A transaction intention identification method and device

By obtaining transaction behavior sequences and using coding networks and graph neural networks to generate transactions and user representations, combined with intent identification networks, the problem of inaccurate identification of large-value transaction intentions in the existing technology is solved, and the effectiveness of transaction risk monitoring and management is improved .

CN114418767BActive Publication Date: 2024-11-12ALIPAY (HANGZHOU) INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210090764.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-25
Publication Date
2024-11-12
Estimated Expiration
2042-01-25

AI Technical Summary

Technical Problem

The existing technology is difficult to accurately identify the types of transaction intentions for large-scale transactions, resulting in insufficient monitoring and control measures for risky customers.

Method used

By obtaining the transaction behavior sequence, using the coding network and the graph neural network to fuse the transaction characteristics and user characteristics, generate transaction representations and user representations, and determine the transaction intent type in combination with the intention identification network.

Benefits of technology

It improves the accuracy of identification of transaction intention types and enhances the monitoring and risk management capabilities of risky transactions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114418767B_ABST
    Figure CN114418767B_ABST
Patent Text Reader

Abstract

The embodiments of the present specification provide a transaction intention recognition method and device, which obtain a transaction behavior sequence, which includes a target transaction to be identified and historical transactions between two users of the target transaction; the transaction behavior sequence is input into a coding network to obtain a corresponding transaction representation; in a pre-constructed user relationship network diagram, the two users are respectively used as target nodes, and the attribute features of the associated nodes are aggregated for the target nodes using a graph neural network to obtain user representations of both users; the user representations and transaction representations of both users are input into an intention recognition network to determine the transaction intention type of the target transaction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of transaction security technology, and more particularly to a transaction intention recognition method and device. Background Art

[0002] In the field of transaction security, large-value transaction identification and control are important components of financial institutions' compliance duties and are also the focus of previous regulatory inspections. Common types of transaction intentions of large-value customers include commodity transactions, service transactions, salary, capital transactions, and illegal transactions. It is crucial to accurately determine the transaction intention type of large-value transactions, and then determine whether the customer of the large-value transaction is a risky customer based on the transaction intention type, and take subsequent control measures against risky customers to effectively reduce transaction security risks within the platform.

[0003] Therefore, how to provide a method to accurately identify the transaction intention type of a transaction becomes an urgent problem to be solved. Summary of the invention

[0004] One or more embodiments of the present specification provide a transaction intention identification method and device to achieve accurate identification of the transaction intention type of a transaction.

[0005] According to a first aspect, a transaction intention recognition method is provided, comprising:

[0006] Acquire a transaction behavior sequence, wherein the transaction behavior sequence includes a target transaction to be identified and historical transactions between two users of the target transaction;

[0007] Inputting the transaction behavior sequence into the encoding network to obtain corresponding transaction representation;

[0008] In the pre-constructed user relationship network graph, the two users are respectively taken as target nodes, and the attribute features of the associated nodes are aggregated on the target nodes using the graph neural network to obtain the user representations of the two users respectively;

[0009] The user representations of the two users and the transaction representation are input into an intention recognition network to determine the transaction intention type of the target transaction.

[0010] In one possible implementation, the obtaining of the transaction behavior sequence includes:

[0011] Acquire transaction information of the target transaction, including the two users;

[0012] Obtain historical transactions between the two users within a preset time period before the target transaction;

[0013] The historical transactions and the target transactions are arranged in chronological order to form the transaction behavior sequence.

[0014] In one possible implementation, the graph neural network includes k aggregation layers;

[0015] The method of using the graph neural network to aggregate attribute features of associated nodes on the target node includes:

[0016] In the k aggregation layers, k-level vector aggregation is performed for the target node in sequence, wherein each level of vector aggregation includes performing neighbor aggregation on the previous level aggregation vector of each neighbor node in the neighbor node set of the target node, and determining the current level aggregation vector of the target node according to the neighbor aggregation result, wherein the primary aggregation vector of each node is determined according to the attribute characteristics of the corresponding user.

[0017] In one possible implementation, the aggregation layer includes a breadth exploration layer and a depth exploration layer;

[0018] Each level of vector aggregation includes:

[0019] In the breadth exploration layer, based on the previous level aggregation vector of the target node and the previous level aggregation vectors of each neighbor node, the weight value of each neighbor node relative to the target node is determined; based on each weight value, a number of target neighbor nodes are determined from the neighbor node set;

[0020] In the deep exploration layer, the current-level aggregation vector of the target node is determined according to the previous-level aggregation vector of the target node and the previous-level aggregation vectors of the plurality of target neighbor nodes.

[0021] In one possible implementation, the coding network includes a coding layer and a mapping layer;

[0022] The obtaining of the corresponding transaction representation includes:

[0023] At the encoding layer, the transaction features of each transaction in the transaction behavior sequence are encoded respectively to obtain an initial representation sequence;

[0024] At the mapping layer, the transaction representation is mapped based on the order relationship of the initial representations in the initial representation sequence and the initial representations therein.

[0025] In one possible implementation, the encoding network is one of the following: a Transformer network, a long short-term memory network (LSTM).

[0026] In one possible implementation, the edges of the user relationship network graph represent association relationships, and the association relationships include one of the following: social relationships, kinship relationships, fund transfer relationships, and same-machine relationships, and the same-machine relationship refers to logging in using the same device.

[0027] According to a second aspect, a transaction intention recognition device is provided, comprising:

[0028] An acquisition module configured to acquire a transaction behavior sequence, wherein the transaction behavior sequence includes a target transaction to be identified and historical transactions between two users of the target transaction;

[0029] A transaction representation acquisition module, configured to input the transaction behavior sequence into the encoding network to obtain a corresponding transaction representation;

[0030] A user representation acquisition module is configured to use a graph neural network to aggregate attribute features of associated nodes on the target nodes in a pre-built user relationship network graph, and obtain user representations of the two users respectively;

[0031] The type determination module is configured to input the user representations of the two users and the transaction representation into the intention recognition network to determine the transaction intention type of the target transaction.

[0032] In one possible implementation, the acquisition module is specifically configured to acquire transaction information of the target transaction, including the two users;

[0033] Obtain historical transactions between the two users within a preset time period before the target transaction;

[0034] The transaction information of the historical transactions and the target transactions are arranged in chronological order to form the transaction behavior sequence.

[0035] In one possible implementation, the graph neural network includes k aggregation layers;

[0036] The user representation acquisition module is specifically configured to: perform k-level vector aggregation on the target node in the k aggregation layers in sequence, wherein each level of vector aggregation includes performing neighbor aggregation on the previous level aggregation vector of each neighbor node in the neighbor node set of the target node, and determining the current level aggregation vector of the target node according to the neighbor aggregation result, wherein the primary aggregation vector of each node is determined according to the corresponding user attribute feature.

[0037] In one possible implementation, the aggregation layer includes a breadth exploration layer and a depth exploration layer;

[0038] The user representation acquisition module is specifically configured to: determine the weight value of each neighbor node relative to the target node based on the previous level aggregation vector of the target node and the previous level aggregation vector of each neighbor node at the breadth exploration layer; determine a number of target neighbor nodes from the neighbor node set based on each weight value;

[0039] In the deep exploration layer, the current-level aggregation vector of the target node is determined according to the previous-level aggregation vector of the target node and the previous-level aggregation vectors of the plurality of target neighbor nodes.

[0040] In one possible implementation, the coding network includes a coding layer and a mapping layer;

[0041] The transaction representation acquisition module is specifically configured to encode the transaction features of each transaction in the transaction behavior sequence at the encoding layer to obtain an initial representation sequence;

[0042] At the mapping layer, the transaction representation is mapped based on the order relationship of the initial representations in the initial representation sequence and the initial representations therein.

[0043] In one possible implementation, the encoding network is one of the following: a Transformer network, a long short-term memory network (LSTM).

[0044] In one possible implementation, the edges of the user relationship network graph represent association relationships, and the association relationships include one of the following: social relationships, kinship relationships, fund transfer relationships, and same-machine relationships, and the same-machine relationship refers to logging in using the same device.

[0045] According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute the method described in the first aspect.

[0046] According to a fourth aspect, a computing device is provided, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method described in the first aspect is implemented.

[0047] According to the method and device provided in the embodiments of this specification, the transaction representation corresponding to the target transaction is determined based on the transaction behavior sequence and the coding network. The obtained transaction representation not only integrates the transaction characteristics of historical transactions and target transactions, but also takes into account the behavioral characteristics of transactions between two users; and in the user relationship network diagram, the target node (two users) is aggregated with the attribute characteristics of the associated nodes using a graph neural network, so that the obtained user representation not only includes the attribute characteristics of the two users themselves, but also includes the attribute characteristics of their associated users. The obtained user representation can more comprehensively and accurately represent the user characteristics; combining the above-mentioned transaction representation and user representation can improve the recognition accuracy of the transaction intention type of the target transaction, thereby effectively monitoring transaction risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0049] Figure 1 A schematic diagram of an implementation framework of an embodiment disclosed in this specification;

[0050] Figure 2 A schematic diagram of a flow chart of a transaction intention identification method provided in an embodiment;

[0051] Figure 3 A schematic diagram of the exploration direction determined for the target node;

[0052] Figure 4 A schematic block diagram of a transaction intention identification device provided in an embodiment. DETAILED DESCRIPTION

[0053] The technical solutions of the embodiments of this specification will be described in detail below with reference to the accompanying drawings.

[0054] The embodiments of this specification disclose a transaction intention recognition method and device. The application scenarios and technical concepts of the method are first introduced as follows:

[0055] At present, the method for identifying the transaction intention type of a transaction is generally to determine the transaction intention type of the transaction based on the attribute characteristics of the two users of the transaction to be identified and the transaction characteristics of the transaction itself. The information used in this identification method is limited and cannot fully characterize the true characteristics of the transaction and its two users, which leads to the low accuracy of the determined transaction intention type.

[0056] In view of this, the inventor proposes a transaction intention recognition method to achieve more accurate recognition of the transaction intention type of a transaction. Figure 1 FIG. 2 shows a schematic diagram of an implementation framework according to an embodiment disclosed in this specification. Figure 1As shown, the transaction information of the target transaction to be identified is obtained, and the transaction information includes the users of both parties (specifically including the user identifications of the users of both parties), and the historical transactions between the users of both parties within a preset time period before the target transaction are obtained, and the historical transactions and the target transactions are arranged in chronological order to form a transaction behavior sequence. The transaction behavior sequence is input into the coding network, so that the transaction features of each transaction in the transaction behavior sequence are encoded through the coding network, and then the key information in each coding result is extracted, and each key information is mapped to generate a representation, so as to obtain the transaction representation corresponding to the target transaction. Accordingly, the transaction representation integrates the transaction features of the historical transactions and the target transaction, and takes into account the behavioral characteristics of the transactions between the users of both parties, and the representation content is more comprehensive.

[0057] In the pre-constructed user relationship network diagram, the two users are respectively taken as the target nodes, and the attribute features of the associated nodes are aggregated on the target nodes using the graph neural network to obtain the user representations of both users. Among them, the associated nodes can be all or part of the nodes in the multi-order (for example, k-order) neighbor nodes of the target node (selected by the graph neural network). The user representation integrates the attribute features of the user itself and the attribute features of the user's associated users (associated nodes). The representation content is more comprehensive and can better represent the user characteristics. The user representations and transaction representations of both users are input into the intention recognition network, so that the intention recognition network combines the user representation and the transaction representation to determine the transaction intention type of the target transaction.

[0058] It can be understood that the user relationship network diagram includes multiple nodes corresponding to multiple users, and edges that represent the association relationship between each user (node), wherein the association relationship may include one of the following: social relationship, kinship relationship, fund transfer relationship, and same-machine relationship, which may refer to (users) logging in using the same device.

[0059] Among them, the encoding network, graph neural network and intent recognition network are pre-trained based on training samples and their corresponding label data. The training samples include sample transaction sequences corresponding to several sample transactions, as well as the attribute characteristics of the sample two users of each sample transaction in the user relationship network graph, and the attribute characteristics of the associated nodes of the sample two users. Each sample transaction sequence includes a sample transaction and a sample historical transaction between its sample two users. The sample historical transaction corresponding to the sample transaction is a transaction between its sample two users within a preset time period before the sample transaction. The label data is used to characterize the real transaction intention type of the corresponding training sample. Transaction intention types include but are not limited to commodity transactions, service transactions, salaries, capital transactions and illegal transactions.

[0060] In this embodiment, the transaction representation not only integrates the transaction characteristics of historical transactions and target transactions, but also takes into account the behavioral characteristics of transactions between two users; the user representation not only includes the attribute characteristics of the two users themselves, but also includes the attribute characteristics of their associated users. The representation content of the two is more comprehensive. By using this type of transaction representation and user representation, the recognition accuracy of the transaction intention type of the target transaction can be improved, the recall rate can be improved, and the transaction risk can be effectively monitored, thereby realizing effective positioning and monitoring of users who conduct illegal transactions.

[0061] The transaction intention recognition method provided in this specification is described in detail below in conjunction with specific embodiments.

[0062] Figure 2 A flow chart of a transaction intention recognition method in one embodiment of this specification is shown. The method can be implemented by any device, equipment, platform, equipment cluster, etc. with computing and processing capabilities. The method includes the following steps S210-S240:

[0063] S210: Acquire a transaction behavior sequence, which includes a target transaction to be identified and historical transactions between two users of the target transaction. Each transaction in the transaction behavior sequence (including the target transaction and historical transactions) is arranged in chronological order. Each transaction corresponds to a number of transaction features, or each transaction is composed of a number of transaction features. The transaction features may be indicators related to the transaction and used to characterize the characteristics of the transaction, and may include but are not limited to: transaction time, transaction amount, transaction ID (transaction identifier), transaction remarks, and transaction product code.

[0064] In one implementation, the two users of the target transaction may be users of a third-party payment platform, and accordingly, the target transaction and historical transactions may be transactions generated by the two users in the third-party payment platform, and the user relationship network diagram is established for each user in the third-party payment platform. Alternatively, the two users may be users of a certain banking institution, and accordingly, the target transaction and historical transactions may be transactions generated by the two users in the banking institution, and the user relationship network diagram is established for each user in the banking institution. Alternatively, the two users may be users of a certain social platform, and accordingly, the target transaction and historical transactions may be transactions generated by the two users in the social platform, and the user relationship network diagram is established for each user in the social platform.

[0065] In another implementation, the platform where the transaction is generated may be different from the platform where the user relationship network diagram is established. For example, the two users of the target transaction may be users of a payment platform and users of another social platform. The payment platform can obtain the user relationship network diagram established by the social platform through cooperation and conduct subsequent transaction analysis.

[0066] In another implementation, when both users of the target transaction are users of a third-party payment platform, the target transaction or historical transaction can be a transaction generated by both users within the third-party payment platform, or a transaction generated by both users within other bound platforms, where the other bound platforms are platforms that authorize the third-party payment platform to obtain corresponding transaction information.

[0067] In one case, the target transaction may be a large-value transaction, and a large-value transaction may refer to a transaction whose transaction amount exceeds a preset amount threshold.

[0068] In one implementation, the S210 may include the following steps 01-03:

[0069] In step 01: the transaction information of the target transaction X is obtained, including both users (specifically including the user identifications of both users, such as user accounts, nicknames, and mobile phone numbers, etc.). The two users refer to the outgoing party and the incoming party of the target transaction.

[0070] Step 02: Obtain historical transactions Y between the two users within a preset time period before the target transaction. The preset time period can be set according to actual needs. The historical transactions can all be large-value transactions, or can include both large-value transactions and small-value transactions, wherein the small-value transaction refers to a transaction whose transaction amount does not exceed the aforementioned preset amount threshold.

[0071] In step 03: historical transactions and target transactions are arranged in chronological order to form a transaction behavior sequence. The historical transactions and target transactions can be arranged in order of transaction time from early to late (or from late to early) to form a transaction behavior sequence, for example, it can be expressed as [Yn…Y1, X], where the transaction time is getting later and later from Yn to Y1 and then to X. The transaction behavior sequence can include the transaction characteristics of each transaction.

[0072] After obtaining the transaction behavior sequence in step S210, in step S220, the transaction behavior sequence is input into the coding network to obtain the corresponding transaction representation. In this step, the transaction behavior sequence is input into the coding network to encode the transaction features of each transaction in the transaction behavior sequence through the coding network, and then the key information in each coding result is extracted, and each key information is mapped to generate a single representation, so as to obtain the transaction representation corresponding to the target transaction. At this time, the transaction representation not only integrates the transaction features of historical transactions and target transactions, but also takes into account the behavioral characteristics of transactions between two users, providing a basis for the accurate identification of subsequent transaction intention types.

[0073] In one embodiment, the coding network may include a coding layer and a mapping layer;

[0074] Correspondingly, at the encoding layer, the transaction features of each transaction in the transaction behavior sequence are encoded respectively to obtain an initial representation sequence. Specifically, at the encoding layer, each transaction feature of each transaction is encoded (e.g., embedding processing), and the encoded vector is used as the initial representation of each transaction. Alternatively, the encoded vector can also be linearly or nonlinearly transformed, and the transformed vector is used as the initial representation of each transaction. In this way, an initial representation sequence composed of the initial representations of each transaction is obtained.

[0075] Subsequently, at the mapping layer, based on the order relationship of each initial representation in the initial representation sequence and each initial representation therein, the transaction representation is mapped. Specifically, at the mapping layer, key information can be extracted from each initial representation according to the order relationship of each initial representation, and then the key information of each initial representation extracted is mapped according to the order relationship of each initial representation to obtain the transaction representation.

[0076] In one embodiment, the encoding network may be a Transformer network. After the initial representation sequence is obtained by encoding through the encoding layer of the Transformer network, each initial representation is integrated into the position encoding information (determined based on the order relationship of each initial representation), and then input into the mapping layer of the Transformer network, so as to capture the context information between each initial representation through the Transformer network, extract the attention weight information between each initial representation, and then map the initial representation sequence into a representation based on the attention weight, so that the output of the Transformer network is used as the transaction representation of the target transaction.

[0077] In another embodiment, the encoding network may be a long short-term memory (LSTM) network. After the initial representation sequence is obtained by encoding through the encoding layer of the LSTM network, each initial representation in the initial representation sequence may be sequentially input into the mapping layer of the LSTM network, so that the output of the LSTM network is used as the transaction representation of the target transaction. In another embodiment, the encoding network may also be other neural networks that process sequence data, such as an RNN network.

[0078] S230: In the pre-constructed user relationship network diagram, the two users are respectively taken as target nodes, and the attribute features of the associated nodes are aggregated on the target nodes using the graph neural network to obtain the user representations of the two users. The user relationship network diagram includes multiple nodes corresponding to multiple users, and edges representing the association relationship between each user (node), wherein the association relationship may include at least one of the following: social relationship, kinship relationship, fund transfer relationship, and same-machine relationship, wherein the same-machine relationship refers to logging in using the same device. The user relationship network diagram may be as follows: Figure 3 As shown, nodes A and B correspond to both users respectively.

[0079] It can be understood that step S220 may be executed before step S230, or step S230 may be executed before step S220, or step S220 and step S230 may be executed in parallel.

[0080] Understandably, Figure 3 As only an example diagram of the user relationship network diagram, the data volume of the user relationship network diagram may be very large, that is, it includes a large number of users and the relationships between them. In order to reduce the amount of calculation to a certain extent, in this step, the relationship network subgraphs corresponding to the two users can be determined in advance from the user relationship network diagram, and then, the two users are respectively taken as target nodes, and in their respective corresponding relationship network subgraphs, the attribute features of the associated nodes are aggregated for the target nodes using the graph neural network to obtain the user representations of the two users. Among them, the relationship network subgraph corresponding to either of the two users at least includes the target node corresponding to the party and the k-order neighbor nodes of the target node, and the associated nodes of the target node can be some or all of the nodes in the k-order neighbor nodes.

[0081] The above-mentioned attribute features may be some indicators related to the user represented by the node and used to characterize the characteristics of the user, and may include but are not limited to the user's hobbies, occupations, and crowd labels.

[0082] In one embodiment of the present specification, the graph neural network may include k aggregation layers;

[0083] The method of using a graph neural network to aggregate the attribute features of associated nodes of a target node may include the following steps: performing k-level vector aggregation on the target node in k aggregation layers in sequence, wherein each level of vector aggregation includes performing neighbor aggregation on the previous level aggregation vector of each neighbor node in the neighbor node set of the target node, and determining the current level aggregation vector of the target node based on the neighbor aggregation result, wherein the primary aggregation vector of each node is determined based on the attribute features of the corresponding user.

[0084] Specifically, in order to perform k-level vector aggregation for the target node v in k aggregation layers in sequence, first, the primary aggregation vector of the target node v is determined according to the attribute characteristics of the target node v. And according to the attribute characteristics of each neighbor node u, determine the primary aggregation vector of each neighbor node u Then based on the primary aggregation vector and the primary aggregation vector Perform k-level vector aggregation for the target node v to obtain the aggregated vector of the k-th aggregation layer As the user representation of the user corresponding to the target node v.

[0085] Among them, in each level of vector aggregation process, each neighbor node u in the above neighbor node set is a node connected to the target node v. It can be understood that at each aggregation layer, in addition to performing vector aggregation on the target node, it is also necessary to perform vector aggregation on each neighbor node to obtain the aggregation vector of this level of each neighbor node. When performing vector aggregation on each neighbor node, it is necessary to use the aggregation vector of the previous level of the neighbor node and the aggregation vector of the previous level of its neighbor node, and so on, so that at the i-th aggregation layer, the aggregation vector of this level of the target node obtained by aggregation contains the attribute characteristics of all or part of the nodes in the i-th order neighbor nodes of the target node. In this way, after aggregation of k aggregation layers, the aggregation vector of the target node is The attribute features of the target node v and its associated nodes are aggregated in .

[0086] In one embodiment, the above determination of the primary aggregation vector The process can be: encode each attribute feature of the target node v, and use the encoded vector as the primary aggregation vector Alternatively, the encoded vector can be linearly or nonlinearly transformed and the transformed vector can be used as the primary aggregation vector

[0087] The above neighbor aggregation operation can be performed by the aggregation function AGG i In one embodiment, for a target node v, its i-th level vector aggregation (i.e., vector aggregation at the i-th aggregation layer) may include first using the aggregation function AGG i , according to the previous level (i.e. level i-1) aggregation vector of the neighbor node u of the target node v Get the neighbor aggregation vector That is, the neighbor aggregation result, where N(v) represents the neighbor node set of the target node v, that is:

[0088]

[0089] Then, according to the neighbor aggregation vector And the previous level (i.e. level i-1) aggregation vector of the target node v Determine the current level (level i) aggregation vector of the target node v Right now:

[0090]

[0091] Where f represents the neighbor aggregation vector and the target node v's previous level vector The applied comprehensive function, W i is the parameter of the i-th level aggregation. In different embodiments, the comprehensive operation in function f may include: and Concatenate, sum, average, etc.

[0092] The aggregation function AGG for the above neighbor aggregation operation i Can take different forms and algorithms, for example: Aggregation function AGG i It can be, but is not limited to, a pooling operation, a summation operation, etc.

[0093] Taking into account that the importance of each neighbor node in the neighbor node set of the target node is different relative to the target node, only the neighbor nodes that are important to the target node can be used for aggregation to obtain a more accurate aggregation vector (of the target node), which can be specifically selected by the graph neural network. In one embodiment, the aggregation layer includes a breadth exploration layer and a depth exploration layer; wherein the breadth exploration layer can determine the importance of each neighbor node in the neighbor node set of the target node relative to the target node, and then determine the target neighbor nodes participating in the aggregation based on the importance, that is, to achieve the exploration and determination of the aggregation direction. The deep exploration layer can retain useful information (relative to the target node) from the upper-level aggregation vector of the target neighbor nodes participating in the aggregation based on its memory characteristics, forget useless information, and obtain an aggregation vector that represents the content more accurately, thereby achieving the exploration and determination of the aggregation depth.

[0094] Accordingly, in each level of vector aggregation, the following steps 11-12 are included:

[0095] In step 11, in the breadth exploration layer, based on the previous level aggregation vector of the target node and the previous level aggregation vector of each neighbor node, determine the weight value of each neighbor node relative to the target node; based on each weight value, determine a number of target neighbor nodes from the neighbor node set; wherein each weight value represents the importance of the corresponding neighbor node relative to the target node, and the larger the weight value, the higher the importance. Accordingly, the M neighbor nodes with the largest relative weight values ​​in the neighbor node set can be used as a number of target neighbor nodes. Then, the previous level aggregation vectors of the several target neighbor nodes and the previous level aggregation vector of the target node are sequentially input into the depth exploration layer. Figure 3 As shown, in the i-th aggregation layer, the target neighbor nodes determined for the target node (node ​​A) are B and C, and the exploration direction is the direction where nodes B and C are located. Accordingly, the previous level aggregation vectors of nodes B and C and the previous level aggregation vector of target node A need to be input into the deep exploration layer in sequence.

[0096] Next, in step 12, in the deep exploration layer, the current-level aggregation vector of the target node is determined based on the previous-level aggregation vector of the target node and the previous-level aggregation vectors of several target neighboring nodes. Specifically, in the deep exploration layer, the useful information in the previous-level aggregation vector of the input node is memorized (retained) by using its memory characteristics, and the useless information in the previous-level aggregation vector of the input node is forgotten, so as to determine the current-level aggregation vector of the target node. Figure 3 As shown, in the deep exploration layer, it is necessary to explore the neighboring nodes of node B and node C. In one implementation, the deep exploration layer can be an LSTM network, a GRU gate control network, or other neural networks for processing sequence data.

[0097] It can be understood that the primary aggregation vector of the target node v can be used as the 0th level aggregation vector. Based on this, in the 1st to the kth aggregation layers, the aggregation is performed on the target node step by step, and the output vector of the kth aggregation layer of the target node v can be obtained. By taking both users as target nodes respectively, the user representations of both users can be obtained respectively.

[0098] S240: Input the user representation and transaction representation of both users into the intention recognition network to determine the transaction intention type of the target transaction. In this step, the user representation and transaction representation of both users can be spliced ​​to obtain a spliced ​​representation, and then the spliced ​​representation is input into the intention recognition network to determine the transaction intention type of the target transaction. In one implementation, the intention recognition network can be a DNN (Deep Neural Networks).

[0099] In this embodiment, the transaction representation not only integrates the transaction characteristics of historical transactions and target transactions, but also takes into account the behavioral characteristics of transactions between two users; the user representation not only includes the attribute characteristics of the two users themselves, but also includes the attribute characteristics of their associated users, which can more comprehensively and accurately represent the two users. The use of transaction representation and user representation with more comprehensive representation content can improve the recognition accuracy of the transaction intention type of the target transaction, thereby effectively monitoring transaction risks and realizing effective positioning monitoring of users who conduct illegal transactions. In addition, in the case where the target transaction is a large-value transaction, the transaction risk monitoring of large-value transactions can be realized.

[0100] The foregoing describes certain embodiments of the present specification, and other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in an order different from that in the embodiments, and the desired results may still be achieved. In addition, the processes depicted in the accompanying drawings do not necessarily have to be performed in the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0101] Corresponding to the above method embodiment, the present specification embodiment provides a transaction intention recognition device 400, whose schematic block diagram is as follows: Figure 4 As shown, including:

[0102] An acquisition module 410 is configured to acquire a transaction behavior sequence, wherein the transaction behavior sequence includes a target transaction to be identified and historical transactions between two users of the target transaction;

[0103] The transaction representation obtaining module 420 is configured to input the transaction behavior sequence into the encoding network to obtain the corresponding transaction representation;

[0104] The user representation acquisition module 430 is configured to use the graph neural network to aggregate the attribute features of the associated nodes on the target nodes in the pre-constructed user relationship network graph, so as to obtain the user representations of the two users respectively;

[0105] The type determination module 440 is configured to input the user representations of the two users and the transaction representation into the intention recognition network to determine the transaction intention type of the target transaction.

[0106] In one possible implementation, the acquisition module 410 is specifically configured to acquire transaction information of the target transaction, including the two users;

[0107] Obtain historical transactions between the two users within a preset time period before the target transaction;

[0108] The transaction information of the historical transactions and the target transactions are arranged in chronological order to form the transaction behavior sequence.

[0109] In one possible implementation, the graph neural network includes k aggregation layers;

[0110] The user representation obtaining module 430 is specifically configured to: perform k-level vector aggregation for the target node in the k aggregation layers in sequence, wherein each level of vector aggregation includes performing neighbor aggregation on the previous level aggregation vector of each neighbor node in the neighbor node set of the target node, and determining the current level aggregation vector of the target node according to the neighbor aggregation result, wherein the primary aggregation vector of each node is determined according to the corresponding user attribute feature.

[0111] In one possible implementation, the aggregation layer includes a breadth exploration layer and a depth exploration layer;

[0112] The user representation acquisition module 430 is specifically configured to: determine the weight value of each neighbor node relative to the target node based on the previous level aggregation vector of the target node and the previous level aggregation vector of each neighbor node at the breadth exploration layer; determine a number of target neighbor nodes from the neighbor node set based on each weight value;

[0113] In the deep exploration layer, the current-level aggregation vector of the target node is determined according to the previous-level aggregation vector of the target node and the previous-level aggregation vectors of the plurality of target neighbor nodes.

[0114] In one possible implementation, the coding network includes a coding layer and a mapping layer;

[0115] The transaction representation acquisition module 420 is specifically configured to encode the transaction features of each transaction in the transaction behavior sequence at the encoding layer to obtain an initial representation sequence; and at the mapping layer, map the transaction representation based on the sequential relationship of each initial representation in the initial representation sequence and each initial representation therein.

[0116] In one possible implementation, the encoding network is one of the following: a Transformer network, a long short-term memory network (LSTM).

[0117] In one possible implementation, the edges of the user relationship network graph represent association relationships, and the association relationships include one of the following: social relationships, kinship relationships, fund transfer relationships, and same-machine relationships, and the same-machine relationship refers to logging in using the same device.

[0118] The above device embodiments correspond to the method embodiments. For specific descriptions, please refer to the description of the method embodiments, which will not be repeated here. The device embodiments are obtained based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments. For specific descriptions, please refer to the corresponding method embodiments.

[0119] The embodiments of the present specification also provide a computer-readable storage medium on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute the method for transferring assets in the payment platform provided in the present specification.

[0120] The embodiment of the present specification also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method for transferring assets in the payment platform provided in the present specification is implemented.

[0121] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the storage medium and computing device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0122] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the embodiments of the present invention can be implemented by hardware, software, firmware, or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0123] The specific implementation methods described above further describe the purpose, technical solutions and beneficial effects of the embodiments of the present invention in detail. It should be understood that the above description is only a specific implementation method of the embodiments of the present invention and is not intended to limit the scope of protection of the present invention. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solution of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A transaction intention recognition method, comprising: Acquire a transaction behavior sequence, wherein the transaction behavior sequence includes a target transaction to be identified and historical transactions between two users of the target transaction; Inputting the transaction behavior sequence into the encoding network to obtain corresponding transaction representation; In the pre-constructed user relationship network graph, the two users are respectively taken as target nodes, and the attribute features of the associated nodes are aggregated on the target nodes using the graph neural network to obtain the user representations of the two users respectively; The user representations of both users and the transaction representation are input into the intention recognition network to determine the transaction intention type of the target transaction, wherein the encoding network, the graph neural network and the intention recognition network are pre-trained based on training samples and their corresponding label data, and the training samples include sample transaction sequences corresponding to several sample transactions, and the label data is used to characterize the actual transaction intention type of the corresponding training samples.

2. The method of claim 1, wherein: The obtaining of the transaction behavior sequence includes: Acquire transaction information of the target transaction, including the two users; Obtain historical transactions between the two users within a preset time period before the target transaction; The historical transactions and the target transactions are arranged in chronological order to form the transaction behavior sequence.

3. The method of claim 1, wherein: The graph neural network includes k aggregation layers; The method of using the graph neural network to aggregate attribute features of associated nodes on the target node includes: In the k aggregation layers, k-level vector aggregation is performed for the target node in sequence, wherein each level of vector aggregation includes performing neighbor aggregation on the previous level aggregation vector of each neighbor node in the neighbor node set of the target node, and determining the current level aggregation vector of the target node according to the neighbor aggregation result, wherein the primary aggregation vector of each node is determined according to the attribute characteristics of the corresponding user.

4. The method of claim 3, wherein: The aggregation layer includes a breadth exploration layer and a depth exploration layer; Each level of vector aggregation includes: In the breadth exploration layer, based on the previous level aggregation vector of the target node and the previous level aggregation vectors of each neighbor node, the weight value of each neighbor node relative to the target node is determined; based on each weight value, a number of target neighbor nodes are determined from the neighbor node set; In the deep exploration layer, the current-level aggregation vector of the target node is determined according to the previous-level aggregation vector of the target node and the previous-level aggregation vectors of the plurality of target neighbor nodes.

5. The method of claim 1, wherein: The coding network includes a coding layer and a mapping layer; The obtaining of the corresponding transaction representation includes: At the encoding layer, the transaction features of each transaction in the transaction behavior sequence are encoded respectively to obtain an initial representation sequence; At the mapping layer, the transaction representation is mapped based on the order relationship of the initial representations in the initial representation sequence and the initial representations therein.

6. The method of claim 5, wherein: The encoding network is one of the following: a Transformer network, a long short-term memory network LSTM.

7. The method according to any one of claims 1 to 6, wherein: The edges of the user relationship network graph represent association relationships, and the association relationships include one of the following: social relationships, kinship relationships, fund transfer relationships, and same-machine relationships, and the same-machine relationship refers to logging in using the same device.

8. A transaction intention recognition device, comprising: An acquisition module configured to acquire a transaction behavior sequence, wherein the transaction behavior sequence includes a target transaction to be identified and historical transactions between two users of the target transaction; A transaction representation acquisition module, configured to input the transaction behavior sequence into the encoding network to obtain a corresponding transaction representation; A user representation acquisition module is configured to use a graph neural network to aggregate attribute features of associated nodes on the target nodes in a pre-built user relationship network graph, and obtain user representations of the two users respectively; The type determination module is configured to input the user representations of the two users and the transaction representation into the intention recognition network to determine the transaction intention type of the target transaction, wherein the encoding network, the graph neural network and the intention recognition network are pre-trained based on training samples and their corresponding label data, and the training samples include sample transaction sequences corresponding to several sample transactions, and the label data is used to characterize the actual transaction intention type of the corresponding training samples.

9. The device of claim 8, wherein: The acquisition module is specifically configured to acquire transaction information of the target transaction, including the two users; Obtain historical transactions between the two users within a preset time period before the target transaction; The transaction information of the historical transactions and the target transactions are arranged in chronological order to form the transaction behavior sequence.

10. The device of claim 8, wherein: The graph neural network includes k aggregation layers; The user representation acquisition module is specifically configured to: perform k-level vector aggregation on the target node in the k aggregation layers in sequence, wherein each level of vector aggregation includes performing neighbor aggregation on the previous level aggregation vector of each neighbor node in the neighbor node set of the target node, and determining the current level aggregation vector of the target node according to the neighbor aggregation result, wherein the primary aggregation vector of each node is determined according to the corresponding user attribute feature.

11. The device of claim 10, wherein: The aggregation layer includes a breadth exploration layer and a depth exploration layer; The user representation acquisition module is specifically configured to: determine the weight value of each neighbor node relative to the target node based on the previous level aggregation vector of the target node and the previous level aggregation vector of each neighbor node at the breadth exploration layer; Based on the respective weight values, determining a number of target neighbor nodes from the neighbor node set; In the deep exploration layer, the current-level aggregation vector of the target node is determined according to the previous-level aggregation vector of the target node and the previous-level aggregation vectors of the plurality of target neighbor nodes.

12. The device of claim 8, wherein: The coding network includes a coding layer and a mapping layer; The transaction representation acquisition module is specifically configured to encode the transaction features of each transaction in the transaction behavior sequence at the encoding layer to obtain an initial representation sequence; At the mapping layer, the transaction representation is mapped based on the order relationship of the initial representations in the initial representation sequence and the initial representations therein.

13. The device of claim 12, wherein: The mapping layer is one of the following: a Transformer network, a long short-term memory network LSTM.

14. The device according to any one of claims 8 to 13, wherein: The edges of the user relationship network graph represent association relationships, and the association relationships include one of the following: social relationships, kinship relationships, fund transfer relationships, and same-machine relationships, and the same-machine relationship refers to logging in using the same device.

15. A computing device comprising a memory and a processor, wherein: The memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 7 is implemented.