Transaction time sequence characteristic data processing method and device

Abstract

The invention discloses a transaction time sequence characteristic data processing method and device. The method comprises the following steps: constructing a current time sequence financial transaction vector; using an associated party transaction update switch to control whether the current time sequence financial transaction vector needs to be updated so as to determine that the current time sequence transaction state needs to retain the transaction information amount of the previous time sequence to obtain a transaction update result; controlling the ignored degree of the implicit transaction state of the previous time sequence by using an associated party transaction reset switch to obtain a transaction reset result; obtaining a candidate implicit transaction state of the previous time sequence according to a reset result; and fusing the updating and resetting results and the candidate implicit transaction state to obtain associated transaction feature data. According to the invention, the transaction control switch is arranged to update or reset the transaction time sequence information and determine whether to retain or discard useless transaction information, so that effective associated transaction feature data in the transaction time sequence data can be better captured to adapt to financial transaction information time sequence changes, and the quality and availability of the associated transaction feature data are improved.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and in particular to a method and device for processing transaction time series feature data. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the invention that are recited in the claims. No statement herein is admitted to be prior art by virtue of its inclusion in this section.

[0003] Existing related-party transaction data is primarily stored in relational databases. Due to the large volume of data, traditional methods often rely on businesses setting business rules based on experience to extract and process related-party transaction data. However, financial transaction data is typically a time series of transactions. Compared to traditional machine learning, recurrent neural network (RNN) algorithms, while setting business rules to extract and process related-party transaction data can extract financial transaction data, it cannot capture long-term transaction contextual information, i.e., encounters a "time series gap" problem. While eliminating distorted related-party transaction data extracted by business rules, some implicit real-world related-party transaction data is often overlooked by the system, making it difficult to effectively capture valid related-party transaction data and unable to adapt to the time series changes in financial institutions' transaction information data. Summary of the Invention

[0004] An embodiment of the present invention provides a method for processing transaction time series feature data, which is used to set a transaction control switch to update or reset transaction time series information, determine whether to retain or discard useless transaction information, and thus better capture valid associated transaction feature data in the transaction time series data to adapt to changes in the time series of financial transaction information. The method includes:

[0005] Obtain financial business transaction data of all dimensions of related parties and construct the current time series financial transaction vector;

[0006] Use the related-party transaction update switch to control whether the current time-series financial transaction vector needs to be updated, so as to determine the amount of transaction information of the previous time-series that needs to be retained in the current time-series transaction state, and obtain the related-party transaction update result;

[0007] The related-party transaction reset switch is used to control the neglect degree of the previous time-series implicit transaction state of the current time-series financial transaction vector, so as to determine the previous time-series implicit transaction state of the related-party transaction information and obtain the related-party transaction reset result;

[0008] According to the reset result of the related-party transaction party, the dependence degree of the related-party transaction on the implicit transaction state of the previous time series is determined, and the candidate implicit transaction state of the previous time series is obtained;

[0009] The related-party transaction update results, related-party transaction reset results and candidate implicit transaction states of the previous time series are fused to obtain the current time series implicit transaction state and its corresponding related-party transaction feature data.

[0010] An embodiment of the present invention further provides a device for processing transaction time series feature data, which is used to set a transaction control switch to update or reset transaction time series information and decide whether to retain or discard useless transaction information, thereby better capturing valid associated transaction feature data in the transaction time series data and adapting to changes in the time series of financial transaction information. The device includes:

[0011] The acquisition unit is used to obtain the financial business transaction data of all dimensions of the related parties and construct the current time series financial transaction vector;

[0012] An updating unit, configured to use a related-party transaction update switch to control whether the current time-series financial transaction vector needs to be updated, so as to determine the amount of transaction information of the previous time-series that needs to be retained in the current time-series transaction state, and obtain a related-party transaction update result;

[0013] A reset unit, configured to control the degree of neglect of the previous time-series implicit transaction state of the current time-series financial transaction vector by using the related-party transaction reset switch, so as to determine the previous time-series implicit transaction state of the related-party transaction information and obtain a related-party transaction reset result;

[0014] a candidate implicit transaction state determination unit, configured to determine the degree of dependence of the related-party transaction on the implicit transaction state of the previous time sequence according to the reset result of the related-party transaction party, and obtain the candidate implicit transaction state of the previous time sequence;

[0015] The extraction unit is used to use the related-party transaction update result, the related-party transaction reset result and the candidate implicit transaction state of the previous time series to fuse and obtain the current time series implicit transaction state and its corresponding related-party transaction feature data.

[0016] An embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method for processing transaction timing feature data when executing the computer program.

[0017] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for processing the transaction time series feature data is implemented.

[0018] An embodiment of the present invention further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the above-mentioned method for processing transaction time series feature data.

[0019] Compared with the existing technical solutions for processing transaction timing feature data that are difficult to effectively capture effective related-party transaction data and cannot adapt to the temporal changes in financial institution transaction information data, the transaction timing feature data processing solution provided by the embodiment of the present invention updates or resets the transaction timing information by setting a transaction control switch, and decides whether to retain or discard useless transaction information, thereby better capturing effective related transaction feature data in the transaction timing data, adapting to the temporal changes in financial transaction information, and improving the quality and availability of related transaction feature data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0021] Figure 1 Schematic diagram of a process for processing transaction time series feature data according to an embodiment of the present invention;

[0022] Figure 2 Schematic diagram of the principle of extracting time series features of related-party transactions in an embodiment of the present invention;

[0023] Figure 3 Schematic diagram of the hyperparameter tuning process of the model in an embodiment of the present invention;

[0024] Figure 4 Schematic diagram of the structure of a device for processing transaction time series feature data in an embodiment of the present invention;

[0025] Figure 5 FIG. 1 is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings. Here, the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0027] The acquisition, transmission, storage, use, and processing of data in the technical solution of this application comply with the relevant provisions of laws and regulations.

[0028] It should be noted that in the embodiments of the present application, certain software, components, models and other existing solutions in the industry may be mentioned. They should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0029] Before introducing the embodiments of the present invention, the terms involved in the embodiments of the present invention are first introduced as follows.

[0030] Related party: A natural person, legal person or unincorporated organization that has an associated relationship with the enterprise as defined in the regulatory rules that the enterprise needs to comply with.

[0031] Related-party transactions: refer to the transfer of resources, services, obligations or other benefits between an enterprise and its subsidiaries and related parties.

[0032] The inventors discovered that due to the large volume of related-party transaction data, traditional methods for processing related-party transaction data primarily rely on rules set based on business experience. For example, these methods extract and process related-party transaction data using business type, transaction time, transaction direction, and transaction method. This sometimes requires business personnel to intervene and conduct screening and analysis to eliminate distorted related-party transaction data extracted by business rules. This often results in some implicit, real related-party transaction data being ignored by the system. Setting corresponding business rules for all related-party transaction data is a crude and simplistic approach, and sometimes the rules are outdated or not updated in a timely manner, resulting in inaccurate related-party transaction data extraction. Consequently, when measuring related-party transaction risk, errors in related-party transaction measurement occur and cannot be corrected in a timely manner, leading to misreporting of related-party transaction data.

[0033] In light of the technical challenges inherent in existing technologies, embodiments of the present invention provide a solution for processing transaction time series feature data. This solution extracts features from related-party transaction details by designing an effective financial transaction model algorithm. This improves the quality and usability of related-party transaction data, provides a better data foundation for related-party transaction regulatory model analysis, and thus enhances the accuracy of related-party transaction identification. The following describes this solution for processing transaction time series feature data in detail.

[0034] Figure 1 FIG. 1 is a flow chart of a method for processing transaction time series feature data in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps:

[0035] Step 101: Obtain financial business transaction data of all dimensions of related parties and construct the current time series financial transaction vector;

[0036] Step 102: Use the related-party transaction update switch to control whether the current time-series financial transaction vector needs to be updated, so as to determine the amount of transaction information of the previous time-series that needs to be retained in the current time-series transaction state, and obtain the related-party transaction update result;

[0037] Step 103: Using the related-party transaction reset switch to control the degree of neglect of the previous-time-sequence implicit transaction state of the current-time-sequence financial transaction vector, to determine the previous-time-sequence implicit transaction state of the related-party transaction information, and obtain the related-party transaction reset result;

[0038] Step 104: Determine the degree of dependence of the related-party transaction on the implicit transaction state of the previous time sequence based on the reset result of the related-party transaction party, and obtain the candidate implicit transaction state of the previous time sequence;

[0039] Step 105: Utilize the related-party transaction update result, the related-party transaction reset result and the candidate implicit transaction state of the previous time series to fuse and obtain the current time series implicit transaction state and its corresponding related-party transaction feature data.

[0040] The method for processing transaction time series feature data provided by an embodiment of the present invention, when in operation: obtains financial business transaction data of all dimensions of related parties and constructs a current time series financial transaction vector; uses a related party transaction update switch to control whether the current time series financial transaction vector needs to be updated, so as to determine the amount of transaction information of the previous time series that needs to be retained in the current time series transaction state, and obtains a related party transaction update result; uses a related party transaction reset switch to control the degree of neglect of the previous time series implicit transaction state of the current time series financial transaction vector, so as to determine the implicit transaction state of the previous time series of the related party transaction information, and obtains a related party transaction reset result; determines the degree of dependence of the related party transaction on the implicit transaction state of the previous time series based on the related party transaction reset result, and obtains a candidate implicit transaction state of the previous time series; and uses the related party transaction update result, the related party transaction reset result, and the candidate implicit transaction state of the previous time series to fuse and obtain the current time series implicit transaction state and its corresponding related party transaction feature data.

[0041] Compared with the existing technical solutions for processing transaction timing feature data that are difficult to effectively capture effective related-party transaction data and cannot adapt to the temporal changes in financial institution transaction information data, the transaction timing feature data processing method provided by the embodiment of the present invention updates or resets the transaction timing information by setting a transaction control switch, and decides whether to retain or discard useless transaction information, thereby better capturing effective related transaction feature data in the transaction timing data, adapting to the temporal changes in financial transaction information, and improving the quality and availability of related transaction feature data. Figure 2 This is a schematic diagram of the principle of extracting time series features of related-party transactions in an embodiment of the present invention. Figure 3 This is a schematic diagram of the hyperparameter tuning process of the model in the embodiment of the present invention. Figures 2 to 3 The method for processing transaction time series feature data involved in an embodiment of the present invention is introduced in detail.

[0042] In the above step 101, a related party transaction data mart is constructed, and all business transaction data of related parties are captured from the financial transaction data mart. The dimensions include transaction counterparty, transaction time, transaction type, transaction amount, transaction price, transaction institution, etc., thereby constructing high-dimensional spatial data information of business transaction data and obtaining transaction data that meets business needs.

[0043] Considering the technical issues existing in existing technologies, this requires adjusting the time series detection of transaction sequences to capture the contextual information of transaction time series data. At the same time, whether to retain or discard useless transaction information as needed to ensure that valid related-party transaction information is captured. To this end, this paper provides a transaction control switch to update or reset transaction time series information, determining whether to retain the memory information at the beginning of the sequence. This allows for better capture of related-party transaction feature data in transaction time series data, adapting to the time series changes of financial transaction information, and constructing related-party transaction feature data that meets regulatory and internal control management requirements.

[0044] In the above step 101, the transaction dimension space is established:

[0045]

[0046] Table 1

[0047] Based on the transaction details data shown in Table 1 above, the transaction details are combined according to the counterparty time to construct the transaction high-dimensional space vector T at time t. t , dimension d, as the subsequent input, that is, Figure 2 The financial transaction vector T in t .

[0048] In the above step 102, a related party transaction update switch is set up, such as Figure 2 Transaction update switch U in t :

[0049] The update switch primarily controls whether related-party transaction information for the current time step (a time step is the smallest unit of discrete continuous time in numerical simulation or sequence analysis, representing the interval between two adjacent state updates) should be updated. It determines how much transaction information from the previous time step should be retained for the current transaction state. Its output value ranges from 0 to 1. A larger value indicates more retention of previous transaction information (the amount of transaction information from the previous time series), while a smaller value indicates greater reliance on current transaction information (the current time series financial transaction vector).

[0050] The calculation formula for the output of the related-party transaction update switch is: In one embodiment, the related-party transaction update switch is used to control whether the current time-series financial transaction vector needs to be updated, so as to determine the amount of transaction information of the previous time-series that needs to be retained in the current time-series transaction state, and obtain the related-party transaction update result. This may include obtaining the related-party transaction update result according to the following related-party transaction update switch:

[0051]

[0052] Among them, M u is the matrix of related-party transaction update switches, ε u is the adjustment vector of the related-party transaction update switch, C t-1 is the implicit transaction state of the previous time step, that is, the implicit transaction state of the previous time series, T t is the current financial transaction input, that is, the current time series financial transaction vector, α is the adjustment parameter, between (0, 1), is the switch activation function, and the updated output of the related-party transaction is U t , that is U t Update results for related party transactions.

[0053] In the above step 103, a related party transaction reset switch is set up, such as Figure 2 Transaction reset switch R in t :

[0054] The reset switch determines the extent to which the implicit transaction state of related-party transaction information from the previous time step is ignored. When the reset switch output is close to 0, the transaction information tends to "forget" the transaction information from the previous time step and rely only on the current input; when the output is close to 1, the transaction information from the previous time step is retained more.

[0055] The calculation formula of the related-party transaction reset switch is: In one embodiment, the related-party transaction reset switch is used to control the neglect degree of the implicit transaction state of the previous time sequence of the current time sequence financial transaction vector to determine the implicit transaction state of the previous time sequence of the related-party transaction information, and obtain the related-party transaction reset result. The calculation formula of the related-party transaction reset switch can include:

[0056]

[0057] Among them, M r is the matrix of related-party transaction reset switches, ε r is the adjustment vector of the related-party transaction reset switch, C t-1 is the implicit transaction state of the previous time step, that is, the implicit transaction state of the previous time series, T tis the current financial transaction input, that is, the current time series financial transaction vector, α is the adjustment parameter, between (0, 1), is the switch activation function, and the related party transaction reset output is R t , that is, R t Reset results for related party transactions.

[0058] In step 104 above, the related-party transaction implicitly measures the transaction status:

[0059] Use the related party transaction reset switch to control the related party transaction implicit transaction status of the previous time series (such as Figure 2 Implicit state C t-1 ), the implicit transaction state combines the influence of the current transaction and the reset switch to calculate the candidate implicit transaction state.

[0060] The calculation formula for the candidate implicit transaction state of the related-party transaction is: In one embodiment, the degree of dependence of the related-party transaction on the implicit transaction state of the previous time sequence is determined based on the reset result of the related-party transaction party, and the candidate implicit transaction state of the previous time sequence is obtained. This can include obtaining the candidate implicit transaction state of the previous time sequence according to the following candidate implicit transaction state formula:

[0061] C t =AdMish(W C ·[R t ⊙C t-1 , T t-1 ]+ε C ])

[0062]

[0063] Among them, W c and ε c is the implicit matrix and adjustment vector. AdMish is the activation function, which is further adjusted based on Mish (a new activation function proposed in 2019). It retains the impact of changes in transaction time series data and can show a better transaction information screening effect on transaction time series data. ⊙ represents element-wise multiplication, C t ' is the candidate implicit transaction state, that is, the candidate implicit transaction state of the previous time series, R t Reset result for related party transactions, C t-1 is the implicit transaction state of the previous time series, T t-1 is the previous time series financial transaction vector.

[0064] In the above step 105, the implicit transaction status of the current time series of the related-party transaction is:

[0065] The last step is to use the previous time step state of the related party transaction information and the candidate implicit transaction state C at the previous momentt , and calculate the current implicit transaction state by performing weighted fusion through updating switches, such as Figure 2 The hidden state C in t .

[0066] The calculation formula for the current implicit transaction state is: In one embodiment, the related-party transaction update result, the related-party transaction reset result, and the candidate implicit transaction state of the previous time series are integrated to obtain the current time series implicit transaction state and its corresponding related-party transaction feature data. The calculation formula may include obtaining the current time series implicit transaction state according to the following implicit transaction state formula:

[0067] C t =C t-1 ⊙(1-U t )+C t ⊙U t ;

[0068] Among them, C t is the current time series implicit transaction status, C t-1 is the implicit transaction state of the previous time step, that is, the implicit transaction state of the previous time series, C t ' is the candidate implicit transaction state, that is, the candidate implicit transaction state of the previous time series, U t To update the switch weight, U t It is the update result of the related party transaction, which controls the previous time series financial transaction C t-1 The retention ratio and new transaction information C t 'Introduction ratio. Finally, we extracted Figure 2 The characteristics of related-party transactions are shown.

[0069] The following describes further preferred embodiments of the present invention.

[0070] The training and tuning steps of the model involving the characteristics of related-party transaction time series transaction data are as follows:

[0071] Due to the large amount of related-party transaction information data, considering the performance of measuring the entire time series of related-party transaction information, it is necessary to tune the hyperparameters to improve the effect of the related-party transaction feature extraction model.

[0072] First, because the number of hidden units selected in the model affects its expressiveness and complexity, it's typically recommended to start with smaller values of 32, 64, or 128 and gradually increase the number of units. For related-party transactions, key considerations include the type of business, direction, method, time, and accounting type. While the number of hidden units can improve the model's ability to represent transaction time series, too many hidden units can lead to overfitting. Experimentation and data analysis have shown that the model with 64 units in this embodiment of the present invention effectively represents related-party transaction data.

[0073] Secondly, the number of layers has a significant impact on the model. Related-party transaction information is not a simple task, and it is difficult to obtain transaction time series data features with a simple layer. In the embodiment of the present invention, two layers are stacked to obtain transaction information time series data, while avoiding too many layers causing a sharp increase in time complexity.

[0074] Finally, the learning rate determines the gradient update step size, which in turn affects the convergence of the entire related-party transaction information model. While there are many learning rate methods, this paper employs a variant of the Adam algorithm, the AdamW algorithm, to address the weight decay of the Adam algorithm, effectively decoupling the weight decay and optimization steps of related-party transactions.

[0075] Related-party transaction data feature extraction training AdamW tuning process:

[0076] In one embodiment, the related party transaction update switch is a related party transaction update neural network model, and the related party transaction reset is a related party transaction reset neural network model; Figure 3 FIG. 1 is a schematic diagram of the hyperparameter tuning process of the model in an embodiment of the present invention, as shown in FIG. Figure 3 As shown, the method for processing transaction time series feature data may further include the following steps of tuning the hyperparameters of the related-party transaction update neural network model and the related-party transaction reset neural network model:

[0077] Step 201: Determine the training learning gradient of related-party transaction features;

[0078] Step 202: Determine the training learning momentum and second-order moment of the related-party transaction feature based on the learning gradient;

[0079] Step 203: Update the current moment parameters of the financial transaction according to the momentum and the second-order moment to obtain an updated related-party transaction update neural network model and a related-party transaction reset neural network model.

[0080] A detailed introduction is given below.

[0081] In the above step 201, the transaction feature training learns the gradient calculation:

[0082] The formula for obtaining the current parameter gradient through back propagation is as follows. In one embodiment, determining the training learning gradient of the related-party transaction feature includes determining the training learning gradient of the related-party transaction feature according to the following formula:

[0083]

[0084] Among them, g t is the gradient of the transaction series at time t, Loss is the loss function, It is the current moment parameter of financial business transaction.

[0085] In step 202 above, the trading feature training learns momentum and second-order moment calculations:

[0086] 1) Momentum:

[0087] M t =β1M t-1 +(1-β1)g t ;

[0088] Among them, M t is the momentum of the transaction at the current moment, M t-1 is the momentum of the previous time series, β1 is the momentum decay rate, and according to the time series characteristics of the transaction business, it can be set to 0.95 in the embodiment of the present invention.

[0089] 2) Second-order moment estimation:

[0090] G t =β2G t-1 +(1-β2)g t ⊙g t ;

[0091] Among them, G t is the second moment of the transaction at the current moment, G t-1 is the second-order moment of the previous time series, β2 is the second-order moment attenuation rate, and according to the transaction business time series characteristics, it can be set to 0.999 in the embodiment of the present invention.

[0092] After the above step 202, the method for processing the above transaction timing feature data may further include a step of deviation correction, that is, in one embodiment, the method for processing the above transaction timing feature data may further include: correcting the momentum according to the momentum decay rate to obtain a corrected momentum, and correcting the second-order moment according to the second-order moment decay rate to obtain a corrected second-order moment; in this way, the subsequent steps update the current moment parameters of the financial transaction according to the momentum and the second-order moment to obtain an updated related-party transaction update neural network model and a related-party transaction reset neural network model, and may include: updating the current moment parameters of the financial transaction according to the corrected momentum and the corrected second-order moment to obtain an updated related-party transaction update neural network model and a related-party transaction reset neural network model, thereby further improving the accuracy of the processing of the transaction timing feature data.

[0093] In one embodiment, correcting the momentum according to the momentum decay rate to obtain the corrected momentum may include correcting the momentum according to the following formula to obtain the corrected momentum:

[0094]

[0095] in, is the modified momentum, Mt is the momentum of the transaction at the current moment, is the momentum decay rate at the current time t of the transaction;

[0096] Correcting the second-order moment according to the second-order moment attenuation rate to obtain a corrected second-order moment may include correcting the second-order moment according to the following formula to obtain a corrected second-order moment:

[0097]

[0098] in, is the corrected second-order moment, G t is the second moment of the transaction at the current moment, is the second-order moment decay rate of the transaction at the current time t.

[0099] In specific implementation, by correcting the momentum and second-order moment, the impact of the initial value on the initial value of the financial transaction feature can be reduced, and the efficiency of processing the transaction time series feature data can be further improved.

[0100] In the above step 202, the parameters are updated:

[0101] In one embodiment, updating the current moment parameters of the financial transaction based on the momentum and the second-order moment may include updating the current moment parameters of the financial transaction according to the following formula:

[0102]

[0103] in, is the current moment parameter of financial transaction, is the parameter of the financial transaction at the previous moment, is the corrected momentum, ε is the adjustment vector, α is the adjustment parameter, which is between (0, 1), and λ is the adjustment weight, which is adjusted according to actual experience. is the corrected second-order moment, is the gradient update part of Adam, It is an independent weight decay term and has nothing to do with gradient updates. The learning rate weight decay of the related-party transaction model directly acts on the current parameters rather than the gradient, thereby achieving robustness of model training.

[0104] In summary, by optimizing the related-party transaction characteristics and extracting the temporal characteristics of the related-party transaction data, the model not only solves the problems of timeliness in processing large amounts of related-party transaction data, model fitting and gradient attenuation, but also can well realize the extraction of related-party transaction data characteristics, effectively improving the related-party transaction identification capabilities and management requirements.

[0105] In summary, the method for processing transaction time series feature data provided by the embodiment of the present invention has the following beneficial technical effects:

[0106] 1. Solve the problem of relying on business experience and rule setting by optimizing the model involved in extracting related-party transaction features, and solve the problem of untimely rule updates and incorrect capture of related-party transaction data.

[0107] 2. By setting update switches and reset switches for related-party transactions, the long-term financial transaction context timing information can be effectively resolved, and the timing feature information of related-party transactions can be effectively captured.

[0108] 3. By adjusting the activation function of the related-party transaction switch, we can comprehensively consider the timing and parameter fine-tuning to more closely fit the timing characteristic data of related-party transactions.

[0109] 4. By using AdMish as the activation function to process the implicit transaction status of related-party transactions, it can not only retain the impact of changes in transaction time series data, but also show a better transaction information screening effect on transaction time series data and more effectively process the time series characteristics of related-party transaction data.

[0110] 5. By optimizing the model involved in extracting related-party transaction features, effective related-party transaction data can be quickly obtained from the vast amount of related-party transaction data, and related-party transaction feature data can be quickly extracted, reducing the need for business personnel to conduct secondary screening and review, while also reducing manual re-entry and declaration, greatly freeing up manpower operations.

[0111] 6. By optimizing the model involved in extracting related-party transaction features, the expression of related-party transaction data information features is effectively achieved while reducing overfitting.

[0112] 7. By optimizing the models involved in extracting related-party transaction features, we can solve problems such as model gradient attenuation, improve the ability to identify related-party transactions, and meet the internal management and regulatory reporting requirements for related-party transactions.

[0113] The present invention also provides a device for processing transaction time series feature data, as described in the following embodiments. Because the principles underlying the device are similar to those of the method for processing transaction time series feature data, the implementation of the device can be found in the implementation of the method for processing transaction time series feature data, and any repetitions will not be repeated.

[0114] Figure 4 FIG. 1 is a structural diagram of a device for processing transaction time series feature data according to an embodiment of the present invention. Figure 4 As shown, the device includes:

[0115] Acquisition unit 01 is used to obtain financial business transaction data of all dimensions of related parties and construct the current time series financial transaction vector;

[0116] Update unit 02, used to control whether the current time series financial transaction vector needs to be updated using the related party transaction update switch, so as to determine the amount of transaction information of the previous time series that needs to be retained in the current time series transaction state, and obtain the related party transaction update result;

[0117] Reset unit 03, used to control the neglect degree of the previous time sequence implicit transaction state of the current time sequence financial transaction vector by using the related party transaction reset switch, so as to determine the previous time sequence implicit transaction state of the related party transaction information and obtain the related party transaction reset result;

[0118] The candidate implicit transaction state determination unit 04 is used to determine the degree of dependence of the related-party transaction on the implicit transaction state of the previous time sequence according to the reset result of the related-party transaction party, and obtain the candidate implicit transaction state of the previous time sequence;

[0119] The extraction unit 05 is used to use the related party transaction update result, the related party transaction reset result and the candidate implicit transaction state of the previous time series to fuse and obtain the current time series implicit transaction state and its corresponding related party transaction feature data.

[0120] In one embodiment, the updating unit is specifically configured to obtain the related party transaction update result according to the following related party transaction update switch:

[0121]

[0122] Among them, M u is the matrix of related-party transaction update switches, ε u is the adjustment vector of the related-party transaction update switch, C t-1 is the implicit transaction state of the previous time series, T t is the current time series financial transaction vector, α is the adjustment parameter, is the switch activation function, U t Update results for related party transactions.

[0123] In one embodiment, the reset unit is specifically configured to obtain the related party transaction reset result according to the following related party transaction reset switch:

[0124]

[0125] Among them, M r is the matrix of related-party transaction reset switches, ε r is the adjustment vector of the related-party transaction reset switch, C t-1 is the implicit transaction state of the previous time series, T t is the current time series financial transaction vector, α is the adjustment parameter, is the switch activation function, R t Reset results for related party transactions.

[0126] In one embodiment, the candidate implicit transaction state determining unit is specifically configured to obtain the candidate implicit transaction state of the previous time sequence according to the following candidate implicit transaction state formula:

[0127] C t '=AdMish(W c ·[R t ⊙C t-1 , T t-1 ]+ε c ]);

[0128]

[0129] Among them, w c is the implicit matrix, ε c is the adjustment vector, AdMish is the activation function, C t ' is the candidate implicit transaction state of the previous time series, R t Reset result for related party transactions, C t-1 is the implicit transaction state of the previous time series, T t-1 is the previous time series financial transaction vector.

[0130] In one embodiment, the extraction unit is specifically configured to obtain the current time series implicit transaction state according to the following implicit transaction state formula:

[0131] C t =C t-1 ⊙(1-U t )+C t ⊙U t ;

[0132] Among them, C t is the current time series implicit transaction status, C t-1 is the implicit transaction state of the previous time series, C t is the candidate implicit transaction state of the previous time series, U t Update results for related party transactions.

[0133] In one embodiment, the related party transaction update switch is a related party transaction update neural network model, and the related party transaction reset is a related party transaction reset neural network model;

[0134] The method for processing transaction time series feature data further includes an optimization unit for performing the following steps to optimize the hyperparameters of the related-party transaction update neural network model and the related-party transaction reset neural network model:

[0135] Determine the training learning gradient for related-party transaction characteristics;

[0136] Determining the learning momentum and second-order moment of the related-party transaction feature training according to the learning gradient;

[0137] According to the momentum and the second-order moment, the current moment parameters of the financial transaction are updated to obtain an updated related-party transaction update neural network model and an updated related-party transaction reset neural network model.

[0138] In one embodiment, the optimization unit is further configured to: correct the momentum according to the momentum decay rate to obtain a corrected momentum, and correct the second-order moment according to the second-order moment decay rate to obtain a corrected second-order moment;

[0139] According to the momentum and the second-order moment, the parameters of the financial transaction at the current moment are updated to obtain an updated related-party transaction update neural network model and a related-party transaction reset neural network model, including: according to the corrected momentum and the corrected second-order moment, the parameters of the financial transaction at the current moment are updated to obtain an updated related-party transaction update neural network model and a related-party transaction reset neural network model.

[0140] In one embodiment, correcting the momentum according to the momentum decay rate to obtain the corrected momentum includes: correcting the momentum according to the following formula to obtain the corrected momentum:

[0141]

[0142] in, is the modified momentum, M t is the momentum of the transaction at the current moment, is the momentum decay rate at the current time t of the transaction;

[0143] Correcting the second-order moment according to the second-order moment attenuation rate to obtain a corrected second-order moment includes correcting the second-order moment according to the following formula to obtain a corrected second-order moment:

[0144]

[0145] in, is the corrected second-order moment, G t is the second moment of the transaction at the current moment, is the second-order moment decay rate of the transaction at the current time t.

[0146] In one embodiment, updating the current moment parameters of the financial transaction based on the momentum and the second-order moment includes updating the current moment parameters of the financial transaction according to the following formula:

[0147]

[0148] in, is the current moment parameter of financial transaction, is the parameter of the financial transaction at the previous moment, is the corrected momentum, is the corrected second-order moment, is the gradient update part of Adam, is an independent weight decay term.

[0149] In one embodiment, determining the training learning momentum and second-order moment of the related-party transaction feature based on the learning gradient includes determining the momentum according to the following formula:

[0150] M t =β1M t-1 +(1-β1)g t ;

[0151] Among them, M t is the momentum of the transaction at the current moment, M t-1 is the momentum of the previous time series, β1 is the momentum decay rate, g t is the gradient of the transaction at the current time t.

[0152] In one embodiment, determining the training learning momentum and second-order moment of the related-party transaction feature based on the learning gradient includes determining the second-order moment according to the following formula:

[0153] G t =β2G t-1 +(1-β2)g t ⊙g t ;

[0154] Among them, G t is the second moment of the transaction at the current moment, G t-1 is the second-order moment of the previous time series, β2 is the second-order moment decay rate, g t is the gradient of the transaction at the current time t.

[0155] Based on the above invention concept, Figure 5 As shown, the present invention also proposes a computer device 500, including a memory 510, a processor 520 and a computer program 530 stored in the memory 510 and executable on the processor 520, wherein the processor 520 implements the aforementioned method for processing transaction timing feature data when executing the computer program 530.

[0156] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for processing the transaction time series feature data is implemented.

[0157] An embodiment of the present invention further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the above-mentioned method for processing transaction time series feature data.

[0158] Compared with the existing technical solutions for processing transaction timing feature data that are difficult to effectively capture effective related-party transaction data and cannot adapt to the temporal changes in financial institution transaction information data, the transaction timing feature data processing solution provided by the embodiment of the present invention updates or resets the transaction timing information by setting a transaction control switch, and decides whether to retain or discard useless transaction information, thereby better capturing effective related transaction feature data in the transaction timing data, adapting to the temporal changes in financial transaction information, and improving the quality and availability of related transaction feature data.

[0159] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0161] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0163] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for processing transaction time series feature data, characterized in that: include: Obtain financial business transaction data of all dimensions of related parties and construct the current time series financial transaction vector; Use the related-party transaction update switch to control whether the current time-series financial transaction vector needs to be updated, so as to determine the amount of transaction information of the previous time-series that needs to be retained in the current time-series transaction state, and obtain the related-party transaction update result; The related-party transaction reset switch is used to control the neglect degree of the previous time-series implicit transaction state of the current time-series financial transaction vector, so as to determine the previous time-series implicit transaction state of the related-party transaction information and obtain the related-party transaction reset result; According to the reset result of the related-party transaction party, the dependence degree of the related-party transaction on the implicit transaction state of the previous time series is determined, and the candidate implicit transaction state of the previous time series is obtained; The related-party transaction update results, related-party transaction reset results and candidate implicit transaction states of the previous time series are fused to obtain the current time series implicit transaction state and its corresponding related-party transaction feature data.

2. The method according to claim 1, wherein The related-party transaction update switch is used to control whether the current time-series financial transaction vector needs to be updated, so as to determine the amount of transaction information of the previous time-series that needs to be retained in the current time-series transaction state, and obtain the related-party transaction update result, including obtaining the related-party transaction update result according to the following related-party transaction update switch: Among them, M u is the matrix of related-party transaction update switches, ε u is the adjustment vector of the related-party transaction update switch, C t-1 is the implicit transaction state of the previous time series, T t is the current time series financial transaction vector, α is the adjustment parameter, is the switch activation function, U t Update results for related party transactions.

3. The method according to claim 1, wherein The related-party transaction reset switch is used to control the neglect degree of the previous time-series implicit transaction state of the current time-series financial transaction vector, so as to determine the previous time-series implicit transaction state of the related-party transaction information, and obtain the related-party transaction reset result, including obtaining the related-party transaction reset result according to the following related-party transaction reset switch: Among them, M r is the matrix of related-party transaction reset switches, ε r is the adjustment vector of the related-party transaction reset switch, C t-1 is the implicit transaction state of the previous time series, T t is the current time series financial transaction vector, α is the adjustment parameter, is the switch activation function, R t Reset results for related party transactions.

4. The method according to claim 1, wherein According to the reset result of the related-party transaction party, the dependence degree of the related-party transaction on the implicit transaction state of the previous time sequence is determined, and the candidate implicit transaction state of the previous time sequence is obtained, including obtaining the candidate implicit transaction state of the previous time sequence according to the following candidate implicit transaction state formula: C t ′=AdMish(W c ·[R t ⊙C t-1 ,T t-1 ]+ε c ]); Among them, W c is the implicit matrix, ε c is the adjustment vector, AdMish is the activation function, C t ′ is the candidate implicit transaction state of the previous time series, R t Reset result for related party transactions, C t-1 is the implicit transaction state of the previous time series, T t-1 is the previous time series financial transaction vector.

5. The method according to claim 1, wherein The related-party transaction update result, the related-party transaction reset result, and the candidate implicit transaction state of the previous time series are integrated to obtain the current time series implicit transaction state and its corresponding related-party transaction feature data, including obtaining the current time series implicit transaction state according to the following implicit transaction state formula: C t =C t-1 ⊙(1-U t )+C t ′⊙U t ; Among them, C t is the current time series implicit transaction status, C t-1 is the implicit transaction state of the previous time series, C t ′ is the candidate implicit transaction state of the previous time series, U t Update results for related party transactions.

6. The method according to claim 1, wherein The related party transaction update switch is for updating the neural network model for related party transactions, and the related party transaction reset is for resetting the neural network model for related party transactions; The method for processing transaction time series feature data further includes the following steps of tuning the hyperparameters of the related-party transaction update neural network model and the related-party transaction reset neural network model: Determine the training learning gradient for related-party transaction characteristics; Determining the learning momentum and second-order moment of the related-party transaction feature training according to the learning gradient; According to the momentum and the second-order moment, the current moment parameters of the financial transaction are updated to obtain an updated related-party transaction update neural network model and an updated related-party transaction reset neural network model.

7. The method according to claim 6, wherein Also includes: Correcting the momentum according to the momentum decay rate to obtain a corrected momentum, and correcting the second-order moment according to the second-order moment decay rate to obtain a corrected second-order moment; According to the momentum and the second-order moment, the parameters of the financial transaction at the current moment are updated to obtain an updated related-party transaction update neural network model and a related-party transaction reset neural network model, including: according to the corrected momentum and the corrected second-order moment, the parameters of the financial transaction at the current moment are updated to obtain an updated related-party transaction update neural network model and a related-party transaction reset neural network model.

8. The method according to claim 7, wherein Correcting the momentum according to the momentum decay rate to obtain a corrected momentum includes correcting the momentum according to the following formula to obtain a corrected momentum: in, is the modified momentum, M t is the momentum of the transaction at the current moment, is the momentum decay rate at the current time t of the transaction; Correcting the second-order moment according to the second-order moment attenuation rate to obtain a corrected second-order moment includes correcting the second-order moment according to the following formula to obtain a corrected second-order moment: in, is the corrected second-order moment, G t is the second moment of the transaction at the current moment, is the second-order moment decay rate of the transaction at the current time t.

9. The method according to claim 7, wherein Updating the current moment parameters of the financial transaction according to the momentum and the second-order moment includes updating the current moment parameters of the financial transaction according to the following formula: in, is the current moment parameter of financial transaction, is the parameter of the financial transaction at the previous moment, is the corrected momentum, is the corrected second-order moment, is the gradient update part of Adam, is an independent weight decay term.

10. The method according to claim 6, wherein Based on the learning gradient, determining the momentum and second-order moment of the related-party transaction feature training learning, including determining the momentum according to the following formula: M t =β1M t-1 +(1-β1)g t ; Among them, M t is the momentum of the transaction at the current moment, M t-1 is the momentum of the previous time series, β1 is the momentum decay rate, g t is the gradient of the transaction at the current time t.

11. The method according to claim 6, wherein Based on the learning gradient, determining the training learning momentum and second-order moment of the related-party transaction feature includes determining the second-order moment according to the following formula: G t =β2G t-1 +(1-β2)g t ☉g t ; Among them, G t is the second moment of the transaction at the current moment, G t-1 is the second-order moment of the previous time series, β2 is the second-order moment decay rate, g t is the gradient of the transaction at the current time t.

12. A device for processing transaction time series feature data, characterized in that: include: The acquisition unit is used to obtain the financial business transaction data of all dimensions of the related parties and construct the current time series financial transaction vector; An updating unit, configured to use a related-party transaction update switch to control whether the current time-series financial transaction vector needs to be updated, so as to determine the amount of transaction information of the previous time-series that needs to be retained in the current time-series transaction state, and obtain a related-party transaction update result; A reset unit, configured to control the degree of neglect of the previous time-series implicit transaction state of the current time-series financial transaction vector by using the related-party transaction reset switch, so as to determine the previous time-series implicit transaction state of the related-party transaction information and obtain a related-party transaction reset result; a candidate implicit transaction state determination unit, configured to determine the degree of dependence of the related-party transaction on the implicit transaction state of the previous time sequence according to the reset result of the related-party transaction party, and obtain the candidate implicit transaction state of the previous time sequence; The extraction unit is used to use the related-party transaction update result, the related-party transaction reset result and the candidate implicit transaction state of the previous time series to fuse and obtain the current time series implicit transaction state and its corresponding related-party transaction feature data.

13. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 11 is implemented.

14. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 11 is implemented.

15. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 11 is implemented.

Images