Risk control method, risk order determination method and related equipment

Abstract

The invention provides a risk control method, a risk order determination method and related equipment. Relates to the technical field of risk control. The method comprises the following steps: acquiring multiple pieces of source address information of a target order and corresponding target address information of a standard structure; performing multi-task address matching on the source address information and the target address information through a multi-task matching model, and determining the source address information matched with the target address information; the multi-task matching model comprises a first sub-task based on text distance matching and a second sub-task based on text semantic matching; and in response to the fact that the quantity of the source address information matched with the target address information exceeds a preset first threshold value, determining the target address information as a risk address so as to perform risk control based on the risk address. According to the method and the device, the problem of low prediction accuracy of the spatial directivity of the multi-source heterogeneous massive address data in related technologies can be solved.

Description

Technical Field

[0001] The present disclosure relates to the technical field of risk control, and in particular to a risk control method, a risk order determination method and related equipment. Background Art

[0002] With the popularization of location services and mobile devices, a large number of vertical industries based on addresses as spatial information have emerged. It is particularly important to accurately identify and control the various risks brought about by this.

[0003] Risks can be identified by analyzing the address information of a large number of orders. However, due to the different input / acquisition channels of address information, it is difficult to accurately determine the spatial directionality of multi-source heterogeneous massive address data.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0005] The purpose of the embodiments of the present disclosure is to provide a risk control method, a risk order determination method and related equipment, thereby solving the problem of low prediction accuracy of the spatial directionality of multi-source heterogeneous massive address data in the related art to a certain extent.

[0006] According to a first aspect of the present disclosure, a risk control method is provided, comprising: obtaining multiple source address information of a target order and corresponding target address information of a standard structure; the source address information comprises address information of different structures from multiple sources; performing multi-task address matching on each of the source address information and the target address information through a multi-task matching model, and determining the source address information that matches the target address information; the matching result of the multi-task matching model is determined based on the matching result of each subtask, the multi-task matching model comprises a first subtask and a second subtask, the first subtask is based on text distance matching, and the second subtask is based on text semantic matching; in response to the number of source address information that matches the target address information exceeding a preset first threshold, determining that the target address information is a risky address, and performing risk control based on the risky address.

[0007] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the address matching of the source address information and the target address information includes: determining a first matching degree based on distance information between each source address information and the target address information through the first subtask; determining a first text similarity based on semantic information between each source address information and the target address information through the second subtask; fusing the first matching degree with the corresponding first text similarity to obtain a matching degree between each source address information and the target address information; in response to the matching degree being greater than a preset threshold, determining that the source address information matches the target address information.

[0008] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the first subtask includes a corresponding feature extraction network and an upper tower network, the network structure of the second subtask is the same as that of the first subtask, and feature information is shared between the feature extraction networks of the first subtask and the second subtask; determining the first distance information through the first subtask includes: extracting text distance features of the source address information and the target address information through the feature extraction network corresponding to the first subtask to obtain corresponding first feature vectors and second feature vectors; determining the first distance information between the first feature vector and the second feature vector through the upper tower network corresponding to the first subtask, and determining the first matching degree based on the first distance information.

[0009] The determining of the first text similarity through the second subtask includes: extracting text semantic features of the source address information and the target address information through a feature extraction network corresponding to the second subtask to obtain corresponding third feature vectors and fourth feature vectors; determining the similarity between the third feature vector and the fourth feature vector through an upper tower network corresponding to the second subtask, and using the similarity as the first text similarity.

[0010] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the method further includes: obtaining a training sample set, the training samples in the training sample set including address pairs consisting of a source address and a target address, a first label corresponding to a first subtask and a second label corresponding to a second subtask, the first label being determined based on the second distance information between the address pairs of the training samples, and the second label being determined based on the second text similarity between the address pairs of the training samples; training the first subtask and the second subtask respectively through the training sample set to determine the model parameters of the multi-task matching model for risk control.

[0011] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the first subtask and the second subtask are trained respectively through the training sample set, including: inputting the training samples in the training sample set into the first subtask and the second subtask respectively, obtaining a first output corresponding to the first subtask and a second output corresponding to the second subtask; determining a first loss function according to the first output and the corresponding first label; determining a second loss function according to the second output and the corresponding second label; performing weighted summation of the first loss function and the second loss function to determine a target loss function; adjusting the model parameters of the multi-task matching model based on the target loss function; wherein, the weights of the weighted summation in different training rounds are updated through weight update factors.

[0012] In an exemplary embodiment of the present disclosure, based on the aforementioned scheme, the method further includes: in response to a ratio of a second distance information between an address pair of a training sample and a longer character length in the address pair being greater than a preset second threshold, determining that the first label indicates a successful match; in response to a second text similarity between an address pair of a training sample being greater than a preset third threshold, determining that the second label indicates a successful match.

[0013] According to a second aspect of the present disclosure, a method for determining risky orders is provided, the method comprising: obtaining address information of each order to be predicted; matching the address information with a risk address in a risk address list through a risk prediction model to obtain a corresponding target matching degree; the risk address list is formed based on the risk addresses determined in any of the aforementioned embodiments; the network structure of the risk prediction model is the same as that of the multi-task matching model, and the risk prediction model is obtained by training with a training sample set consisting of risk addresses and risk address lists; in response to the target matching degree being greater than a preset fourth threshold, determining that the order corresponding to the address information is a risky order.

[0014] According to a third aspect of the present disclosure, a risk control device is provided, the device comprising: an acquisition module, used to acquire multiple source address information of a target order and corresponding target address information of a standard structure; the source address information comprises address information of different structures from multiple sources; a matching module, used to perform multi-task address matching on each of the source address information and the target address information through a multi-task matching model, and determine the source address information that matches the target address information; the matching result of the multi-task matching model is determined based on the matching result of each subtask, the multi-task matching model comprises a first subtask and a second subtask, the first subtask is based on text distance matching, and the second subtask is based on text semantic matching; a control module, used to determine that the target address information is a risky address in response to the number of source address information that matches the target address information exceeding a preset first threshold, and perform risk control based on the risky address.

[0015] According to a fourth aspect of the present disclosure, a risk control device is provided, comprising: an acquisition module for acquiring address information of each order to be predicted; a matching module for matching the address information with a risk address in a risk address list through a risk prediction model to obtain a corresponding target matching degree; the risk address list is formed based on the risk addresses determined in any of the aforementioned embodiments; the network structure of the risk prediction model is the same as that of the multi-task matching model, and the risk prediction model is obtained by training a training sample set consisting of risk addresses and risk address lists; a determination module for determining that the order corresponding to the address information is a risky order in response to the target matching degree being greater than a preset fourth threshold.

[0016] According to a fifth aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method described in any one of the above embodiments is implemented.

[0017] According to a sixth aspect of the present disclosure, an electronic device is provided, comprising: one or more processors; and a storage device for one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the method described in any one of the above embodiments.

[0018] The exemplary embodiments of the present disclosure may have some or all of the following beneficial effects:

[0019] In the risk control method provided in the exemplary implementation of the present disclosure, on the one hand, the text matching of the first subtask can be combined with the semantic matching of the second subtask to mine the text information and semantic information in the source address information and the target address information, reduce the interference caused by the different sources of the source address information, and improve the accuracy of address matching. On the other hand, introducing a multi-task matching model into the risk control process based on address matching can simultaneously solve the problem of partial semantics that cannot be captured due to word-by-word matching, and the problem of low matching accuracy due to high semantic relevance but inconsistent spatial directivity, thereby improving the accuracy of risk control.

[0020] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification are used to explain the principles of the present disclosure. Obviously, the accompanying drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0022] Figure 1 The flowchart of the risk control method according to one embodiment of the present disclosure is schematically shown.

[0023] Figure 2 The following schematically shows a schematic diagram of an address matching process based on a multi-task matching model according to an embodiment of the present disclosure.

[0024] Figure 3 The following schematically shows a process diagram of determining the first text similarity of the second subtask according to an embodiment of the present disclosure.

[0025] Figure 4 The flowchart of the training process of the multi-task matching model according to one embodiment of the present disclosure is schematically shown.

[0026] Figure 5 A flowchart of the implementation process of the risk control method according to an embodiment of the present disclosure is schematically shown.

[0027] Figure 6 The following schematically shows a flowchart of the implementation process of a risk order determination method according to an embodiment of the present disclosure.

[0028] Figure 7 The structural block diagram of a risk control device according to an embodiment of the present disclosure is schematically shown.

[0029] Figure 8The structural block diagram of a risk order determination device according to an embodiment of the present disclosure is schematically shown.

[0030] Fig. 9 A block diagram of an electronic device suitable for implementing the embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0031] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as being limited to the examples set forth herein; on the contrary, these embodiments are provided so that the present disclosure will be more comprehensive and complete, and the concepts of the example embodiments are fully conveyed to those skilled in the art. The described features, structures, or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced while omitting one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known technical solutions are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0032] In addition, the accompanying drawings are only schematic illustrations of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor devices and / or microcontroller devices.

[0033] The present disclosure provides a risk control method. This embodiment uses the method applied to a server as an example. It is understandable that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms, or a node in a blockchain. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, a car-mounted device, etc., but is not limited thereto. When the risk control method provided in this embodiment is implemented through the interaction between a terminal and a server, the terminal and the server can be directly or indirectly connected by wired or wireless communication, and the present disclosure does not limit this.

[0034] The risk control method provided in the embodiment of the present disclosure may be executed on a server, and accordingly, the risk control device is generally disposed in the server.

[0035] Address consistency verification is a key link in the fields of financial risk control, and can be applied to many related fields such as anti-fraud and credit map construction. Through the consistency verification of order-related addresses, it can be determined whether there is a fraud risk, and then timely discovered and handled to reduce platform risks. This disclosure designs a risk control method based on address consistency verification for network fraud.

[0036] refer to Figure 1 As shown, a risk control method according to an exemplary embodiment of the present disclosure may include the following steps.

[0037] Step S110, obtaining multiple source address information of the target order and corresponding target address information of the standard structure; the source address information includes address information of different structures from multiple sources.

[0038] Step S120, performing multi-task address matching on each source address information and the target address information through a multi-task matching model to determine the source address information that matches the target address information; the matching result of the multi-task matching model is determined based on the matching result of each subtask, and the multi-task matching model includes a first subtask and a second subtask, the first subtask is based on text distance matching, and the second subtask is based on text semantic matching.

[0039] Step S130: In response to the number of source address information matched by the target address information exceeding a preset first threshold, the target address information is determined to be a risky address, and risk control is performed based on the risky address.

[0040] In the risk control method provided in this example implementation, on the one hand, the text matching of the first subtask can be combined with the semantic matching of the second subtask to mine the text information and semantic information in the source address information and the target address information, reduce the interference caused by the different sources of the source address information, and improve the accuracy of address matching. On the other hand, introducing a multi-task matching model into the risk control process based on address matching can simultaneously solve the problem of partial semantics that cannot be captured due to word-by-word matching, and the problem of low matching accuracy due to high semantic relevance but inconsistent spatial directivity, thereby improving the accuracy of risk control.

[0041] The following describes each step of the present disclosure in more detail.

[0042] In step S110, multiple source address information of the target order and corresponding target address information of the standard structure are obtained.

[0043] In this example implementation, the target order may be an order to be evaluated within a specified period. Such as an application platform order in the past month or the past week. The source address information may include address information of different structures from multiple sources. The source address information refers to address information directly input or obtained from various sources, such as addresses recognized by OCR (Optical Character Recognition), addresses manually input by business personnel, addresses input by online orders, etc. The target address information refers to an address of a standard text structure stored in a database, such as XX Province XX City XX District / County XX Street / Township XXX (detailed address). The source address information and the target address information may be a rough address of a larger range (such as XX District, XX City), or a detailed address of a smaller range (such as XX Street XX Town XX House Number, XX District, XX City), which is not limited in this example. The source address information has different address sources and different structures, such as different address details, different address text formats (such as different uppercase and lowercase letters, full-width and half-width differences), etc., which may interfere with the address matching process. The source address information and the target address information form an address pair to be matched.

[0044] Step S120, performing multi-task address matching on each source address information and the target address information through a multi-task matching model to determine the source address information that matches the target address information; the matching result of the multi-task matching model is determined based on the matching result of each subtask, and the multi-task matching model includes a first subtask and a second subtask, the first subtask is based on text distance matching, and the second subtask is based on text semantic matching.

[0045] In this example implementation, the multi-task matching model is a trained neural network based on the PLE (Progressive Layered Extraction) framework. By separating the shared part (Shared Experts) and the characteristic part of each task (Task-Specific Experts), deep information is extracted layer by layer, which is conducive to solving the negative transfer problem and the seesaw problem (the phenomenon that the performance of some tasks is improved and the performance of some tasks is reduced) in multi-task learning.

[0046] In this example implementation, the multi-task matching model may include at least a first subtask and a second subtask, wherein the first subtask may be used to extract the features of the characters of the address pair to be matched, and determine the corresponding matching result based on the distance information between the characters of the address pair to be matched. The second subtask may be used to extract the semantic features of the address pair to be matched, and determine the corresponding matching result based on the similarity between the semantic features of the address pair to be matched. The final matching result is determined based on the matching results of each subtask.

[0047] For example, reference Figure 2, the multi-task matching model may include a first subtask and a second subtask, the first subtask may include a feature extraction network composed of a first subtask expert and a first gating network and a first upper tower network, and the second subtask has the same structure as the first subtask, that is, the second subtask may include a feature extraction network composed of a second subtask expert and a second gating network and a second upper tower network. Address matching of source address information and target address information may be performed through the following steps.

[0048] The first subtask determines a first degree of match between each source address information and the target address information based on distance information.

[0049] In this example implementation, the input data (source address information and target address information) can be input into the first subtask, and the text distance features of the input address pair can be extracted through the feature extraction network corresponding to the first subtask (first subtask expert) to obtain a first feature vector corresponding to the source address information and a second feature vector corresponding to the target address information. The feature extraction network can include multiple expert layers and multiple shared layers, and the shared layer gradually extracts the feature representation formed by the feature layer. Each layer of the extraction network combines the shared experts and unique experts (i.e., subtask experts) to obtain the shared experts and unique experts of the next layer. Both shared experts and unique experts can be various deep neural networks, such as DNN. Feature screening / filtering is performed between different experts through a gating network (such as a first gating network and a second gating network).

[0050] Then, the extracted first feature vector and the second feature vector are processed through the upper tower network corresponding to the first subtask (such as the first upper tower network), the first distance information between the two is calculated, and the first matching degree is determined based on the first distance information.

[0051] In this example implementation, the first distance information can be determined by the minimum edit distance method or by using Euclidean distance, cosine similarity, etc., which is not limited in this example. The first matching degree is determined based on the first distance information to determine the matching probability between the two. For example, the minimum edit distance method is to find the shortest path for adding, deleting, modifying and checking through dynamic programming, that is, the minimum distance. For the source address of Tongzhou District, Beijing, and the target address of Maquqiao Town, Tongzhou District, Beijing, the edit distance of the address pair to be matched is 4, and the main editing operation is the addition operation. The corresponding matching probability is edit distance / length of the longer address 4 / 10=0.4.

[0052] For example, the output of the first subtask can be expressed as:

[0053] y 1 (x) = t 1 (g 1，N (x)

[0054] =t 1 (w 1，N g 1，N-1 (x))S 1，N (x)

[0055] In the above formula, t 1 represents the upper tower network of the first subtask, x represents the input data, that is, the vector expression of the source address information and the target address information, S 1，N (x) represents the selection matrix that connects shared experts and unique experts in series, w 1，N Corresponding to the softmax activation function of the gated network in the Nth extraction network, g 1，N (x) is the gating network corresponding to the Nth extraction network of the first subtask. The selection matrices of shared experts and subtask-specific experts are different, that is, shared experts are obtained by combining all experts in the previous layer structure, and unique experts are obtained by combining shared experts and corresponding unique experts in the previous layer structure.

[0056] The second subtask determines a first text similarity between each source address information and the target address information based on semantic information.

[0057] In this example implementation, the feature extraction network corresponding to the second subtask (second subtask expert) can be used to extract text semantic features of the source address information and the target address information to obtain the corresponding third feature vector and fourth feature vector; then, the upper-level tower network corresponding to the second subtask (second upper-level tower network) can be used to determine the similarity between the third feature vector and the fourth feature vector, and the cosine similarity is used as the first text similarity.

[0058] In this example implementation, the upper tower network corresponding to the second subtask can match short texts based on a pre-trained model, and can also match based on DSSM (Deep Structured Semantic Model), ESIM (Enhanced Sequential Inference Model), word2vec and other related sentence vector representation methods, which are not limited in this example. For example, Figure 3 As shown in the figure, short text matching based on the pre-training model is used as an example for explanation. The vectors to be matched (the third eigenvector and the fourth eigenvector) are respectively input into the Bert pre-training module (such as the first pre-training module and the second pre-training module) for vector representation, and then pooled through the pooling layer (such as the first pooling layer and the second pooling layer), and then the cosine similarity (vector similarity) of the pooled vectors is calculated, and finally classified through the softmax function (such as the classification layer) to determine the matching probability of the two.

[0059] The first matching degree and the corresponding first text similarity are fused to obtain the matching degree between each source address information and the target address information.

[0060] In this example implementation, the fusion process may be to perform various merging processes on the first matching degree and the corresponding first text similarity, such as product, weighted product, sum, weighted sum, etc., which is not limited in this example.

[0061] In response to the matching degree being greater than a preset threshold, it is determined that the source address information matches the target address information.

[0062] In this example implementation, the preset threshold may be set and adjusted based on experience and actual conditions. For example, the preset threshold may be set to 0.85 or 0.9.

[0063] In some embodiments, the multi-task matching model needs to be trained before it is used, such as Figure 4 As shown, the training process of the multi-task matching model may include the following steps.

[0064] Step S410, obtaining a training sample set.

[0065] In this example implementation, the corresponding source address information can be obtained from historical order data within a specified time period (such as the past six months or a month), and the corresponding target address information can be determined from the inventory record data in the business database. The source address information and the corresponding target address information form an address pair of a training sample. Each training sample can also include label information of each subtask, such as a first label corresponding to the first subtask and a second label corresponding to the second subtask.

[0066] In this example implementation, the first label is determined based on the second distance information between the address pairs of the training sample. For example, in response to the ratio of the second distance information between the address pairs of a training sample and the length of the longer character in the address pair being greater than a preset second threshold, it is determined that the first label indicates a successful match; the second threshold can be set according to actual conditions, such as the second threshold can be set to 0.8. For example, the first label can be defined as:

[0067] Label1(query,document)=1

[0068] if d(query,document) / max(len(query),len(document))>0.8else 0

[0069] In the above formula, query represents the source address, document represents the target address, d(query, document) represents the second distance between the source address and the target address, len(query) represents the character length of the source address, len(document) represents the character length of the target address, max represents the maximum value operation, and the value of the first label can be 0 or 1.

[0070] In this example implementation, the second label is determined based on the second text similarity between the address pairs of the training samples. For example, in response to the second text similarity between the address pairs of a training sample being greater than a preset third threshold, the second label is determined to indicate a successful match. The third threshold can be set according to actual conditions, such as the third threshold can be set to 0.9. For example, the second label can be defined as: Label2(query, document) = 1 if b(query, document) > 0.9 else 0

[0071] In the above formula, b(query, document) represents the second text similarity between the source address and the target address, and the value of the second label can be 0 or 1.

[0072] Step S420: input the training samples in the training sample set into the first subtask and the second subtask respectively, and obtain a first output corresponding to the first subtask and a second output corresponding to the second subtask.

[0073] In this example implementation, the first subtask and the second subtask both use the same training sample set, that is, the first subtask and the second subtask share a sample space. The model structures of the first subtask and the second subtask are as follows: Figure 2 As shown in the figure, the size of the hidden unit (i.e., output) of the last fully connected layer of the gating network is equal to the number of experts. The final output of the gating network is normalized. The number of gating networks in the second layer is the same as the number of subtasks, and the first layer has one more shared gating network than other layers.

[0074] Step S430: determine a first loss function according to the first output and the corresponding first label.

[0075] Step S440: determine a second loss function according to the second output and the corresponding second label.

[0076] Step S450: performing weighted summation on the first loss function and the second loss function to determine a target loss function.

[0077] In the above implementation, the first loss function and the second loss function can be determined by using a logarithmic loss function, or other types of loss functions, which are not limited in this example. The calculation formula of the target loss function can be:

[0078]

[0079] In the above formula, subscripts 1 and 2 represent the first subtask and the second subtask respectively, L 1 (θ 1 ,θ s ) represents the first loss function, L 2 (θ 2 ,θ s ) represents the second loss function; ω 1,0 represents the initial weight of the first subtask, represents the weight update factor of the tth training round of the first subtask, Indicates whether the i-th sample belongs to the sample space of the first subtask, and its value is [0,1]. The two subtasks of the present invention share the sample space, so This is equivalent to simplifying the loss function calculation formula and reducing the calculation complexity. represents the loss function of the i-th sample in the first subtask, θ 1 ,θ s They represent the hyperparameters of the first subtask and the shared parameters between the first and second subtasks, respectively. They respectively represent the estimated result and the true label (i.e., the first label) of the i-th sample in the first subtask.

[0080] Step S460: adjusting the model parameters of the multi-task matching model based on the target loss function.

[0081] In this example implementation, the model parameters can be adjusted using a stochastic gradient descent algorithm until the target loss function is small enough or the preset training rounds are reached, that is, the model training process is completed. In the above embodiment, the weights of the first subtask and the second subtask are learnable parameters, and the weights of the weighted sums in different training rounds are updated by the weight update factor.

[0082] Based on the above process, the first subtask and the second subtask are trained respectively through the training sample set to determine the model parameters of the multi-task matching model for risk control, and the trained multi-task matching model is used to evaluate and analyze the order data to be analyzed.

[0083] Step S130, in response to the number of source address information matched by the target address information exceeding a preset first threshold, determining that the target address information is a risky address, so as to perform risk control based on the risky address.

[0084] In this example implementation, when multiple source address information matches the same target address information, it indicates that the multiple source address information points to the same space, that is, the multiple source address information is essentially the same address. It can be considered that a large number of source address information points to the same target address within a specified period, indicating that there may be fraudulent behavior, and the target address information can be used as a risky address. Based on the determined risky address, risk control can be performed on the orders or credit behaviors of the target address, such as restricting transactions and purchase behaviors related to the address, so as to reduce losses caused by fraud.

[0085] For example, the implementation process of the risk control method disclosed in the present invention is as follows: Figure 5 As shown, the following steps may be included:

[0086] Step S501, obtaining a training sample set and constructing a multi-task matching model. The training samples in the training sample set include address pairs consisting of a source address and a target address, a first label corresponding to a first subtask and a second label corresponding to a second subtask, the first label is determined based on the second distance information between the address pairs of the training samples, and the second label is determined based on the second text similarity between the address pairs of the training samples.

[0087] Step S502: input the training samples in the training sample set into the first subtask and the second subtask respectively, and obtain a first output corresponding to the first subtask and a second output corresponding to the second subtask.

[0088] Step S503, determining a first loss function according to the first output and the corresponding first label; determining a second loss function according to the second output and the corresponding second label; performing weighted summation on the first loss function and the second loss function to determine a target loss function.

[0089] Step S504: adjusting the model parameters of the multi-task matching model based on the target loss function.

[0090] The above steps S502-S504 may be repeated until the training cutoff condition is reached to complete the model training process.

[0091] Step S505, obtaining multiple source address information of the target order and corresponding target address information of the standard structure; the source address information includes address information of different structures from multiple sources.

[0092] Step S506, performing multi-task address matching on each source address information and target address information through a multi-task matching model, and outputting the matching degree between the source address and the target address.

[0093] In this example, source address information that matches target address information is determined; the matching result of the multi-task matching model is determined based on the matching result of each subtask, and the multi-task matching model includes a first subtask and a second subtask, the first subtask is based on text distance matching, and the second subtask is based on text semantic matching.

[0094] Step S507, determine whether the matching degree is greater than a preset threshold, if so, proceed to step S508, otherwise, proceed to step S506 to match subsequent address pairs.

[0095] Step S508, determining whether the current source address matches the target address information.

[0096] Step S509, determining whether the number of source address information matched by a target address information exceeds a preset first threshold, if so, proceeding to step S510. Otherwise, proceeding to step S506 to continue matching.

[0097] Step S510: determine that the target address information is a risky address.

[0098] Step S511, performing risk control based on the risk address.

[0099] Experimental verification

[0100] According to the above specific implementation process, a specific embodiment is provided. In this embodiment, 100 data (which may be order-related data) are randomly extracted from the credit risk address data, and the disclosed method, semantic matching and minimum edit distance matching methods are respectively used. Under the same experimental conditions, the extracted data are predicted for risk addresses. The results are shown in Table 1.

[0101] Table 1 Prediction accuracy of risky addresses corresponding to different matching methods

[0102] index Minimum edit distance Semantic Matching The disclosed method Accuracy 0.77 0.78 0.83

[0103] From the results in Table 1, it can be seen that the prediction accuracy of the disclosed method is improved by 6% to 7% compared with the single-task matching algorithm used in the related art, which shows that the disclosed method can improve the address matching accuracy in the field of risk control.

[0104] refer to Figure 6 The present disclosure also provides a risk order determination method, which may include the following steps S610-S630.

[0105] Step S610, obtaining address information of each order to be predicted.

[0106] Step S620, matching an address information with a risk address in a risk address list through a risk prediction model to obtain a corresponding target matching degree; the risk address list is formed based on the risk addresses determined in the above-mentioned embodiment.

[0107] Step S630: In response to the target matching degree being greater than a preset fourth threshold, determining that the order corresponding to the address information is a risky order.

[0108] In this example implementation, the network structure of the risk prediction model is the same as the multi-task matching model, and the risk prediction model is obtained by training a training sample set consisting of a risk address and a risk address list, that is, the risk prediction model can also adopt a multi-task target model, but the model needs to be trained using different training sample sets consisting of risk addresses and risk address lists. In this example, a risk address list, i.e., a blacklist, can be formed by the risk addresses determined in the aforementioned risk control method, and the real-time order data of the application mall can be matched with the risk addresses in the blacklist. If the matching degree of the two is greater than the preset fourth threshold, the order is determined to be a risky order, i.e., the order may contain fraudulent behavior.

[0109] The present disclosure aims to solve the problem of low accuracy in the process of address consistency verification in the field of financial risk control by introducing the PLE framework, and constructs a multi-task matching model for identifying online fraud behaviors based on the framework. By combining mechanical character matching with semantic matching, two sub-tasks are formed, corresponding to address distance and semantic relevance respectively. This can simultaneously solve the problem that some semantics cannot be captured due to mechanical word-by-word matching, and the problem that the semantic vector has a high correlation but is actually inconsistent in spatial directionality. On the other hand, the present application designs labels for each sub-task to ensure the training effect of the multi-task matching model and the use effect of the model. In addition, the present disclosure reduces computational complexity and improves running speed and practicality by sharing sample space.

[0110] Furthermore, in this exemplary embodiment, a risk control device 700 is also provided. The risk control device 700 can be applied to a server. Figure 7As shown, the risk control device 700 may include: an acquisition module 710, used to acquire multiple source address information of a target order and corresponding target address information of a standard structure; the source address information includes address information of different structures from multiple sources; a matching module 720, used to perform multi-task address matching on each source address information and the target address information through a multi-task matching model, and determine the source address information that matches the target address information; the matching result of the multi-task matching model is determined based on the matching result of each subtask, and the multi-task matching model includes a first subtask and a second subtask, the first subtask is based on text distance matching, and the second subtask is based on text semantic matching; a control module 730, used to determine that the target address information is a risky address in response to the number of source address information matching the target address information exceeding a preset first threshold, so as to perform risk control based on the risk address.

[0111] In an exemplary embodiment of the present disclosure, the matching module 720 can also be used to: determine a first matching degree based on distance information between each source address information and the target address information through a first subtask; determine a first text similarity based on semantic information between each source address information and the target address information through a second subtask; fuse the first matching degree with the corresponding first text similarity to obtain a matching degree between each source address information and the target address information; in response to the matching degree being greater than a preset threshold, determine that the source address information matches the target address information.

[0112] In an exemplary embodiment of the present disclosure, the first subtask includes a corresponding feature extraction network and an upper tower network, the network structure of the second subtask is the same as that of the first subtask, and feature information is shared between the feature extraction networks of the first subtask and the second subtask; the matching module 720 can also be used to: extract text distance features of source address information and target address information through the feature extraction network corresponding to the first subtask to obtain corresponding first feature vectors and second feature vectors; determine first distance information between the first feature vector and the second feature vector through the upper tower network corresponding to the first subtask, and determine a first matching degree based on the first distance information; extract text semantic features of the source address information and the target address information through the feature extraction network corresponding to the second subtask to obtain corresponding third feature vectors and fourth feature vectors; determine the similarity between the third feature vector and the fourth feature vector through the upper tower network corresponding to the second subtask, and use the similarity as the first text similarity.

[0113] In an exemplary embodiment of the present disclosure, the device 700 may also include a training module, which can be used to: obtain a training sample set, the training samples in the training sample set include address pairs consisting of source addresses and target addresses, a first label corresponding to the first subtask and a second label corresponding to the second subtask, the first label is determined based on the second distance information between the address pairs of the training samples, and the second label is determined based on the second text similarity between the address pairs of the training samples; train the first subtask and the second subtask respectively through the training sample set to determine the model parameters of the multi-task matching model for risk control.

[0114] In an exemplary embodiment of the present disclosure, the training module can also be used to: input the training samples in the training sample set into the first subtask and the second subtask respectively, to obtain the first output corresponding to the first subtask and the second output corresponding to the second subtask; determine the first loss function according to the first output and the corresponding first label; determine the second loss function according to the second output and the corresponding second label; perform weighted summation of the first loss function and the second loss function to determine the target loss function; based on the target loss function, adjust the model parameters of the multi-task matching model; wherein the weights of the weighted summation in different training rounds are updated through the weight update factor.

[0115] In an exemplary embodiment of the present disclosure, the device also includes a label determination module, which can be used to: determine that the first label indicates a successful match in response to a ratio of a second distance information between an address pair of a training sample and a longer character length in the address pair being greater than a preset second threshold; determine that the second label indicates a successful match in response to a second text similarity between an address pair of a training sample being greater than a preset third threshold.

[0116] The specific details of each module or unit in the above-mentioned risk control device have been described in detail in the corresponding risk control method, so they will not be repeated here.

[0117] Furthermore, in this exemplary embodiment, a risk order determination device 800 is also provided. The risk order determination device 800 can be applied to a server. Figure 8As shown, the risk order determination device 800 may include: an acquisition module 810, used to obtain address information of each order to be predicted; a matching module 820, used to match an address information with a risk address in a risk address list through a risk prediction model to obtain a corresponding target matching degree; the risk address list is formed based on the risk addresses determined in the above-mentioned embodiment; the network structure of the risk prediction model is the same as the multi-task matching model, and the risk prediction model is obtained by training a training sample set consisting of a risk address and a risk address list; a determination module 830, used to determine that the order corresponding to the address information is a risk order in response to the target matching degree being greater than a preset fourth threshold.

[0118] The specific details of each module or unit in the above-mentioned risk order determination device have been described in detail in the corresponding risk order determination method, so they will not be repeated here.

[0119] As another aspect, the present application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the method in the following embodiment. For example, the electronic device may implement Figure 1 to Figure 5 The various steps shown, etc.

[0120] It should be noted that the computer-readable medium shown in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0121] Refer to the following Fig. 9 hereinafter describes an electronic device 900 according to such an embodiment of the present disclosure. Fig. 9 The electronic device 900 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0122] like Fig. 9 As shown, the electronic device 900 is in the form of a general computing device. The components of the electronic device 900 may include, but are not limited to: the at least one processing unit 910, the at least one storage unit 920, a bus 930 connecting different system components (including the storage unit 920 and the processing unit 910), and a display unit 940.

[0123] The storage unit stores program codes, which can be executed by the processing unit 910, so that the processing unit 910 executes the steps according to various exemplary embodiments of the present disclosure described in the above “Exemplary Method” section of this specification.

[0124] The storage unit 920 may include a readable medium in the form of a volatile storage unit, such as a random access storage unit (RAM) 9201 and / or a cache storage unit 9202 , and may further include a read-only storage unit (ROM) 9203 .

[0125] The storage unit 920 may also include a program / utility 9204 having a set (at least one) of program modules 9205, such program modules 9205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0126] Bus 930 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0127] The electronic device 900 may also communicate with one or more external devices 970 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may also communicate with one or more devices that enable a user to interact with the electronic device 900, and / or communicate with any device that enables the electronic device 900 to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via an input / output (I / O) interface 950. Furthermore, the electronic device 900 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter 960. As shown, the network adapter 960 communicates with other modules of the electronic device 900 via a bus 930. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 900, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RA identification systems, tape drives, and data backup storage systems.

[0128] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the embodiment of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the embodiment of the present disclosure.

[0129] In addition, the above-mentioned figures are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present disclosure, and are not intended to be limiting. It is easy to understand that the processes shown in the above-mentioned figures do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be performed synchronously or asynchronously, for example, in multiple modules.

[0130] It should be noted that, although the steps of the method in the present disclosure are described in a specific order in the drawings, this does not require or imply that the steps must be performed in the specific order, or that all the steps shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps, etc., which shall all be considered as part of the present disclosure.

[0131] It should be understood that the disclosure disclosed and defined in this specification extends to all alternative combinations of two or more individual features mentioned or evident in the text and / or the accompanying drawings. All these different combinations constitute multiple alternative aspects of the disclosure. The embodiments of this specification illustrate the best mode known for implementing the disclosure and will enable those skilled in the art to utilize the disclosure.

Claims

1. A risk control method, characterized in that: include: Obtain multiple source address information of the target order and corresponding target address information of the standard structure; The source address information includes address information of different structures from multiple sources; Performing multi-task address matching on each of the source address information and the target address information through a multi-task matching model to determine the source address information that matches the target address information; the matching result of the multi-task matching model is determined based on the matching result of each subtask, the multi-task matching model includes a first subtask and a second subtask, the first subtask is based on text distance matching, and the second subtask is based on text semantic matching; In response to the number of source address information matched by the target address information exceeding a preset first threshold, the target address information is determined to be a risky address, and risk control is performed based on the risky address.

2. The risk control method according to claim 1, characterized in that: The performing address matching on the source address information and the target address information includes: Determine, through the first subtask, a first degree of match between each of the source address information and the target address information based on distance information; Determine, by means of the second subtask, a first text similarity between each of the source address information and the target address information based on semantic information; Performing a fusion process on the first matching degree and the corresponding first text similarity to obtain a matching degree between each source address information and the target address information; In response to the matching degree being greater than a preset threshold, it is determined that the source address information matches the target address information.

3. The risk control method according to claim 2, characterized in that: The first subtask includes a corresponding feature extraction network and an upper tower network, the network structure of the second subtask is the same as that of the first subtask, and feature information is shared between the feature extraction networks of the first subtask and the second subtask; The determining the first distance information through the first subtask includes: Extracting text distance features of the source address information and the target address information through a feature extraction network corresponding to the first subtask to obtain corresponding first feature vectors and second feature vectors; Determining first distance information between the first feature vector and the second feature vector through an upper tower network corresponding to the first subtask, and determining a first matching degree based on the first distance information; The determining the first text similarity through the second subtask includes: Extracting text semantic features of the source address information and the target address information through a feature extraction network corresponding to the second subtask to obtain corresponding third feature vectors and fourth feature vectors; The similarity between the third feature vector and the fourth feature vector is determined by the upper tower network corresponding to the second subtask, and the similarity is used as the first text similarity.

4. The risk control method according to claim 1, characterized in that: The method further comprises: Acquire a training sample set, wherein the training samples in the training sample set include an address pair consisting of a source address and a target address, a first label corresponding to the first subtask, and a second label corresponding to the second subtask, wherein the first label is determined based on second distance information between the address pairs of the training samples, and the second label is determined based on second text similarity between the address pairs of the training samples; The first subtask and the second subtask are trained respectively by using the training sample set to determine model parameters of the multi-task matching model for risk control.

5. The risk control method according to claim 4, characterized in that: The training of the first subtask and the second subtask respectively by using the training sample set includes: Inputting the training samples in the training sample set into the first subtask and the second subtask respectively, obtaining a first output corresponding to the first subtask and a second output corresponding to the second subtask; Determine a first loss function according to the first output and the corresponding first label; Determine a second loss function according to the second output and the corresponding second label; Performing a weighted summation on the first loss function and the second loss function to determine a target loss function; Based on the objective loss function, adjusting model parameters of the multi-task matching model; The weights of the weighted sums in different training rounds are updated by weight update factors.

6. The risk control method according to claim 4, characterized in that: The method further comprises: In response to a ratio of second distance information between an address pair of a training sample and a longer character length in the address pair being greater than a preset second threshold, determining that the first label indicates a successful match; In response to a second text similarity between an address pair of a training sample being greater than a preset third threshold, it is determined that the second label indicates a successful match.

7. A method for determining a risk order, characterized in that: The method comprises: Obtain address information for each order to be predicted; The address information is matched with a risk address in the risk address list by a risk prediction model to obtain a corresponding target matching degree; the risk address list is formed based on the risk address determined in any one of claims 1 to 6; the network structure of the risk prediction model is the same as that of the multi-task matching model, and the risk prediction model is obtained by training a training sample set consisting of risk addresses and risk address lists; In response to the target matching degree being greater than a preset fourth threshold, it is determined that the order corresponding to the address information is a risky order.

8. A risk control device, characterized in that: The device comprises: An acquisition module, used to acquire multiple source address information of a target order and corresponding target address information of a standard structure; the source address information includes address information of different structures from multiple sources; A matching module, configured to perform multi-task address matching on each of the source address information and the target address information through a multi-task matching model, and determine the source address information that matches the target address information; the matching result of the multi-task matching model is determined based on the matching result of each subtask, the multi-task matching model includes a first subtask and a second subtask, the first subtask is based on text distance matching, and the second subtask is based on text semantic matching; The control module is used to determine that the target address information is a risky address in response to the number of source address information matched by the target address information exceeding a preset first threshold, so as to perform risk control based on the risky address.

9. A risk order determination device, characterized in that: The device comprises: An acquisition module is used to obtain the address information of each order to be predicted; A matching module, used for matching the address information with a risk address in a risk address list through a risk prediction model to obtain a corresponding target matching degree; the risk address list is formed based on the risk address determined in any one of claims 1 to 6; the network structure of the risk prediction model is the same as that of the multi-task matching model, and the risk prediction model is obtained by training a training sample set consisting of a risk address and a risk address list; A determination module is used to determine that the order corresponding to the address information is a risky order in response to the target matching degree being greater than a preset fourth threshold.

10. An electronic device, characterized in that: include: one or more processors; as well as A storage device, used to store one or more programs, when the one or more programs are executed by the one or more processors, enables the one or more processors to implement the method according to any one of claims 1 to 7.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images