WO2019154108A1

WO2019154108A1 - Method and apparatus for processing transaction data

Info

Publication number: WO2019154108A1
Application number: PCT/CN2019/073104
Authority: WO
Inventors: 赵科科; 赵星
Original assignee: 阿里巴巴集团控股有限公司
Priority date: 2018-02-12
Filing date: 2019-01-25
Publication date: 2019-08-15
Also published as: TW201935386A; CN108446978A

Abstract

A method and apparatus for processing transaction data. The method comprises: first obtaining n data sets respectively corresponding to continuous n time periods, wherein each data set i comprises transaction detail data of a user in a corresponding time period; then obtaining a derivation variable on the basis of the transaction detail data in the corresponding data set i, and forming, on the basis of the derivation variable, a feature factor corresponding to each data set; and on this basis, inputting each feature factor into a time-recursive neural network in a time sequence, and obtaining a processing result from the neural network. Therefore, transaction data is mined and analyzed more effectively.

Description

Method and device for processing transaction data

Cross-reference to related applications

The present application claims priority to Chinese Patent Application No. 201101146777.7, filed on Feb. 12, 20, the entire disclosure of which is incorporated herein by reference. in.

Technical field

One or more embodiments of the present specification relate to the field of computer technology, and more particularly to a method and apparatus for processing transaction data.

Background technique

With the development of Internet technology, people use the Internet and e-wallets more and more frequently for various transactions, thereby forming transaction data. Transaction data is a high-value data asset. Especially in the current big data background, how to dig deep into transaction data and extract the value of data has important significance in technology improvement and business improvement.

Further, while deep digging data, it is also necessary to consider application scenarios and security issues. The transaction data generally reflects the user's transaction history. If the transaction data is mined and processed, the transaction data information can be applied to a wider range of application scenarios, such as credit business scenarios, which will further improve data utilization. In addition, in many cases, there is the possibility of co-building models with other agencies, which requires the initial processing of data to be sent to other agencies. At this point, it is hoped that the data to be mined has higher data value and data meaning, and the system risk of data leakage and the protection of user privacy should be considered, and the business meaning should be hidden as much as possible. In this way, high requirements are placed on data mining and processing.

Therefore, it is hoped that there will be an improved solution to process transaction data more efficiently.

Summary of the invention

One or more embodiments of the present specification describe a method and apparatus for processing transaction data more efficiently by combining preliminary data mining derived from variables and further data analysis of neural networks.

According to a first aspect, a method of processing transaction data is provided, comprising:

Obtaining n data sets respectively corresponding to consecutive n preset time segments, wherein each data set i includes transaction detail data of the user in the corresponding time period;

Forming n feature vectors respectively corresponding to the n data sets, wherein each feature vector Fi includes a derivative variable derived based on transaction detail data in the corresponding data set i;

The n feature vectors are input to the time recursive neural network in chronological order, and the processing results are obtained from the time recursive neural network.

According to one embodiment, the transaction detail data includes a plurality of fields including at least a transaction time field, a transaction amount field, and at least one category field.

In a possible design, the step of forming a feature vector includes: acquiring the plurality of fields of the transaction detail data in the data set i; performing an aggregation operation on the data of the plurality of fields to obtain a derivative variable; The derived variable is used as a vector element of the feature vector Fi.

According to an embodiment, the aggregating the data in the plurality of fields comprises: selecting at least a part of the plurality of fields to be combined to obtain a combined field; and performing an operation operation on the data of the combined field to obtain a derived variable.

Further, in an embodiment, the operation operation includes one or more of the following: numerical value judgment, counting, summation, averaging, standard deviation, grading, and distribution statistics.

According to a possible design, the step of forming a feature vector further comprises: acquiring content of the at least one category field in the data set i; converting the content of the at least one category field into a word vector by using a word embedding model; The predicate vector is part of the feature vector Fi.

In one embodiment, the time recursive neural network employs one of a recurrent neural network RNN, a long-term and short-term memory neural network LSTM, and a gated loop unit neural network GRU.

In one embodiment, the time recursive neural network further includes at least one fully connected layer.

According to one embodiment, the time recursive neural network is trained using a calibrated data set that includes historical transaction data and has a label for whether a credit default has occurred.

Accordingly, in one embodiment, obtaining the processing result from the time recursive neural network includes obtaining, from the output layer of the time recursive neural network, a probability that the user has a credit default as a processing result.

Moreover, in one embodiment, obtaining the processing result from the time recursive neural network may further comprise: obtaining the node feature value from the hidden layer of the neural network as a processing result.

According to a second aspect, an apparatus for processing transaction data is provided, comprising:

a data set obtaining unit, configured to acquire n data sets respectively corresponding to consecutive n preset time segments, where each data set i of the n data sets includes transaction detail data of the user in the corresponding time period;

a vector forming unit configured to form n feature vectors respectively corresponding to the n data sets, wherein each feature vector Fi includes a derivative variable derived based on transaction detail data in the corresponding data set i;

And a processing unit configured to input the n feature vectors into the time recursive neural network in chronological order, and obtain the processing result from the time recursive neural network.

According to a third aspect, there is provided a computer readable storage medium having stored thereon a computer program for causing a computer to perform the method of the first aspect when the computer program is executed in a computer.

According to a fourth aspect, there is provided a computing device comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, implementing the method of the first aspect .

Through the method and device provided by the embodiments of the present specification, variable derivative of the transaction detail data is firstly performed, preliminary data mining is performed, and then the feature vector based on the derivative variable is input to the neural network for further processing, which significantly improves the network performance, and It can be applied to a variety of scenarios according to network training conditions. In addition, user privacy and security can be guaranteed.

DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly described below. It is obvious that the drawings in the following description are only some embodiments of the present invention, Those skilled in the art can also obtain other drawings based on these drawings without any creative work.

1 is a schematic diagram of an implementation scenario of an embodiment disclosed in the present specification;

2 shows a flow chart of a method of processing transaction data, in accordance with one embodiment;

Figure 3 illustrates a schematic diagram of transaction detail data in accordance with one embodiment;

4 shows a process diagram of a neural network in accordance with one embodiment;

FIG. 5 shows a schematic diagram of processing of a neural network according to another embodiment;

Figure 6 shows a schematic block diagram of a processing device in accordance with one embodiment.

Detailed ways

The solution provided in this specification will be described below with reference to the accompanying drawings.

FIG. 1 is a schematic diagram of an implementation scenario of an embodiment disclosed in the present specification. As shown in the figure, firstly, the historical transaction detail data of the user within a long period of time, for example, one year, is obtained, and the longer time period is equally divided into consecutive time periods, for example, 12 months. These historical data are organized into corresponding data sets S1, S2, ... S12 according to respective time periods. For example, each month corresponds to a data set, and each data set contains transaction details of the user of the current month. Then, the transaction detail data in each data set is subjected to lateral variable derivation, that is, the transaction detail data in each data set is laterally cross-aggregated and the like, and is not vertically aggregated across the data set. Thus, derived variables are obtained for each data set, and then feature vectors F1, F2...F12 corresponding to the data set are formed based on these derived variables. The feature vectors corresponding to the respective data sets thus obtained are input into the time recursive neural network in chronological order, and the processing results are obtained from the neural network. In this way, the data mining derived from variables is combined with the data processing of the neural network, and the transaction data is processed more effectively.

2 shows a flow diagram of a method of processing transaction data, in accordance with one embodiment. The executive body of the method flow can be any device, device, platform or system with computing and processing capabilities, such as a server, and more specifically, for example, an Alipay server. As shown in FIG. 2, the method includes the following steps: Step 21: Obtain n data sets respectively corresponding to consecutive n preset time segments, where each data set i includes transaction detail data of a user in a corresponding time period; 22, forming n feature vectors respectively corresponding to the n data sets, wherein each feature vector Fi includes a derivative variable derived based on the transaction detail data in the corresponding data set i; step 23, the n features The vector inputs a time recursive neural network in chronological order, and the processing results are obtained from the time recursive neural network. The specific implementation of each of the above steps is described below.

First, in step 21, a transaction detail data set corresponding to each of the n consecutive preset time periods is acquired.

In one embodiment, the historical transaction detail data of the user within a long period of time is first obtained, and the longer time period is equally divided into shorter, consecutive n time periods, and the historical transaction detail data is obtained. According to each time period, it is organized into corresponding data sets.

In one example, the longer time period is 1 year, the shorter time period is month, and the continuous n time periods are 12 months. In another example, the longer time period described above is 3 months and the consecutive n time periods are 12 weeks. In other examples, the length of the time period and the size of n can be set according to actual business needs.

Correspondingly, the data set corresponding to each time period includes the transaction detail data of the user during the time period. For example, in the case where the above time period is a month, each month corresponds to one data set, and the data set contains the transaction detail data of the user in the month.

In one example, the transaction detail data is embodied in several transaction records. And, the transaction detail data can include multiple fields. Generally, these fields include at least: a transaction time field, a transaction amount field. Typically, the transaction data will also include the transaction location field. In one embodiment, the transaction data further includes at least one category field. The category fields include, for example, a seller category, a product category, an order source category, and the like.

FIG. 3 shows a schematic diagram of transaction detail data in accordance with one embodiment. In the example of Figure 3, the transaction detail data contains multiple transaction records and contains the following fields: transaction time, transaction location, transaction amount, and three category fields, namely the seller category, the product category, and the order category. The value of the category field is determined according to a preset rule, for example, the commodity is divided into food (Class A), clothing (Class B), book audio (C), household items (D), virtual goods (E) Class) A total of 5 classes, the order category is divided into online (On) class and offline class (off), and so on. It can be understood that FIG. 3 is just an example. Transaction detail data can take other formats, including more, fewer, or other field content, depending on the business needs.

On the basis of acquiring the transaction detail data sets corresponding to the respective time segments, in step 22, feature vectors corresponding to the respective data sets are formed based on the derived variables of the transaction detail data in the respective data sets.

Specifically, in one embodiment, in order to form the feature vector Fi for the data set i, first acquiring a plurality of fields of the transaction detail data in the data set i, and then performing aggregation operations on the data of the plurality of fields, thereby obtaining a derivative variable; Next, the derived variable is taken as a vector element of the feature vector Fi, thereby forming a feature vector Fi based on the derived variable. The acquisition of the above derived variables is actually a horizontal variable derivation, that is, a derivative operation such as horizontal cross-aggregation for the transaction detail data in each data set, without vertically merging across the data set. For example, in the case where each data set corresponds to one month, the horizontal variable derivation is only aggregated for the transaction details of the user within the one month, and the vertical and the next month are not crossed across the data set. derivative.

More specifically, in one embodiment, the horizontal variable derivation specifically includes: selecting at least a part of the plurality of fields of the transaction detail data to be combined to obtain a combined field; and performing an operation operation on the data of the combined field to obtain a derivative variable . In one embodiment, the arithmetic operations include: numerical determination, counting, summation, averaging, standard deviation, quantification, distribution statistics, and the like.

Table 1 shows examples of derived variables and their derivatives.

Table 1

As shown in Table 1, in the variable derivation process, it is possible to select only one field of the transaction detail data (in this case, the combination field is the field itself), and perform arithmetic operations such as variables X1 and X2; Select at least two fields of the transaction detail data, combine them, and perform operations on the data of the combined fields to obtain derivative variables. Specifically, in the example of Table 1, the derived variable X1 represents the total amount of the transaction. In the case where a data set corresponds to one month, X1 represents the total amount of transactions of the user within one month. The process of obtaining the derived variable is that, in the data set corresponding to one month, the amount field is selected, and the data in the field is summed. For another example, the derivative variable X2 represents the average amount of transactions in a month. The process of obtaining the derived variable X2 may include selecting an amount field and averaging the data in the field.

X3-X6 is a derived variable determined based on a combination of multiple fields. Specifically, the derived variable X3 represents, for example, within one month, the total amount spent on the category A commodity. In order to obtain the derivative variable X3, the amount field of the transaction detail data is combined with the field of the commodity category, and the transaction record of the commodity category field A is determined by numerical judgment, and then the data of the amount field in the records is summed, The derived variable X3 is obtained. The derivative variable X4 represents the number of consumers in the offline two types of shops. In order to obtain the derivative variable X4, the seller category field and the order category field of the transaction detail data are combined, and the transaction record whose seller category field value is “second class” and the order category field value is “off” is selected by numerical judgment. And count these transaction records, thus obtaining X4. Other derived variables are similarly obtained through field combination and arithmetic operations. It should be noted that the derivative variable X6 represents the distribution of the number of purchases during the week. A little different from other derived variables that are embodied as a single value, such derived variables of the representation distribution can be represented in the form of sub-vectors of multiple values, for example, X6 can be expressed as (p1, p2, ... p7), Where pi represents the proportion of the number of pens purchased on the ith day of the week.

Using each derived variable as a vector element, a feature vector corresponding to the data set can be formed. For example, based on the derived variables of Table 1, a feature vector Fi = (X1, X2, X3, X4, X5, X6) can be formed. It can be understood that Table 1 is only an example. It is also possible to use more kinds of field combinations, and use more kinds of operations to perform lateral variable derivation in the same data set, thereby obtaining more derivative variables and forming feature vectors of more elements. .

In general, the value of the category field used for variable derivation should not be too large. For example, for the commodity category, the derivative variable X3 shown in Table 1 indicates the total amount of consumption in the category A commodity. If the case of different values in the same category is treated the same, then the total consumption of goods in Class B, Class C, ... Class E is also counted in correspondence with X3. If the value of the commodity category field is too large, such as 20 classes, then the number of derived variables may be too large. Therefore, variable derivation is generally based on a category field whose value is within a certain range.

In one embodiment, the transaction detail data in the data set includes a category field having a value greater than a predetermined threshold, such as twenty. For such a category field, in one embodiment, the content of the category field is converted to a word vector using a word embedding model, with the word vector as part of the feature vector of the data set.

It can be understood that the word embedding model is a model used in natural language processing NLP to convert a single word into a vector. In the simplest model, a set of features is constructed for each word as its corresponding vector. Further, in order to embody the relationship between words, such as category relationship, affiliation, the language model can be trained in various ways to optimize vector expression. For example, the word2vec tool contains a variety of word embedding methods, which can quickly get the vector expression of a word, and the vector expression can reflect the analogy between words. There are also other word embedding algorithms, such as one hot encoding algorithms.

The word field can be converted to a word vector by the word embedding model. The word vector can be spliced together with the derived variables to form a feature vector corresponding to the data set.

By performing the above lateral variable derivation for each data set i, and optional word embedding transformation, a corresponding feature vector Fi is formed for each data set. Then, next in step 23, the thus formed n feature vectors are input into the time recursive neural network in chronological order, and the processing result is obtained from the time recursive neural network.

In one embodiment, the time recursive neural network employs a Recurrent Neural Network (RNN). The recurrent neural network RNN is a typical time recurrent neural network that can be used to process sequence data. In the RNN, the current output of a sequence is associated with its previous output. Specifically, the RNN memorizes the previous information and applies it to the calculation of the current output, that is, the nodes between the hidden layers are connected, and the input of the hidden layer includes not only the output of the input layer but also the previous moment. The output of this hidden layer. That is to say, the t-th hidden layer state can be expressed as:

St=f(U*Xt+W*St-1)

Where Xt is the state of the tth input layer, St-1 is the t-1th hidden layer state, f is the calculation function, and W and U are the weights. As such, the RNN loops back to the current input, taking into account the timing effects of the input sequence.

In one embodiment, the time recursive neural network uses a Long Short Term Memory (LSTM). LSTM is an improved model based on the cyclic neural network RNN. Cyclic neural network RNN has long-term dependence problems when dealing with long-term memory, and training is difficult (such as the problem of gradient overflow). The LSTM model proposed on the basis of RNN can further solve the problem of long-term dependence.

In the repetitive network module of the LSTM model, three gate calculations are implemented, namely input gate, output gate and "forgotten gate". The "Forgetting Gate" setting allows information to be selectively passed, thereby discarding some information that is no longer needed, thus judging and shielding the input unnecessary interference information, thereby better analyzing and processing the long-term data.

In one embodiment, the time recursive neural network employs a Gated Recurrent Unit (GRU) neural network. The gated loop unit GRU neural network can be considered as a simplification and variant of the LSTM neural network. The GRU model modifies the settings of the input, output, and forgetting gates in the LSTM to two gates: the update gate and the reset gate. The update gate is used to control the degree to which the status information of the previous moment is brought into the current state. The larger the value of the update gate is, the more the status information is brought in at the previous moment. The reset gate is used to control the degree of ignoring the status information of the previous moment. The smaller the value of the reset gate, the more it is ignored. By updating the gate and resetting the gate, the control of the influence of the state information of the previous moment on the current state is realized.

There are some other variants of the circulatory neural network RNN and the long- and short-term memory neural network LSTM. Such variant models can be used as a time recursive neural network in the embodiment.

To further enhance the performance of the neural network, in one embodiment, the time recursive neural network includes a number of fully connected layers in addition to the temporal recursive layers implemented by the various models above. Each node in the fully connected layer is connected to all nodes of the upper layer to integrate the features extracted from the front. In this way, a comprehensive analysis and processing of the characteristics of the previously hidden layers is achieved.

The neural networks constructed in various ways by various models as described above are time recursive neural networks, and sequences having time series information can be processed. Correspondingly, in step 23, the feature vectors Fi corresponding to the respective data sets i are sequentially input to the above-described neural network in chronological order.

In one embodiment, the n feature vectors are sequentially input into the time recursive neural network according to the time sequence of the corresponding data set, and further analyzed.

In another embodiment, the n feature vectors are organized into a feature matrix according to the chronological order of the corresponding data sets. For example, assuming that each feature vector is an m-dimensional vector, n feature vectors can be organized into a matrix of m*n, and the row vector of the i-th row in the matrix corresponds to the feature vector Fi. Such a feature matrix is input to the above time recursive neural network. Since the row vectors in the feature matrix are organized in chronological order, the neural network can still obtain the timing relationship between the feature vectors, and then perform analysis processing.

4 shows a process diagram of a neural network in accordance with one embodiment. In the example of FIG. 4, a total of 12 data sets are acquired, each of which corresponds to one month of transaction detail data, and each data set i forms a corresponding feature vector Fi based on the derived variables. Moreover, the time recurrent neural network in Figure 4 uses a cyclic neural network RNN. As shown in FIG. 4, the feature vectors F1, F2, ..., F12 are sequentially input to the cyclic neural network RNN. The horizontal arrow between the RNNs in Fig. 4 indicates the transition of the time state, and H1-H12 indicates the state change of the RNN. After inputting F12, the cyclic neural network can output the processing result.

FIG. 5 shows a process diagram of a neural network in accordance with another embodiment. In the example of FIG. 5, the data sets respectively corresponding to one month, and the corresponding feature vectors F1-F12 are still employed, but the time recurrent neural network uses the long- and short-term memory neural network LSTM. Moreover, the neural network also includes a fully connected layer based on the LSTM architecture. In the example of Fig. 5, feature vectors F1, F2, ... F12 are sequentially input to the LSTM neural network, and H1-H12 indicate state changes of the LSTM. Moreover, the result of processing the LSTM to F12 is spliced again with F12 itself, that is, H12 and F12 are spliced together, and then input to the fully connected layer together. Finally, the processing result is output by the fully connected layer. Although in Fig. 5, the F12 and H12 are spliced and then input to the fully connected layer, in other embodiments, H12 can also be directly input to the fully connected layer without being spliced with F12.

It will be understood that Figures 4 and 5 are merely examples. Time recursive neural networks can be constructed in a variety of ways, including the above examples and other models that are not exhaustively enumerated.

In order for the constructed neural network to better process the transaction data, it is necessary to train the neural network in advance to optimize the network model parameters. In one embodiment, a two-class supervised learning algorithm is employed to train the neural network described above. Specifically, the time recursive neural network is trained using the calibrated data set, the calibrated data set including historical transaction data, and a tag having a credit default. More specifically, a data set containing historical transaction data is obtained, which may be from the same user or a different user. Correspondingly, the data set is derived from variables to form a sequence of feature vectors. On the other hand, a record of whether the user has a credit default in the period corresponding to the historical transaction data set is also acquired, and based on the record, the historical transaction data set is given a tag value indicating whether a credit default has occurred. A historical transaction data set with such a tag becomes a calibration data set that can be used to train a neural network. In the process of training, the calibration data set (especially the sequence of feature vectors) is processed by the neural network to give a prediction result of whether credit default will occur, and the prediction result is compared with the actual tag value, thereby calculating The loss function, the gradient transfer according to the loss function, the modification and optimization of the model parameters, and thus repeated training, thereby obtaining the trained neural network.

The neural network thus trained can be used to give a probability of a user's credit default by analyzing and processing the transaction data. That is, in step 23, when the n feature vectors corresponding to the n data sets are input to the trained, time recursive neural network, the neural network may output the probability that the user has a credit default as an output result. Accordingly, in one embodiment, obtaining the processing result from the time recursive neural network includes, from the output layer of the neural network, obtaining a probability that the user has a credit default as a processing result. In such cases, transaction data is applied to the business scenario of credit evaluation through variable derivative and neural network training.

In another embodiment, the node feature values may also be obtained from the hidden layer of the time recursive neural network as a result of the processing. It can be understood that, according to the construction manner of the neural network, the node feature values in some of the hidden layers also have strong data value and meaning. For example, some neural networks contain a bottleneck layer with a significantly reduced number of nodes. The node feature value of the bottleneck layer can be regarded as a low-dimensional representation of the input feature, which can reflect strong data logic and meaning. Other hidden layers near the output layer, the node feature values also have a certain data meaning. Therefore, the node feature values can also be extracted from the hidden layer of the neural network as a result of the processing. Such processing results can be used to input into further models, such as models built with other organizations, for further data analysis and processing.

Thus, in the above embodiment, variable derivative is first performed on the transaction detail data, preliminary data mining is performed, and then the feature vector based on the derivative variable is input to the neural network for further processing, which makes the network performance significantly improved and can be applied to Contains a variety of scenarios for credit business. In addition, since the processing of data by the neural network is a process of nonlinear transformation, the data thus processed has both clear data meaning and user privacy and security.

In another aspect, embodiments of the specification also provide an apparatus for processing transaction data. Figure 6 shows a schematic block diagram of a processing device in accordance with one embodiment. As shown in FIG. 6, the processing device 600 includes: a data set obtaining unit 610, configured to acquire n data sets respectively corresponding to consecutive n preset time segments, where each data set i of the n data sets includes a corresponding a transaction detail data of the user in the time period; the vector forming unit 620 is configured to form n feature vectors respectively corresponding to the n data sets, wherein each feature vector Fi includes, respectively, based on the transaction details in the corresponding data set i Derived variables derived from the data; and processing unit 630 configured to input the n feature vectors into the time recursive neural network in chronological order, and obtain the processing results from the time recursive neural network.

According to an embodiment, the transaction detail data includes a plurality of fields, the plurality of fields including at least: a transaction time field, a transaction amount field, and at least one category field.

In an embodiment, the vector forming unit 620 includes: a field obtaining module 621 configured to acquire the plurality of fields of the transaction detail data in the data set i; and an aggregation operation module 622 configured to the plurality of fields The data is subjected to an aggregation operation to obtain a derived variable; and an element forming module 623 is configured to use the derived variable as a vector element of the feature vector Fi.

According to an embodiment, the aggregating operation module 622 is further configured to: select at least a part of the plurality of fields to be combined to obtain a combined field; and perform an operation operation on the data of the combined field to obtain a derivative variable.

According to an embodiment, the vector forming unit 620 further includes a word embedding module 624 configured to: acquire content of the at least one category field in the data set i; convert the content of the at least one category field by using a word embedding model Is a word vector; the word vector is taken as part of the feature vector Fi.

According to one embodiment, the time recursive neural network described above employs one of a recurrent neural network RNN, a long- and short-term memory neural network LSTM, and a gated loop unit neural network GRU.

In one embodiment, the processing unit 630 is configured to obtain, from the output layer of the time recursive neural network, a probability that the user has a credit default as a result of the processing.

In one embodiment, processing unit 630 may also obtain node feature values from the hidden layer of the time recursive neural network as a result of the process.

Through the above device, variable derivative of transaction detail data is firstly carried out, preliminary data mining is performed, and then the feature vector based on the derivative variable is input into the neural network for further processing, which makes the network performance significantly improved and can be applied according to the network training situation. In a variety of scenarios. In addition, user privacy and security can be guaranteed.

According to another embodiment, there is also provided a computer readable storage medium having stored thereon a computer program for causing a computer to perform the method described in connection with FIG. 2 when the computer program is executed in a computer.

According to still another embodiment, there is also provided a computing device comprising a memory and a processor, the memory storing executable code, and when the processor executes the executable code, implementing the method described in connection with FIG. 2 method.

Those skilled in the art will appreciate that in one or more examples described above, the functions described herein can be implemented in hardware, software, firmware, or any combination thereof. When implemented in software, the functions may be stored in a computer readable medium or transmitted as one or more instructions or code on a computer readable medium.

The specific embodiments of the present invention have been described in detail with reference to the preferred embodiments of the present invention. The scope of the protection, any modifications, equivalent substitutions, improvements, etc., which are made on the basis of the technical solutions of the present invention, are included in the scope of the present invention.

Claims

A method of processing transaction data, including:

Obtaining n data sets respectively corresponding to consecutive n preset time segments, where each data set i in the n data sets includes transaction detail data of the user in the corresponding time period;

Forming n feature vectors respectively corresponding to the n data sets, wherein each feature vector Fi includes a derivative variable derived based on transaction detail data in the corresponding data set i;

Importing the n feature vectors into a time recursive neural network in chronological order, obtaining a processing result from the time recursive neural network, wherein obtaining the processing result includes obtaining the result from an output layer of the time recursive neural network The probability that a user has a credit default is the result of the processing.
The method of claim 1 wherein said transaction detail data comprises a plurality of fields, said plurality of fields comprising at least: a transaction time field, a transaction amount field, and at least one category field.
The method of claim 2, wherein forming n feature vectors respectively corresponding to the n data sets comprises:

Obtaining the plurality of fields of the transaction detail data in the data set i;

Aggregating data of the plurality of fields to obtain derived variables;

The derived variable is taken as a vector element of the feature vector Fi.
The method of claim 3 wherein the aggregating the data in the plurality of fields comprises:

Selecting at least a part of the plurality of fields to be combined to obtain a combined field;

The data of the combined field is operated to obtain the derived variable.
The method of claim 4 wherein said computing operation comprises one or more of the following: numerically determining, counting, summing, averaging, seeking standard deviation, finding quantile, and distributing statistics.
The method of claim 2, wherein forming n feature vectors respectively corresponding to the n data sets comprises:

Obtaining content of the at least one category field in the data set i;

Converting the content of the at least one category field into a word vector using a word embedding model;

The word vector is taken as part of the feature vector Fi.
The method of claim 1 wherein said time recursive neural network employs one of a recurrent neural network RNN, a long- and short-term memory neural network LSTM, and a gated loop unit neural network GRU.
The method of claim 7 wherein said time recursive neural network comprises at least one fully connected layer.
The method of claim 1, the time recursive neural network training with a calibrated data set comprising historical transaction data and having a label of whether a credit default has occurred.
The method of claim 1, wherein obtaining the processing result from the time recursive neural network comprises obtaining a node feature value from the hidden layer of the time recursive neural network as a processing result.
A device for processing transaction data, comprising:

a data set obtaining unit, configured to acquire n data sets respectively corresponding to consecutive n preset time segments, where each data set i of the n data sets includes transaction detail data of the user in the corresponding time period;

a vector forming unit configured to form n feature vectors respectively corresponding to the n data sets, wherein each feature vector Fi includes a derivative variable derived based on transaction detail data in the corresponding data set i;

a processing unit configured to input the n feature vectors into a time recursive neural network in chronological order, obtain a processing result from the time recursive neural network, wherein obtaining the processing result includes output from the time recursive neural network The layer obtains the probability that the user has a credit default as a processing result.
The apparatus of claim 11, wherein the transaction detail data comprises a plurality of fields, the plurality of fields comprising at least: a transaction time field, a transaction amount field, and at least one category field.
The apparatus of claim 12 wherein said vector forming unit comprises:

a field obtaining module configured to acquire the plurality of fields of the transaction detail data in the data set i;

An aggregation operation module configured to perform aggregation operations on data of the plurality of fields to obtain a derivative variable;

An element forming module configured to use the derived variable as a vector element of the feature vector Fi.
The apparatus of claim 13 wherein said aggregation operation module is further configured to:

Selecting at least a part of the plurality of fields to be combined to obtain a combined field;

The data of the combined field is operated to obtain the derived variable.
The apparatus of claim 14, wherein the arithmetic operation comprises one or more of the following: numerical value determination, counting, summation, averaging, standard deviation, grading, distribution statistics.
The apparatus of claim 12, wherein the vector forming unit further comprises a word embedding module configured to:

Obtaining content of the at least one category field in the data set i;

Converting the content of the at least one category field into a word vector using a word embedding model;

The word vector is taken as part of the feature vector Fi.
The apparatus of claim 11 wherein said time recursive neural network employs one of a recurrent neural network RNN, a long- and short-term memory neural network LSTM, and a gated loop unit neural network GRU.
The apparatus of claim 17 wherein said time recursive neural network comprises at least one fully connected layer.
The apparatus of claim 12, said time recursive neural network training with a calibrated data set comprising historical transaction data and having a label of whether a credit default has occurred.
The apparatus of claim 11, wherein the processing unit is configured to obtain a node feature value from the hidden layer of the time recursive neural network as a result of the processing.
A computer readable storage medium having stored thereon a computer program for causing a computer to perform the method of any of claims 1-10 when the computer program is executed in a computer.
A computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, implementing the method of any one of claims 1-10 method.