese characters "transaction unsuccessful, deducted", "transaction repeat submission clearing", "cancelled transaction", "return (credit adjusted) transaction funds not submitted clearing", "transaction not approved", "cardholder disputed for transaction amount", "payment otherwise". The ticket refund identification code may also be the numbers "001", "002", "003", "004", "005", "006", "007", each of which corresponds to a reason for the refund.

For another example, the result of the classification of the reason for the return is [ -0.987231734.6105857-2.279877-2.2601516-2.2902958-2.2136965-1.6816981 ], -0.98723173, 4.6105857, -2.279877, -2.2601516, -2.2902958, -2.2136965 and-1.6816981 are the predicted values respectively corresponding to seven reasons for the return, wherein 4.6105857 is the largest predicted value and the reason for the return indicating the sample event for the return is the "transaction repeat filing clearing" second reason.

The trained order-returning processing model can be used for predicting the order-returning reason of the order-returning event. In some examples, in the case that a part of the destaging sample vectors is selected to perform iterative training on the classification model, another part of the destaging sample vectors can be used to test the destaging processing model, and the prediction effect of the destaging processing model is tested.

In the embodiment of the application, the description content of the recorded order-returning sample event can be converted into an order-returning sample vector, so that at least part of the order-returning sample vector is used as a training set to perform iterative training on the initialized classification model, and the classification model after the iterative training is used as an order-returning processing model. The input of the order-returning processing model comprises an order-returning sample vector, and the output of the order-returning processing model comprises the sorting result of the order-returning reasons. And the reason for the receipt corresponding to the vector of the input receipt event can be predicted according to the classification result of the receipt reason, so that the receipt processing is realized, the manual flow of the receipt processing is omitted, the flow of the receipt processing is simplified, and the efficiency of the receipt processing is improved. Moreover, the analysis of the order-returning event by the order-returning processing model has a uniform standard, so that the subjectivity of the order-returning processing can be reduced, and a more objective order-returning processing result can be obtained.

The following description will be given by taking an example in which the order-returning processing model includes a decision tree and is obtained by performing iterative training. FIG. 2 is a flowchart of another embodiment of a return processing model training method provided in the present application. Fig. 2 is different from fig. 1 in that step S101 in fig. 1 may be specifically detailed as step S1011 to step S1013, and step S103 in fig. 1 may be specifically detailed as step S1031 and step S1032.

In step S1011, the description content of the entered return sample event is subjected to normalization processing.

In some cases, the description content of the destage sample event may be in an unstructured form or a semi-structured form, and the description content of different destage sample events may be in different forms, and a normalization process is required to be performed, so that the description content of the destage sample event is in a unified standard form, that is, the normalization process is performed. The file formats of the description contents of the receipt rejection sample events can be unified, and then the description contents of the receipt rejection sample events with the unified file formats are converted into a data table with the unified format.

Specifically, the input unstructured file or semi-structured file with the description content of the receipt sample event can be converted into a standard template with the description content of the receipt sample event. And analyzing the standard template in which the description content of the order-returning sample event is recorded to obtain the structured data corresponding to the description content of the order-returning sample event. And carrying out data cleaning on the structured data to obtain the description content of the return order sample event after the standardized processing.

The unstructured file may specifically include a PDF scan file, a Word file, and the like, and is not limited herein. The semi-structured document may specifically include, but is not limited to, an XML document, a mail document, and the like. The structured data may be data specifying the content of a specific field, and is not limited herein.

For example, the originally entered description content of the callback sample event is a PDF scan file, each page of the PDF scan file may be converted into a picture, and the picture may be converted into a word format file or an excel format file by using an Optical Character Recognition (OCR) technique. Reading the content in the word format file or the excel format file, and converting the word format file or the excel format file into a fixed word template or an excel template through a written script. The word template or the excel template can be further analyzed through the script, and converted into structured data to be stored in the database. The structured data may include fields for sequence number, scene description, reason for chargeback code, nature of service, etc.

And misoperation can exist in the description content of the entered returned sample event, so that the identification of the description content is influenced. The content of the structured data of the description content can be standardized through data cleaning, and the readability and the accuracy of the structured data are improved. Data cleansing specifically may include removing the extraordinary punctuation "? The operations of "", "/", "@", "^", "&" etc., and may also include the operation of removing line break "\\ n", "\ r \ n", etc., without limitation.

For example, for "# my 10:09 today's telephone contact cardholder recorded in scenario description field, cardholder says that payment is made on self-service payment machine of hospital using credit card on the same day

The fee, after the transaction, the implement displays the transaction timeout, but the cardholder is confident that the deduction has been made. The cardholder believes the transaction at that time

And the card is automatically corrected at the later stage, so that the card is swiped again to pay the money. The cardholder only enjoys the color Doppler with one-time value of 312.67 yuan at the same day

The service is checked. The cardholder tries to contact the hospital, but the hospital is said not to receive the transaction and cannot refund the transaction. "the content recorded in the scene description field after data cleaning is" I contact the cardholder by phone 10:09 today, the cardholder calls that the self-service payment machine of the hospital uses credit card to pay the fee on the same day, the machine displays that the transaction is overtime after the transaction, but the cardholder inquires the deducted money by WeChat. At that time, the cardholder thinks that the transaction will be automatically corrected in the later period, so that the cardholder swipes the card again to pay the payment. The cardholder enjoys the color Doppler check service with a value of 312.67 dollars only once on the day. The cardholder tries to contact the hospital, but the hospital is said not to receive the transaction and cannot refund the transaction. "

In step S1012, the normalized description content of the return sample event is segmented to obtain sample words in the description content of the return sample event.

The word segmentation method may be set according to the scene and the requirement, and is not limited herein. For example, the word segmentation method may include a forward maximum matching method, a reverse maximum matching method, a two-way matching segmentation method, a Hidden Markov Model (HMM) method, a deep learning method, and the like, which are not limited herein.

In step S1013, a destage sample vector is generated based on the frequency of occurrence of sample words existing in the vocabulary library in the description content of the destage event.

The vocabulary library can be obtained by utilizing sample vocabularies obtained after the description contents of the order-returning sample events are segmented. Specifically, the generation of the vocabulary library can be realized by adopting a word frequency-reverse file frequency (tfidf) method, so that the vocabulary library is used for generating the order-withdrawing sample vector. The vocabulary in the vocabulary library obtained by the tfidf method can represent the importance degree of the vocabulary in the vocabulary library.

Specifically, the word frequency and the inverse text frequency index of each sample vocabulary are calculated. And obtaining the importance index of each sample vocabulary according to the word frequency and the inverse text frequency index of each sample vocabulary. And according to the sequence of the importance indexes from large to small, taking a preset number of sample vocabularies to form a vocabulary library.

The word frequency of the sample vocabulary is the word frequency of the sample vocabulary appearing in the description content of the fallback sample vector, and can be represented by TF (t, d). Wherein t refers to sample vocabulary, and d is description content.

The inverse text frequency index of the sample vocabulary may be specifically calculated according to equation (1):

where IDF (t, D) is an inverse text frequency index, DF (t, D) is the number of descriptions including the sample vocabulary t, and | D | is the total number of descriptions.

The importance index of the sample vocabulary is tfidf index, and can be specifically calculated according to formula (2):

TFIDF(t,d,D)＝TF(t,d)×IDF(t,D) (2)

wherein TFIDF (t, D) is the importance index of the sample vocabulary. The larger the importance index of the sample vocabulary is, the higher the degree of importance of the sample vocabulary is represented. The vocabulary library may be formed by taking a predetermined number of sample vocabularies with importance index sizes ranked first. For example, a vocabulary library is formed by taking the sample vocabulary with the importance index size ranked 500 above.

The description content of the return sample event can be converted into a return sample vector according to whether the description content of the return sample event has sample words in a word library and the frequency of the sample words in the word library appearing in the description content of the return sample event.

For example, the vocabulary library includes 500 sample vocabularies, and the destage sample vector may be a 500-dimensional vector. The first 25 dimensions of a certain singleton sample vector are: 0.2780091,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.53908061,0,0. Correspondingly, the frequency of the 23 rd dimension corresponding sample vocabulary of the destage sample vector appearing in the description content of the destage sample event is about twice as frequent as the frequency of the 1 st dimension corresponding sample vocabulary of the destage sample vector appearing in the description content of the destage sample event. Sample vocabularies corresponding to the 2 nd to 22 nd, 24 th and 25 th dimensions of the destaged sample vector do not appear in the description content of the destaged sample event.

In this embodiment of the application, the description content of the policy return sample event may also be converted into a policy return event vector by using a text vectorization method such as word2vec and bert, which is not limited herein.

The converted order-returning sample vector is more convenient for a machine to process, the speed of training the order-returning processing model can be increased, and the efficiency of training the order-returning processing model is improved.

In step S1031, a target function of the classification model is obtained according to the first partial derivative sum, the second partial derivative sum, the number of leaf nodes of the decision tree, and the regularization term constant parameter.

The first partial derivative sum is a sum of first order partial derivatives of the loss function of the first target vector with respect to the classification model of the last iteration. The first target vector comprises a singleton sample vector corresponding to a leaf node of the decision tree. The second partial derivative sum is a second order partial derivative sum of the loss function of the first target vector with respect to the classification model of the last iteration.

For example, using the XGBoost classification algorithm, the initial objective function may be 0. The objective function is represented by Obj, Obj^(m)For the objective function obtained by the mth iteration, the objective function obtained by the mth iteration can be specifically calculated according to equation (3):

wherein, Obj^(m)For the objective function obtained in the mth iteration, G_jIs the sum of the first partial derivatives, H_jAnd T is the sum of the second partial derivatives, T is the leaf node number of the decision tree, and lambda and gamma are regularization term constant parameters.

The sum of the first partial derivatives and the sum of the second partial derivatives can be calculated according to equations (4) to (7):

wherein G is_jG which may represent all the singleton sample vectors falling on leaf node j of the decision tree_iSum of (a), (b), g_iClassification model for m-1 iterations for a loss function

First order partial derivative of (H)_jMay represent the sum of all the singleton sample vectors, h, that fall on leaf node j of the decision tree_iClassification model for m-1 iterations for a loss function

Second order partial derivatives of (d).

In step S1032, the model parameters are adjusted, and the classification model corresponding to the objective function having the expected value is used as the order rejection processing model.

And through multiple iterations, taking the classification model corresponding to the objective function with the expected value as a receipt processing model. Specifically, the objective function value may be iterated to a converged classification model as a de-ordering processing model. The function expression of the destaging processing model is shown as equation (8):

the output of the classification model after m iterations, namely the order-withdrawing processing model, can be calculated through the formula (8). Wherein f is_k() A kth class decision tree generated for the iteration. Vector x of samples to be singled out_iInput receipt processing model

Can obtain and demote a sample vector x_iAnd sorting the result of the order-returning reason predicted by the corresponding order-returning processing model.

The order-returning reason of the order-returning event can be predicted by using the order-returning processing model obtained by multiple times of iterative training. The application also provides a receipt-returning processing method, which utilizes the receipt-returning processing model trained in the embodiment to process the receipt-returning event. Fig. 3 is a flowchart of an embodiment of a method for processing a return order. As shown in fig. 3, the order rejection processing method may include steps S201 to S203.

In step S201, the description content of the entered policy return event is converted into a policy return event vector.

The order-returning event is an event which needs to judge the reason for returning the order.

In some examples, the description of the entered rollout event is normalized. And performing word segmentation on the description content of the return event after the normalization processing to obtain a description vocabulary in the description content of the return event. And generating a return event vector according to the occurrence frequency of the description vocabularies in the preset vocabulary library in the description content of the return event.

For the specific content of the vocabulary library, reference may be made to the related description in the above embodiments, and further description is omitted here.

The normalized processing of the description content of the entered return event can be specifically executed as: converting the input unstructured file or semi-structured file recorded with the description content of the order-returning event into a standard template recorded with the description content of the order-returning event; analyzing a standard template in which the description content of the order-returning event is recorded to obtain structured data corresponding to the description content of the order-returning event; and carrying out data cleaning on the structured data to obtain the description content of the return order event after the standardized processing.

The description content of the entered ticket refund event is converted into the specific content of the ticket refund event vector, reference may be made to the description for converting the description content of the entered ticket refund sample event into the relevant description of the ticket refund sample vector in the above embodiment, and the conversion methods are basically the same and will not be described herein again.

In step S202, the policy return event vector is used as an input of a preset policy return processing model, and a policy return reason classification result output by the policy return processing model is obtained.

The order-returning processing model is a classification model obtained by training by using a classification algorithm based on an order-returning sample vector. For the training method of the leave-single processing model, reference may be made to the related description in the above embodiments, and details are not repeated herein.

The sorting result of the order-returning reasons is used for representing the order-returning reasons of the order-returning events predicted by the order-returning processing model. In some examples, the reason for. In other examples, the reason for returning the order classification result includes the reason identification codes of a plurality of reason for returning the order and the predicted value of each reason for returning the order. The predicted value is used for representing the matching degree of the order-returning reason and the order-returning event.

For the specific content of the sorting result of the order-returning reason, reference may be made to the related description in the above embodiments, and further description is omitted here.

In step S203, the reason for the return of the return event is output based on the reason for the return classification result.

The reason for returning orders can represent the reason for returning orders of the event for returning orders, and the reason for returning orders of the event for returning orders can be output according to the classification result of the reason for returning orders, so that the reason for returning orders can be displayed more intuitively, and the staff can conveniently obtain the reason for returning orders.

In the embodiment of the application, the order-returning processing model is obtained by training the order-returning sample vector in advance by using a classification algorithm. The recorded order-returning event vector converted from the order-returning event is used as the input of the order-returning processing model, and the order-returning reason of the order-returning event is output through the sorting result of the order-returning reason output by the order-returning processing model, so that the order-returning processing is realized, the manual process of the order-returning processing is omitted, the flow of the order-returning processing is simplified, and the efficiency of the order-returning processing is improved. Moreover, the analysis of the order-returning event by the order-returning processing model has a uniform standard, so that the subjectivity of the order-returning processing can be reduced, and a more objective order-returning processing result can be obtained.

In other embodiments, the above-described policy return reason classification results of the policy return processing model may also be interpreted using model-independent methods. The model-independent approach can be flexibly applied to any classification model, for example, the model-independent approach can be applied to random forest models, deep neural network models, and the like. The explanation calculation of the policy rejection processing model by the model independent method may be calculated based on the policy rejection reason classification result predicted by the policy rejection processing model, or may be calculated in the process of calculating the policy rejection reason classification result by the policy rejection processing model, which is not limited herein. Fig. 4 is a flowchart of another embodiment of a method for processing a return order provided by the present application. Fig. 4 is different from fig. 3 in that the order rejection processing method shown in fig. 4 may further include step S204.

In step S204, based on the policy return processing model and the policy return event vector, the contribution value of each element in the policy return event vector to the policy return reason is obtained.

The destage processing model and the destage event vector may be used as inputs to calculate a destage reason contribution value. Based on the input returned order processing model and returned order event vector, the contribution value of each element to each returned order reason can be obtained.

Each element in the backspacing event vector has a corresponding sample vocabulary in the vocabulary library, and the value of each element can represent the frequency of the sample vocabulary corresponding to the element appearing in the description content of the backspacing event. The description vocabulary corresponding to each element in the return event vector is the same as the sample vocabulary corresponding to each element in the return event vector in the vocabulary library. The contribution value of the element in the backspace event vector to the backspace reason is the contribution value of the sample vocabulary corresponding to the element in the backspace event vector to the backspace reason.

The sample vocabulary corresponding to the elements in the backspace event vector has positive, negative or zero contribution to the backspace reason. The contribution values of the sample vocabularies corresponding to the elements to the order-returning reason are added, and the predicted value of the order-returning reason of the order-returning event can be obtained.

For example, the first reason for the return is predicted to be-0.98723173. The contribution value of the 239 th sample word to the first reason for the receipt is-2.008363, the contribution value of the 85 th sample word to the first reason for the receipt is 0.5134999, and the predicted value of the receipt processing model to the first reason for the receipt is-0.98723173 by adding the contribution values of the sample words corresponding to the elements in the receipt event vector. This indicates that 2.008363 is reduced in the predicted value of the first reason for returning a bill when the number 239 word appears 3 times, and 0.5134999 is added to the predicted value of the first reason for returning a bill when the number 85 word appears 1 time.

FIG. 5 is a diagram illustrating an example of contribution values of sample vocabularies to a reason for returning orders according to an embodiment of the present disclosure. As shown in FIG. 5, the predicted value of the reason for the return is about-1. The contribution value bar on the left side of-1 is a positive contribution and the contribution value bar on the right side of-1 is a negative contribution. Each section of the contribution value strip corresponds to a sample vocabulary, the length of the contribution value strip represents the size of the contribution value, and the longer the length of the contribution value strip is, the greater the contribution value of the sample vocabulary to the predicted value of the reason for returning the order is. The contribution value bars on the left side of-1 represent the contribution value of sample vocabulary # 85, the contribution value of sample vocabulary # 277, the contribution value of sample vocabulary # 96, the contribution value of sample vocabulary # 150, … …, respectively; the contribution value bar on the right side of-1 characterizes the contribution value of sample word No. 437, the contribution value of sample word No. 458, the contribution value of sample word No. 463, the contribution value of sample word No. 245, … …, respectively.

The reason for the return can be explained by the corresponding sample vocabularies with the largest positive contributions and the sample vocabularies with the largest negative contributions of each return reason. The rechecking personnel can recheck the reason for the receipt according to the contribution value of each element in the event vector for the receipt, thereby avoiding the analysis work of manually repeating the reason for the receipt and improving the rechecking efficiency.

The following description will take the example that the order-returning processing model includes a decision tree and the SHAP algorithm is adopted to explain the sorting result of the order-returning reason of the order-returning processing model. Fig. 6 is a flowchart of another embodiment of a method for processing a return order. Fig. 6 is different from fig. 4 in that step S204 in fig. 4 can be specifically subdivided into steps S2041 to S2043 in fig. 6.

In step S2041, a target vocabulary set from which target vocabularies are removed is determined based on the first vocabulary set.

The first vocabulary set is a set comprising description vocabularies corresponding to elements in the event vector of the return order. The target vocabulary is a description vocabulary corresponding to any element in the backspace event vector.

For example, the first vocabulary set N ═ a1, a2, a3, a4, a1, a2, a3, and a4 are description vocabularies corresponding to elements in the singleton vector. In the case where the target vocabulary is a2, the set of target vocabularies may be

Namely, one of the null sets, { a1}, { a3}, { a4}, { a1, a3}, { a1, a4}, { a3, a4}, and { a1, a3, a4 }.

In step S2042, a first average value and a second average value corresponding to the target vocabulary in the event vector of the leave are calculated according to the values of the leaf nodes in the decision tree and the weights of the edges corresponding to the leaf nodes.

The first average value is the average value of the values of the target vocabulary set in the decision tree. The second average value is the average value of the target vocabulary set and the union of the target vocabularies in the decision tree.

According to the values of the elements in the single event vector, the leaf node of the decision tree where the single event vector falls and the decision path of the single event vector in the decision tree can be determined. The decision path of the single event vector in the decision tree is the path from the root node to the leaf node of the decision tree where the single event vector falls. The weight of each edge in the decision tree may be determined from top to bottom, starting from the root node of the decision tree.

And when the vocabulary corresponding to the node of the policy-returning event vector in the decision tree is in the second vocabulary set, setting the weight of the edge corresponding to the decision path as 1 and setting the weight of the edge corresponding to the non-decision path as 0. And under the condition that the vocabulary corresponding to the node of the policy returning event vector in the decision tree is not in the second vocabulary set, traversing the left and right nodes of the node. The left node of the node may be set to a quotient of the number of destage sample vectors falling at the left node and the number of destage sample vectors falling at the parent node of the left node. The right node of the node may be set to a quotient of the number of destage sample vectors falling at the right node and the number of destage sample vectors falling at the parent node of the right node.

And taking the product of the value of the leaf node and the weight of the edge corresponding to the leaf node as the value of the parent node of the leaf node. Taking the product of the value of the father node and the weight of the edge corresponding to the father node as the value of the father node of the upper layer, and so on until the value of the root node of the decision tree is obtained, wherein the value of the root node of the decision tree is the average value. If the second vocabulary set is combined into the target vocabulary set in the process of calculating the value of the root node, the value of the root node is the first average value. If the second vocabulary set is a union of the target vocabulary set and the target vocabulary in the process of calculating the value of the root node, the value of the root node is the second average value.

In step S2043, a contribution value of the target vocabulary to the reason for the return is determined based on the first average, the second average, the number of elements in the target vocabulary set, and the number of elements in the first vocabulary set.

And obtaining the difference value of the contribution value of the description vocabulary corresponding to the elements of the backspacing event vector to the backspacing reason under the condition that the description vocabulary comprises the target vocabulary and under the condition that the description vocabulary does not comprise the target vocabulary. Specifically, the contribution value of the target vocabulary i to the reason for returning the order can be calculated according to equation (9):

wherein S is a target vocabulary set, N is a first vocabulary set, M is the number of elements in the first vocabulary set, i is a target vocabulary, | S | is the number of elements in the target vocabulary set, f_x(S) is a first average value, f_x(S { I }) is a second average value.

The application also provides a training device for the return processing model. Fig. 7 is a schematic structural diagram of an embodiment of a return processing model training apparatus provided in the present application. As shown in FIG. 7, the return processing model training apparatus 300 may include a vector transformation module 301, an initialization module 302, and a training module 303.

The vector conversion module 301 may be configured to convert the description content of the logged sample event into a sample vector.

The initialization module 302 may be configured to set model parameters of the classification model, and initialize the classification model according to the model parameters.

The training module 303 may be configured to select at least part of the order-returning sample vectors as a training set, and perform iterative training on the classification model to obtain an order-returning processing model.

The input of the order-returning processing model comprises an order-returning sample vector, the output of the order-returning processing model comprises an order-returning reason classification result, and the order-returning reason classification result is used for representing the order-returning reason of the order-returning sample event predicted by the order-returning processing model.

In some examples, the rollout processing model includes a decision tree.

The training module 303 may be configured to obtain an objective function of the classification model according to a first partial derivative sum, a second partial derivative sum, the number of leaf nodes of the decision tree, and a regularization term constant parameter, where the first partial derivative sum is a sum of first-order partial derivatives of a loss function of a first objective vector with respect to the classification model of a previous iteration, the first objective vector includes a demotion sample vector corresponding to a leaf node of the decision tree, and the second partial derivative sum is a sum of second-order partial derivatives of the loss function of the first objective vector with respect to the classification model of the previous iteration; and adjusting model parameters, and taking the classification model corresponding to the objective function with the expected value as a receipt-returning processing model.

In some examples, the vector transformation module 301 may be configured to normalize the description of the logged destage sample event; performing word segmentation on the description content of the return sample event after the normalization processing to obtain a sample word in the description content of the return sample event; and generating a destaging sample vector according to the occurrence frequency of the sample vocabularies in the vocabulary library in the description content of the destaging event.

Further, the vector conversion module 301 may be configured to convert the input unstructured file or semi-structured file recorded with the description content of the receipt sample event into a standard template recorded with the description content of the receipt sample event; analyzing a standard template in which the description content of the order-returning sample event is recorded to obtain structured data corresponding to the description content of the order-returning sample event; and carrying out data cleaning on the structured data to obtain the description content of the return order sample event after the standardized processing.

Fig. 8 is a schematic structural diagram of another embodiment of the order rejection processing model training apparatus provided in the present application. FIG. 8 is different from FIG. 7 in that the return processing model training apparatus 300 shown in FIG. 8 may further include a vocabulary library generating module 304 and a testing module 305.

The vocabulary library generating module 304 may be configured to calculate a word frequency and an inverse text frequency index for each sample vocabulary; obtaining an importance index of each sample vocabulary according to the word frequency and the inverse text frequency index of each sample vocabulary; and according to the sequence of the importance indexes from large to small, taking a preset number of sample vocabularies to form a vocabulary library.

The test module 305 may be configured to test the de-ordering processing model using a portion of the de-ordering sample vectors, while selecting another portion of the de-ordering sample vectors to iteratively train the classification model.

The application provides a receipt processing device. Fig. 9 is a schematic structural diagram of an embodiment of an order rejection processing apparatus provided in the present application. As shown in fig. 9, the order rejection processing apparatus 400 may include a vector conversion module 401, a calculation module 402, and an output module 403.

The vector conversion module 401 may be configured to convert the description content of the entered policy returning event into a policy returning event vector;

the calculation module 402 may be configured to use the policy return event vector as an input of a preset policy return processing model to obtain a policy return reason classification result output by the policy return processing model.

The sorting result of the order-returning reasons is used for representing the order-returning reasons of the order-returning events predicted by the order-returning processing model, and the order-returning processing model is a classification model obtained by training through a classification algorithm based on the order-returning sample vectors.

The output module 403 may output the reason for the invoice of the invoice event according to the reason for invoice classification result.

In some examples, the vector conversion module 401 may be configured to perform normalization processing on the description content of the logged leave-single event; performing word segmentation on the description content of the return event after the normalization processing to obtain a description vocabulary in the description content of the return event; and generating a return event vector according to the occurrence frequency of the description vocabularies in the preset vocabulary library in the description content of the return event.

Further, the vector conversion module 401 may be configured to convert the input unstructured file or semi-structured file recorded with the description content of the receipt-returning event into a standard template recorded with the description content of the receipt-returning event; analyzing a standard template in which the description content of the order-returning event is recorded to obtain structured data corresponding to the description content of the order-returning event; and carrying out data cleaning on the structured data to obtain the description content of the return order event after the standardized processing.

In some examples, the reason for.

In other examples, the reason for returning the order classification result includes the reason identification codes of a plurality of reason for returning the order and the predicted value of each reason for returning the order. The predicted value is used for representing the matching degree of the order-returning reason and the order-returning event.

Fig. 10 is a schematic structural diagram of another embodiment of an order rejection processing apparatus provided in the present application. Fig. 10 is different from fig. 9 in that the order rejection processing apparatus 400 shown in fig. 10 may further include a contribution analyzing module 404.

The contribution analysis module 404 may be configured to obtain a contribution value of each element in the return event vector to the return reason based on the return processing model and the return event vector.

In some examples, the rollout processing model includes a decision tree.

Further, the contribution analysis module 404 may be configured to determine a target vocabulary set for removing the target vocabulary based on a first vocabulary set, where the first vocabulary set is a set including description vocabularies corresponding to elements in the return event vector, and the target vocabulary is a description vocabulary corresponding to any element in the return event vector; calculating to obtain a first average value and a second average value corresponding to the target vocabulary in the policy withdrawing event vector according to the value of the leaf node in the policy making tree and the weight of the edge corresponding to the leaf node, wherein the first average value is the average value of the target vocabulary set in the policy making tree, and the second average value is the average value of the target vocabulary set and the target vocabulary union in the policy making tree; and determining the contribution value of the target vocabulary to the reason for returning the order based on the first average value, the second average value, the number of elements in the target vocabulary set and the number of elements in the first vocabulary set.

The order rejection processing model training device 300 and the order rejection processing device 400 in the above embodiments may also be implemented as the same device, and are not limited herein.

The application also provides a training device for the return processing model. Fig. 11 is a schematic structural diagram of an embodiment of a return processing model training apparatus provided in the present application. As shown in fig. 11, the destaging process model training apparatus 500 includes a memory 501, a processor 502, and a computer program stored on the memory 501 and executable on the processor 502.

In one example, the processor 502 may include a Central Processing Unit (CPU), or an Application Specific Integrated Circuit (ASIC), or may be configured to implement one or more Integrated circuits of the embodiments of the present Application.

The Memory 501 may include Read-Only Memory (ROM), Random Access Memory (RAM), magnetic disk storage media devices, optical storage media devices, flash Memory devices, electrical, optical, or other physical/tangible Memory storage devices. Thus, in general, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., a memory device) encoded with software comprising computer-executable instructions and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the rollout process model training method according to the present application.

The processor 502 runs a computer program corresponding to the executable program code by reading the executable program code stored in the memory 501, for implementing the return processing model training method in the above-described embodiment.

In one example, the drop off processing model training device 500 may also include a communication interface 503 and a bus 504. As shown in fig. 11, the memory 501, the processor 502, and the communication interface 503 are connected to each other via a bus 504 to complete communication therebetween.

The communication interface 503 is mainly used for implementing communication between modules, apparatuses, units and/or devices in the embodiments of the present application. Input devices and/or output devices may also be accessed through communication interface 503.

The bus 504 includes hardware, software, or both to couple the components of the destaging processing model training device 500 to each other. By way of example, and not limitation, Bus 504 may include an Accelerated Graphics Port (AGP) or other Graphics Bus, an Enhanced Industry Standard Architecture (EISA) Bus, a Front-Side Bus (FSB), a HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) Bus, an InfiniBand interconnect, a Low Pin Count (LPC) Bus, a memory Bus, a Micro Channel Architecture (MCA) Bus, a Peripheral Component Interconnect (PCI) Bus, a PCI-Express (PCI-X) Bus, a Serial Advanced Technology Attachment (SATA) Bus, a Video Electronics Standards Association Local Bus (VLB) Bus, or other suitable Bus, or a combination of two or more of these. Bus 504 may include one or more buses, where appropriate. Although specific buses are described and shown in the embodiments of the application, any suitable buses or interconnects are contemplated by the application.

The embodiment of the application also provides an order-returning processing device. Fig. 12 is a schematic structural diagram of an embodiment of an order rejection processing apparatus provided in the present application. As shown in fig. 12, the order rejection processing apparatus 600 includes a memory 601, a processor 602, and a computer program stored on the memory 601 and executable on the processor 602.

In one example, the processor 602 may include a Central Processing Unit (CPU), or an Application Specific Integrated Circuit (ASIC), or may be configured to implement one or more Integrated circuits of the embodiments of the present Application.

The Memory 601 may include Read-Only Memory (ROM), Random Access Memory (RAM), magnetic disk storage media devices, optical storage media devices, flash Memory devices, electrical, optical, or other physical/tangible Memory storage devices. Thus, in general, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software comprising computer-executable instructions and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the rollout processing methods according to the present application.

The processor 602 runs a computer program corresponding to the executable program code by reading the executable program code stored in the memory 601 for implementing the order rejection processing method in the above-described embodiment.

In one example, the de-ordering processing device 600 may also include a communication interface 603 and a bus 604. As shown in fig. 12, the memory 601, the processor 602, and the communication interface 603 are connected via a bus 604 to complete communication therebetween.

The communication interface 603 is mainly used for implementing communication between modules, apparatuses, units and/or devices in the embodiments of the present application. Input devices and/or output devices are also accessible through communication interface 603.

The bus 604 includes hardware, software, or both to couple the components of the destaging processing device 600 to each other. By way of example, and not limitation, Bus 604 may include an Accelerated Graphics Port (AGP) or other Graphics Bus, an Enhanced Industry Standard Architecture (EISA) Bus, a Front-Side Bus (FSB), a HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) Bus, an InfiniBand interconnect, a Low Pin Count (LPC) Bus, a memory Bus, a Micro Channel Architecture (MCA) Bus, a Peripheral Component Interconnect (PCI) Bus, a PCI-Express (PCI-X) Bus, a Serial Advanced Technology Attachment (SATA) Bus, a Video Electronics Standards Association Local Bus (VLB) Bus, or other suitable Bus, or a combination of two or more of these. Bus 604 may include one or more buses, where appropriate. Although specific buses are described and shown in the embodiments of the application, any suitable buses or interconnects are contemplated by the application.

The order rejection processing model training apparatus 500 and the order rejection processing apparatus 600 in the above embodiments may also be implemented as the same apparatus, and are not limited herein.

An embodiment of the present application further provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the method for processing a drop-out order and/or the method for processing a drop-out order in the foregoing embodiments can be implemented, and the same technical effect can be achieved. The computer-readable storage medium may include a non-transitory computer-readable storage medium, such as a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and the like, which is not limited herein.

It should be clear that the embodiments in this specification are described in a progressive manner, and the same or similar parts in the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. For apparatus embodiments, device embodiments, computer-readable storage medium embodiments, reference may be made in the descriptive section to method embodiments. The present application is not limited to the particular steps and structures described above and shown in the drawings. Those skilled in the art may make various changes, modifications and additions or change the order between the steps after appreciating the spirit of the present application. Also, a detailed description of known process techniques is omitted herein for the sake of brevity.

Aspects of the present application are described above with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions/acts specified in the flowchart and/or block diagram block or blocks. Such a processor may be, but is not limited to, a general purpose processor, a special purpose processor, an application specific processor, or a field programmable logic circuit. It will also be understood that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware for performing the specified functions or acts, or combinations of special purpose hardware and computer instructions.

It will be appreciated by persons skilled in the art that the above embodiments are illustrative and not restrictive. Different features which are present in different embodiments may be combined to advantage. Other variations to the disclosed embodiments can be understood and effected by those skilled in the art upon studying the drawings, the specification, and the claims. In the claims, the term "comprising" does not exclude other means or steps; the word "a" or "an" does not exclude a plurality; the terms "first" and "second" are used to denote a name and not to denote any particular order. Any reference signs in the claims shall not be construed as limiting the scope. The functions of the various parts appearing in the claims may be implemented by a single hardware or software module. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

Claims

1. A method for processing an order return, comprising:

converting the description content of the recorded order-returning event into an order-returning event vector;

taking the order-returning event vector as the input of a preset order-returning processing model to obtain an order-returning reason classification result output by the order-returning processing model, wherein the order-returning reason classification result is used for representing the order-returning reason of the order-returning event predicted by the order-returning processing model, and the order-returning processing model is a classification model obtained by training through a classification algorithm based on the order-returning sample vector;

and outputting the reason for the.

2. The method according to claim 1, wherein the converting the description of the logged de-ordering event into a de-ordering event vector comprises:

carrying out normalized processing on the recorded description content of the order refunding event;

performing word segmentation on the description content of the return event after the normalization processing to obtain a description vocabulary in the description content of the return event;

and generating the return event vector according to the occurrence frequency of the description vocabulary in a preset vocabulary library in the description content of the return event.

3. The method according to claim 2, wherein the normalizing the description of the entered rollout event comprises:

converting the input unstructured file or semi-structured file recorded with the description content of the order-returning event into a standard template recorded with the description content of the order-returning event;

analyzing a standard template recorded with the description content of the order-returning event to obtain structured data corresponding to the description content of the order-returning event;

and carrying out data cleaning on the structured data to obtain the description content of the order-returning event after the standardized processing.

4. The method of claim 1,

the reason sorting result comprises the reason identification code of the reason of the order returning event;

alternatively, the first and second electrodes may be,

the reason identification codes of a plurality of reason for returning the order and the predicted value of each reason for returning the order are included in the classification result of the reason for returning the order, and the predicted value is used for representing the matching degree of the reason for returning the order and the event for returning the order.

5. The method of claim 1, further comprising:

and obtaining the contribution value of each element in the return order event vector to the return order reason based on the return order processing model and the return order event vector.

6. The method of claim 5, wherein the rollout processing model comprises a decision tree;

the obtaining of the contribution value of each element in the return order event vector to the return order reason based on the return order processing model and the return order event vector comprises:

determining a target vocabulary set for removing target vocabularies based on a first vocabulary set, wherein the first vocabulary set is a set comprising description vocabularies corresponding to elements in the return event vector, and the target vocabularies are the description vocabularies corresponding to any element in the return event vector;

calculating to obtain a first average value and a second average value corresponding to the target vocabulary in the policy returning event vector according to the value of the leaf node in the policy making tree and the weight of the edge corresponding to the leaf node, wherein the first average value is the average value of the target vocabulary set in the policy making tree, and the second average value is the average value of the target vocabulary set and the target vocabulary union in the policy making tree;

determining a contribution value of the target vocabulary to a reason for returning the order based on the first average value, the second average value, the number of elements in the target vocabulary set and the number of elements in the first vocabulary set.

7. A method for training a return processing model is characterized by comprising the following steps:

converting the description content of the recorded order-returning sample event into an order-returning sample vector;

setting model parameters of a classification model, and initializing the classification model according to the model parameters;

selecting at least part of the order-returning sample vectors as a training set, and performing iterative training on the classification model to obtain an order-returning processing model, wherein the input of the order-returning processing model comprises the order-returning sample vectors, the output of the order-returning processing model comprises an order-returning reason classification result, and the order-returning reason classification result is used for representing the order-returning reasons of the order-returning sample events predicted by the order-returning processing model.

8. The method of claim 7, wherein the rollout processing model comprises a decision tree;

selecting at least part of the order-returning sample vectors as a training set, and performing iterative training on the classification model to obtain an order-returning processing model, wherein the method comprises the following steps:

obtaining an objective function of the classification model according to a first partial derivative sum, a second partial derivative sum, the number of leaf nodes of the decision tree and a regularization term constant parameter, wherein the first partial derivative sum is the sum of first-order partial derivatives of a loss function of a first objective vector about the classification model of the previous iteration, the first objective vector comprises the regression sample vector corresponding to the leaf node of the decision tree, and the second partial derivative sum is the sum of second-order partial derivatives of the loss function of the first objective vector about the classification model of the previous iteration;

and adjusting the model parameters, and taking the classification model corresponding to the objective function with the expected value as the order rejection processing model.

9. The method according to claim 7, wherein the converting the description of the logged destage sample event into a destage sample vector comprises:

carrying out normalized processing on the description content of the entered refund sample event;

performing word segmentation on the description content of the return sample event after the normalization processing to obtain a sample word in the description content of the return sample event;

and generating the order-returning sample vector according to the occurrence frequency of the sample words in the vocabulary library in the description content of the order-returning event.

10. The method according to claim 9, wherein the normalizing the description of the entered refund sample event comprises:

converting the input unstructured file or semi-structured file recorded with the description content of the receipt sample event into a standard template recorded with the description content of the receipt sample event;

analyzing a standard template in which the description content of the order-returning sample event is recorded to obtain structured data corresponding to the description content of the order-returning sample event;

and carrying out data cleaning on the structured data to obtain the description content of the order-returning sample event after the normalized processing.

11. The method according to claim 9, wherein after the tokenizing the normalized description content of the return sample event to obtain the sample vocabulary in the description content of the return sample event, further comprising:

calculating the word frequency and the inverse text frequency index of each sample vocabulary;

obtaining an importance index of each sample vocabulary according to the word frequency and the inverse text frequency index of each sample vocabulary;

and according to the sequence of the importance indexes from large to small, taking a preset number of sample vocabularies to form the vocabulary library.

12. The method of claim 7, further comprising:

and under the condition that one part of the order-withdrawing sample vectors are selected to carry out iterative training on the classification model, testing the order-withdrawing processing model by using the other part of the order-withdrawing sample vectors.

13. An order return processing apparatus, comprising:

the vector conversion module is used for converting the recorded description content of the order returning event into an order returning event vector;

the calculation module is used for taking the order rejection event vector as the input of a preset order rejection processing model to obtain an order rejection reason classification result output by the order rejection processing model, the order rejection reason classification result is used for representing the order rejection reason of the order rejection event predicted by the order rejection processing model, and the order rejection processing model is a classification model obtained by training through a classification algorithm based on the order rejection sample vector;

and the output module outputs the reason for returning the order of the order-returning event according to the classification result of the reason for returning the order.

14. A training device for a return processing model, comprising:

the vector conversion module is used for converting the description content of the recorded order-returning sample event into an order-returning sample vector;

the initialization module is used for setting model parameters of a classification model and initializing the classification model according to the model parameters;

the training module is used for selecting at least part of the order-returning sample vectors as a training set, performing iterative training on the classification model, and obtaining an order-returning processing model, wherein the input of the order-returning processing model comprises the order-returning sample vectors, the output of the order-returning processing model comprises an order-returning reason classification result, and the order-returning reason classification result is used for representing the order-returning reasons of the order-returning sample events obtained through prediction of the order-returning processing model.

15. An order return processing apparatus, characterized in that the apparatus comprises: a processor and a memory storing computer program instructions;

the processor, when executing the computer program instructions, implements a method of de-ordering in accordance with any of claims 1 to 6.

16. An order-return processing model training apparatus, the apparatus comprising: a processor and a memory storing computer program instructions;

the processor, when executing the computer program instructions, implements the drop-out processing model training method of any of claims 7 to 12.

17. A computer-readable storage medium having computer program instructions stored thereon, which when executed by a processor implement the method of processing a drop-out order as recited in any one of claims 1 to 6 or the method of training a drop-out order model as recited in any one of claims 7 to 12.