Abnormal transaction detection methods, devices, computer equipment, and storage media

By using an outlier detection method based on historical transaction data in facial recognition devices, abnormal transaction devices and objects are identified, thus solving the problem of misjudgment in the detection of abnormal transaction operations in facial recognition devices and achieving highly accurate abnormal transaction identification.

CN116993342BActive Publication Date: 2026-11-17TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202210429294.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-22
Publication Date
2026-11-17
Estimated Expiration
2042-04-22

AI Technical Summary

Technical Problem

In existing technologies, the detection of abnormal transaction operations by facial recognition devices carries the risk of misjudgment, resulting in low accuracy of detection results.

Method used

By using historical transaction data, we determine the characteristic data of each transaction device and object in multiple dimensions, and use outlier detection methods to extract outlier dimensions from multiple dimensions. Then, we identify the set of abnormal transaction devices and risky objects from the transaction devices and objects, and finally identify abnormal transaction operations.

Benefits of technology

It improves the accuracy of detecting abnormal trading operations, reduces the risk of normal trading operations being misjudged as abnormal, and achieves fast and accurate identification of abnormal transactions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116993342B_ABST
    Figure CN116993342B_ABST
Patent Text Reader

Abstract

This application relates to a method, apparatus, computer device, storage medium, and computer program product for detecting abnormal transaction operations. The method includes: determining a first outlier dimension from the plurality of first dimensions based on device feature data of each transaction device under multiple first dimensions; acquiring target device feature data of each transaction device under the first outlier dimension, and determining an abnormal transaction device set from the plurality of transaction devices based on the target device feature data; determining a second outlier dimension from the plurality of second dimensions based on object feature data of each object under multiple second dimensions; acquiring target object feature data of each object under the second outlier dimension, and determining an abnormal risk object set from the plurality of objects based on the target object feature data; and determining abnormal transaction operations based on the abnormal transaction device set and the abnormal risk object set, thereby improving the accuracy of the detection results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a method, apparatus, computer equipment, storage medium, and computer program product for detecting abnormal transaction operations. Background Technology

[0002] With the development of internet technology, consumers have more diverse payment options. Currently, hardware devices supporting consumer payments include mobile phones, facial recognition devices, card readers, palm-sized readers, and computers. Taking facial recognition devices as an example, after purchasing these devices, service providers deploy them in stores. Consumers can choose to pay using facial recognition devices at checkout after selecting their purchases. Once payment is complete, the device generates corresponding transaction data. The device seller then compiles transaction data from all facial recognition devices and provides commissions to devices exceeding a certain transaction threshold within a certain period. If there is premeditated manipulation of facial recognition devices to generate abnormal transactions, these abnormal transactions are referred to as "brushing" (or "fake order" operations).

[0003] Traditional techniques determine whether a device experiences a sudden surge in transactions at a particular location, or whether a small number of users conduct a large number of transactions within a short period. If so, these transactions are considered fraudulent. However, this method carries the risk of misclassifying normal transactions as abnormal ones, resulting in low accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide an abnormal transaction operation detection method, device, computer equipment, storage medium, and computer program product that can improve the accuracy of detecting fraudulent transactions in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for detecting abnormal transaction operations. The method includes:

[0006] Based on historical transaction data, determine the device characteristic data of each transaction device under multiple first dimensions, and the object characteristic data of each object under multiple second dimensions;

[0007] Based on the device characteristic data of each trading device under multiple first dimensions, the first outlier dimension is determined from the multiple first dimensions.

[0008] Obtain the target device feature data of each trading device under the first outlier dimension, and determine the set of abnormal trading devices from multiple trading devices based on the target device feature data;

[0009] Based on the object feature data of each object under multiple second dimensions, the second outlier dimension is determined from the multiple second dimensions;

[0010] Obtain the target object feature data of each object under the second outlier dimension, and determine the set of abnormal risk objects from multiple objects based on the target object feature data;

[0011] Abnormal transaction operations are identified based on the set of abnormal transaction devices and the set of abnormal risk objects.

[0012] Secondly, this application also provides an abnormal transaction operation detection device. The device includes:

[0013] The first determining module is used to determine, based on historical transaction data, the device characteristic data of each transaction device under multiple first dimensions, and the object characteristic data of each object under multiple second dimensions.

[0014] The second determining module is used to determine a first outlier dimension from multiple first dimensions based on the device feature data of each transaction device under multiple first dimensions; and to determine a second outlier dimension from multiple second dimensions based on the object feature data of each object under multiple second dimensions.

[0015] The third determination module is used to obtain the target device feature data of each trading device under the first outlier dimension, and determine the abnormal trading device set from multiple trading devices based on the target device feature data; and to obtain the target object feature data of each object under the second outlier dimension, and determine the abnormal risk object set from multiple objects based on the target object feature data.

[0016] The fourth determination module is used to determine abnormal transaction operations based on the set of abnormal transaction devices and the set of abnormal risk objects.

[0017] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0018] Based on historical transaction data, determine the device characteristic data of each transaction device under multiple first dimensions, and the object characteristic data of each object under multiple second dimensions;

[0019] Based on the device characteristic data of each trading device under multiple first dimensions, the first outlier dimension is determined from the multiple first dimensions.

[0020] Obtain the target device feature data of each trading device under the first outlier dimension, and determine the set of abnormal trading devices from multiple trading devices based on the target device feature data;

[0021] Based on the object feature data of each object under multiple second dimensions, the second outlier dimension is determined from the multiple second dimensions;

[0022] Obtain the target object feature data of each object under the second outlier dimension, and determine the set of abnormal risk objects from multiple objects based on the target object feature data;

[0023] Abnormal transaction operations are identified based on the set of abnormal transaction devices and the set of abnormal risk objects.

[0024] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0025] Based on historical transaction data, determine the device characteristic data of each transaction device under multiple first dimensions, and the object characteristic data of each object under multiple second dimensions;

[0026] Based on the device characteristic data of each trading device under multiple first dimensions, the first outlier dimension is determined from the multiple first dimensions.

[0027] Obtain the target device feature data of each trading device under the first outlier dimension, and determine the set of abnormal trading devices from multiple trading devices based on the target device feature data;

[0028] Based on the object feature data of each object under multiple second dimensions, the second outlier dimension is determined from the multiple second dimensions;

[0029] Obtain the target object feature data of each object under the second outlier dimension, and determine the set of abnormal risk objects from multiple objects based on the target object feature data;

[0030] Abnormal transaction operations are identified based on the set of abnormal transaction devices and the set of abnormal risk objects.

[0031] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0032] Based on historical transaction data, determine the device characteristic data of each transaction device under multiple first dimensions, and the object characteristic data of each object under multiple second dimensions;

[0033] Based on the device characteristic data of each trading device under multiple first dimensions, the first outlier dimension is determined from the multiple first dimensions.

[0034] Obtain the target device feature data of each trading device under the first outlier dimension, and determine the set of abnormal trading devices from multiple trading devices based on the target device feature data;

[0035] Based on the object feature data of each object under multiple second dimensions, the second outlier dimension is determined from the multiple second dimensions;

[0036] Obtain the target object feature data of each object under the second outlier dimension, and determine the set of abnormal risk objects from multiple objects based on the target object feature data;

[0037] Abnormal transaction operations are identified based on the set of abnormal transaction devices and the set of abnormal risk objects.

[0038] The aforementioned abnormal transaction detection method, apparatus, computer equipment, storage medium, and computer program product, based on the device feature data of each trading device under multiple first dimensions, performs outlier detection, and can initially extract the first outlier dimension from the multiple first dimensions. Correspondingly, based on the object feature data of each object under multiple second dimensions, outlier detection can initially extract the second outlier dimension from the multiple second dimensions. Thus, for trading devices, the set of abnormal trading devices can be quickly and accurately determined from multiple trading devices based primarily on the target device feature data under the first outlier dimension. For trading objects, the set of abnormal risk objects can be quickly and accurately determined from multiple objects based primarily on the target object feature data under the second outlier dimension. In this way, transaction operations completed by abnormal risk objects on abnormal trading devices can be identified as abnormal transaction operations, reducing the risk of normal transaction operations being misjudged as abnormal transaction operations and greatly improving the accuracy of the detection results. Attached Figure Description

[0039] Figure 1 This is a framework diagram of an abnormal transaction operation detection system in one embodiment;

[0040] Figure 2 This is a flowchart illustrating an abnormal transaction operation detection method in one embodiment;

[0041] Figure 3 This is a flowchart illustrating the steps of determining a first outlier dimension from multiple first dimensions in one embodiment.

[0042] Figure 4 This is a schematic diagram of a normal distribution in one embodiment;

[0043] Figure 5 This is a flowchart illustrating the steps for determining a set of abnormal trading devices from multiple trading devices in one embodiment.

[0044] Figure 6 This is a flowchart illustrating the steps for determining a set of abnormal trading devices from multiple trading devices in another embodiment.

[0045] Figure 7 This is a schematic diagram of the structure of the first decision tree model in one embodiment;

[0046] Figure 8 This is a flowchart illustrating the steps for determining a second outlier dimension from multiple second dimensions in one embodiment.

[0047] Figure 9 This is a flowchart illustrating the steps for determining a set of abnormal risk objects from multiple objects in one embodiment.

[0048] Figure 10 This is a flowchart illustrating the steps for determining a set of abnormal risk objects from multiple objects in another embodiment;

[0049] Figure 11 This is a schematic diagram of the structure of the second decision tree model in one embodiment;

[0050] Figure 12 This is a structural block diagram of an abnormal transaction operation detection device in one embodiment;

[0051] Figure 13 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] Figure 1 This application provides an abnormal transaction operation detection system as an embodiment. Figure 1As shown, the abnormal transaction operation detection system includes multiple transaction devices 110 and a server 120. The multiple transaction devices 110 communicate with the server 120 via a network. A data storage system can store the data that the server 120 needs to process. The data storage system can be integrated onto the server 120 or placed in the cloud or on another server. In one possible scenario, the service provider can deploy multiple transaction devices 110 in stores. For example, one transaction device 110 can be deployed in store 1, two in store 2, and two in store 3. After a consumer selects goods at a store, they can choose any of the transaction devices 110 deployed in that store to complete the transaction at checkout. Once the transaction device 110 detects the completion of the transaction, it generates transaction data and sends it to the server 120. This transaction data may include: transaction time, transaction device, object, transaction amount, transaction duration, and the location of the transaction device. Server 120 can detect abnormal transaction operations based on the transaction data sent by transaction device 110. When calculating the transaction volume of each transaction device 110 in the future, abnormal transaction operations can be removed, thereby reducing the possibility of malicious order brushing.

[0054] The transaction device 110 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, facial recognition devices, card readers, palm readers, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 120 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0055] In one embodiment, such as Figure 2 As shown, a method for detecting abnormal transaction operations is provided. This method can be executed by a terminal or a server, and can be applied to... Figure 1 Taking server 120 as an example, the explanation includes the following steps:

[0056] S202. Based on historical transaction data, determine the device characteristic data of each transaction device under multiple first dimensions, and the object characteristic data of each object under multiple second dimensions.

[0057] Historical transaction data includes data generated by multiple objects performing transaction operations on different trading devices.

[0058] In one embodiment, as described above, each transaction device sends the transaction data to the server after generating transaction data each time. The server can use all received transaction data as historical transaction data, or, based on the transaction time point, extract transaction data generated by multiple historical transaction operations within a first time period from all received transaction data, and use the extracted transaction data as historical transaction data. The aforementioned first time period can be a preset historical time period, such as the past week, the past month, or the past year, etc., and this embodiment of the application does not limit this.

[0059] In one embodiment, multiple first dimensions may include transaction device identifier, transaction device location, transaction device model, transaction device origin, number of transactions per day per device, transaction amount per day per device, number of trading partners on the device, number of trading partners on the device sharing common social groups, and average transaction amount. Among these, transaction device identifier, transaction device location, transaction device model, and transaction device origin belong to the device attribute dimension; the number of transactions per day per device belongs to the transaction quantity dimension; the transaction amount per day per device belongs to the transaction resource dimension; the number of trading partners on the device and the number of trading partners on the device sharing common social groups belong to the trading partner dimension; and the average transaction amount belongs to the statistical dimension. The first dimensions can be flexibly set according to actual needs, and this embodiment does not limit them.

[0060] In one embodiment, after obtaining historical transaction data and multiple first dimensions, the server can count the transaction devices included in the historical transaction data, and for each transaction device, calculate the device characteristic data of that transaction device under multiple first dimensions.

[0061] For example, as shown in Table 1, assume that the first dimension includes: the number of transactions per device per day, the transaction amount per device per day, the number of transaction objects on the device, the number of transaction objects on the device that share a common social group, and the average transaction amount per device. Assume that the historical transaction data is the transaction data within the past month. The server counts the transaction devices included in the transaction data within the past month. Assume that the counted transaction devices include: abc1, abc2, abc3, abc4, and abc5. For each transaction device, the server extracts the transaction data for the past month from historical transaction data, counts the number of transactions for that device in the past month, divides this number by 30 to obtain the average number of transactions per day, and uses this average number of transactions per day as the device's characteristic data in the "number of transactions per day per device" dimension. The server also calculates the sum of the transaction amounts for the past month, divides this total amount by 30 to obtain the average daily amount, and uses this average daily amount as the device's characteristic data in the "number of transaction objects per device" dimension. Finally, the server counts the number of objects in the past month's transaction data for that device and uses this number as the device's characteristic data in the "number of transaction objects per device" dimension. The server then searches for the transaction data for the past month for that device. The system calculates the number of shared social groups among the transaction objects and uses this number as the device characteristic data for the dimension of "number of shared social groups among transaction objects on the device". For example, if the transaction device has 100 transaction objects in the past month, and two or more of these 100 objects are in the same social group, then that social group is considered a shared social group. Assuming that the count shows 10 shared social groups among these 100 objects, then 10 is used as the device characteristic data for the dimension of "number of shared social groups among transaction objects on the device". The total amount is then divided by the number of transactions in the past month to obtain the average transaction amount, which is used as the device characteristic data for the dimension of "average transaction amount per device". Following this method, the device characteristic data for abc1, abc2, abc3, abc4, and abc5 in multiple first dimensions can be obtained, as shown in Table 1.

[0062] It should be noted that: in Table 1, abc1, abc2, abc3, abc4, and abc5 are merely examples of unique device identifiers; other forms of unique device identifiers can also be used, and this application embodiment does not limit them. Similarly, in Table 1, the number of trading devices and multiple first dimensions are also examples, and this application embodiment does not limit them.

[0063] Table 1

[0064]

[0065]

[0066] In one embodiment, the object may be a user, a company, or a transaction account, etc., and this application embodiment does not limit this.

[0067] In one embodiment, multiple second dimensions may include: object identifier, object social group information, object's region, number of transactions per day, number of devices used for transactions per day, number of device locations, average transaction interval duration, and average transaction time. Among these, object identifier, object social group information, and object's region belong to object attribute dimensions; the number of transactions per day belongs to the transaction quantity dimension; the number of devices used for transactions per day and the number of device locations belong to the transaction device dimension; the average transaction interval duration belongs to the transaction interval dimension; and the average transaction time belongs to the statistical dimension. Similar to the first dimension, the second dimensions can be flexibly set according to actual needs, and this embodiment does not limit this.

[0068] In one embodiment, after obtaining historical transaction data and multiple second dimensions, the server can count the objects contained in the historical transaction data and, for each object, calculate the object feature data of that object under multiple second dimensions.

[0069] For example, as shown in Table 2, assume that multiple second dimensions include: the number of transactions per day for an object, the number of devices involved in transactions per day, the average transaction interval, the number of device locations, and the average transaction time for an object. Assume that the historical transaction data is from the past month. The server counts the objects included in the transaction data from the past month, assuming the counted objects include: 10001, 10002, 10003, 10004, and 10005. For each object, the server extracts the transaction data from the historical transaction data for the past month, counts the number of transactions for that object in the past month, divides this number by 30 to obtain the average number of transactions per day, and uses this average number of transactions per day as the object's characteristic data under the "number of transactions per day for an object" dimension; it also counts the total number of devices involved in the transactions for that object in the past month, divides this total number of devices by 30 to obtain the average number of devices involved in the transactions, and uses this average number of devices involved in the transactions per day as the object's characteristic data under the "number of devices involved in the transactions per day" dimension; finally, the server sorts the transaction data for the past month for that object according to the transaction time point. Based on the sorting results, the time interval between two adjacent transactions is calculated, and then the average time is taken as the object feature data under the dimension of "average transaction interval duration". The locations of the transaction devices included in the object's transaction data for the past month are counted, and the number of locations is taken as the object feature data under the dimension of "number of device locations". The sum of the transaction durations included in the object's transaction data for the past month is calculated, and the total transaction duration is divided by the number of transactions for the object in the past month to obtain the average duration. This average duration is taken as the object feature data under the dimension of "average transaction time". Following this method, object feature data for 10001, 10002, 10003, 10004, and 10005 in multiple second dimensions can be obtained, as shown in Table 2.

[0070] It should be noted that: in Table 1, 10001, 10002, 10003, 10004, and 10005 are merely examples of object unique identifiers; other forms of object unique identifiers can also be used, and this application embodiment does not limit them. Similarly, in Table 1, the number of objects and multiple second dimensions are also examples, and this application embodiment does not limit them.

[0071] Table 2

[0072]

[0073] S204. Based on the device characteristic data of each trading device under multiple first dimensions, determine the first outlier dimension from the multiple first dimensions.

[0074] In one embodiment, the server can acquire device feature data for each transaction device under each first dimension, then identify outlier data from this device feature data and count the number of outliers. Based on the number of outliers corresponding to each first dimension, a first outlier dimension is determined from multiple first dimensions.

[0075] S206. Obtain the target device feature data of each trading device under the first outlier dimension, and determine the set of abnormal trading devices from multiple trading devices based on the target device feature data.

[0076] Specifically, the target device feature data for each trading device under the first outlier dimension can be data found in the preset historical trading data under the first outlier dimension, or it can be obtained by filtering data under the first outlier dimension from historical trading data within a specified preset historical time period based on actual needs.

[0077] Furthermore, for each trading device, after obtaining the target device feature data of the trading device in the first outlier dimension, the server can determine whether the trading device is an abnormal trading device based on the target device feature data through statistical methods, or through a pre-established decision tree model. After obtaining the abnormal judgment results of each trading device, the server will take the set of abnormal trading devices among multiple trading devices as the abnormal trading device set.

[0078] Taking Table 1 as an example, assuming that the first outlier dimension determined from multiple first dimensions includes: the number of transactions per device per day, the number of transaction objects on the device, and the average transaction amount per device, the server obtains the device characteristic data of abc1, abc2, abc3, abc4, and abc5 under these dimensions, that is, the data in the first, third, and fifth columns of Table 1. Based on these three columns of data, the set of devices with abnormal transactions is determined from abc1, abc2, abc3, abc4, and abc5.

[0079] S208. Based on the object feature data of each object under multiple second dimensions, determine the second outlier dimension from the multiple second dimensions.

[0080] In one embodiment, the server can acquire object feature data for each object in each second dimension, then identify outlier data from this object feature data and count the number of outliers. Based on the number of outliers corresponding to each second dimension, a second outlier dimension is determined from multiple second dimensions.

[0081] S210. Obtain the target object feature data of each object under the second outlier dimension, and determine the set of abnormal risk objects from multiple objects based on the target object feature data.

[0082] Specifically, the target device feature data for each object under the second outlier dimension can be data found in the preset historical transaction data under the second outlier dimension, or it can be obtained by filtering the data under the second outlier dimension from historical transaction data within a specified preset historical time period based on actual needs.

[0083] Furthermore, for each object, after obtaining the target device feature data of the object in the second outlier dimension, the server can determine whether the object is an abnormal risk object based on the target device feature data through statistical methods, or through a pre-established decision tree model. After obtaining the abnormal judgment results of each object, the server uses the set of abnormal objects among multiple objects as the abnormal risk object set.

[0084] Taking Table 2 as an example, assume that the second outlier dimensions determined from multiple second dimensions include: the number of transactions per day, the number of devices used for transactions per day, and the number of device locations. The server obtains the object feature data for 10001, 10002, 10003, 10004, and 10005 under these dimensions, namely the data in columns 1, 2, and 4 of Table 2. Based on these three columns of data, an abnormal risk object set is determined from 10001, 10002, 10003, 10004, and 10005.

[0085] It should be noted that: S204-S206 can be executed before, after, or simultaneously with S208-S210. Figure 2 The execution order shown is merely an example, and the execution order of these two processes is not limited in the embodiments of this application.

[0086] S212. Based on the set of abnormal transaction devices and the set of abnormal risk objects, determine the abnormal transaction operation.

[0087] In one embodiment, as described above, each transaction device sends the transaction data to the server after generating transaction data each time. The server can extract transaction data for a second time period from all received transaction data. Based on the transaction data for the second time period, the set of abnormal transaction devices, and the set of abnormal risk objects, the server determines the abnormal transaction operation. The aforementioned second time period can be the same as the first time period, a portion of the first time period, or a time period that includes the first time period but is longer than the first time period. This embodiment of the application does not limit this.

[0088] The abnormal transaction operation detection method provided in this application performs outlier detection based on device feature data of each transaction device under multiple first dimensions, which can initially extract the first outlier dimension from the multiple first dimensions. Correspondingly, based on object feature data of each object under multiple second dimensions, outlier detection can initially extract the second outlier dimension from the multiple second dimensions. In this way, for transaction devices, the set of abnormal transaction devices can be quickly and accurately determined from multiple transaction devices based on the target device feature data under the first outlier dimension. For transaction objects, the set of abnormal risk objects can be quickly and accurately determined from multiple objects based on the target object feature data under the second outlier dimension. In this way, transaction operations completed by abnormal risk objects on abnormal transaction devices can be identified as abnormal transaction operations, reducing the risk of normal transaction operations being misjudged as abnormal transaction operations and greatly improving the accuracy of detection results.

[0089] In one embodiment, see Figure 3 As shown, based on the device characteristic data of each trading device under multiple first dimensions, the first outlier dimension is determined from the multiple first dimensions, including:

[0090] S301. Based on the device feature data of each trading device under multiple first dimensions, determine the distribution of device feature data under each first dimension, and determine the abnormal data under each first dimension based on the distribution.

[0091] For each first dimension, the server can determine the distribution in the following ways: the distribution can be obtained by calculating statistical parameters (such as mean, variance, median, mode, etc.) using statistical methods; it can also be obtained by plotting points and drawing the corresponding distribution map by establishing data axes or coordinate systems; or it can be obtained by drawing heat maps based on device feature data and obtaining the corresponding distribution based on the heat maps.

[0092] S302. Based on the number of outliers under each first dimension, select the first dimension from multiple first dimensions whose number of outliers meets the first preset condition, and use it as the first outlier dimension.

[0093] In one embodiment, after obtaining the number of outliers in each first dimension, the server can filter out the first dimension from multiple first dimensions whose number of outliers is greater than a preset value, and use it as the first outlier dimension. The preset value can be flexibly set according to the actual situation, and this embodiment does not limit it.

[0094] In one embodiment, after obtaining the number of outliers under each first dimension, the server can sort the multiple first dimensions in descending order of the number of outliers, and take the top N first dimensions as the first outlier dimensions. N can be flexibly set according to the actual situation, and this embodiment does not limit it.

[0095] In the above embodiments, the distribution of device feature data in the first dimension can reflect the degree of deviation of a certain device feature data from other device feature data. The abnormal data determined based on this distribution is more accurate. Furthermore, the first outlier dimension selected based on the number of abnormal data is also more accurate.

[0096] In one embodiment, based on the device feature data of each trading device under multiple first dimensions, the distribution of device feature data under each first dimension is determined, and abnormal data under each first dimension is determined based on the distribution. This includes: determining the mean and standard deviation corresponding to each first dimension based on the device feature data of each trading device under multiple first dimensions; determining the abnormal threshold corresponding to each first dimension based on the mean and standard deviation corresponding to each first dimension; and determining the abnormal data under each first dimension based on the device feature data under each first dimension and the abnormal threshold corresponding to each first dimension.

[0097] In one embodiment, see Figure 4 As shown, in a normal distribution, the probability of a value falling within (μ-σ, μ+σ) is 0.6526; the probability of a value falling within (μ-2σ, μ+2σ) is 0.9544; and the probability of a value falling within (μ-3σ, μ+3σ) is 0.9974, where σ represents the standard deviation and μ represents the mean. Since "low-probability events" typically refer to events with a probability of less than 5%, and from... Figure 4 It can be seen that the probability of the value distribution outside (μ-3σ, μ+3σ) is less than three per thousand. Therefore, μ-3σ and μ+3σ can be used as the threshold for judging low probability events.

[0098] Therefore, in this embodiment, the server can determine the mean and standard deviation corresponding to each first dimension based on the device feature data of each transaction device under multiple first dimensions. After obtaining the mean and standard deviation corresponding to each first dimension, μ-3σ and μ+3σ can be calculated for each first dimension. μ-3σ and μ+3σ are used as the anomaly thresholds for the first dimension. For each device feature data under the first dimension, it is determined whether the device feature data is within the interval (μ-3σ, μ+3σ). If it is, the device feature data is determined to be normal data; if it is not, the device feature data is determined to be abnormal data. Thus, the abnormal data under the first dimension can be obtained.

[0099] The following example illustrates this:

[0100] Referring to Table 1, which shows five first dimensions, taking the dimension of "number of transactions per device per day" as an example, the device characteristic data of each trading device in this dimension in Table 1 are 1000, 2000, 1500, 5000, and 2000. The server calculates the mean μ and standard deviation σ of these device characteristic data, and uses μ-3σ and μ+3σ as outlier thresholds. For each device characteristic data in 1000, 2000, 1500, 5000, and 2000, it is determined whether the device characteristic data is within the range of (μ-3σ, μ+3σ). If not, the device characteristic data is determined to be outlier data. Thus, the outlier data in the dimension of "number of transactions per device per day" can be obtained.

[0101] In the above embodiments, in statistical work, the mean and standard deviation are the two most important measures to describe the central tendency and dispersion of data. The outlier threshold determined based on the mean and standard deviation satisfies the basic assumption of low probability events and has high accuracy.

[0102] In one embodiment, there are multiple first outlier dimensions identified from multiple first dimensions, where "multiple" refers to two or more, see [link to relevant documentation]. Figure 5 As shown, based on the target device characteristic data, a set of abnormal trading devices is determined from multiple trading devices, including:

[0103] S501. Obtain the outlier threshold corresponding to each first outlier dimension.

[0104] In one embodiment, the server can determine the mean and standard deviation of each first dimension based on the device feature data of each trading device under multiple first dimensions; determine the outlier threshold of each first dimension based on the mean and standard deviation of each first dimension; determine the outlier data of each first dimension based on the device feature data of each first dimension and the outlier threshold of each first dimension; and select the first dimension whose number of outlier data meets a first preset condition from the multiple first dimensions as the first outlier dimension based on the number of outlier data in each first dimension. The outlier thresholds corresponding to each first dimension have already been obtained in the above process. After determining the first outlier dimension, the server searches for the outlier threshold corresponding to each first outlier dimension from the outlier thresholds corresponding to each first dimension.

[0105] S502. Based on the target device feature data under each first outlier dimension and the anomaly threshold corresponding to each first outlier dimension, determine the anomaly judgment result of each trading device under each first outlier dimension.

[0106] In one embodiment, for each first outlier dimension, the server can determine the anomaly range based on the anomaly threshold corresponding to the first outlier dimension, and can determine the anomaly judgment result of the trading device in each first outlier dimension based on the target device feature data of the trading device in each first outlier dimension and the anomaly range corresponding to each first outlier dimension.

[0107] In one embodiment, the outlier threshold corresponding to the first outlier dimension may include μ-3σ and μ+3σ, where μ is the mean of the device feature data under the first outlier dimension, σ is the standard deviation of the device feature data under the first outlier dimension, and the outlier range determined based on the outlier threshold may include (-∞, μ-3σ) and (μ+3σ, +∞).

[0108] In one embodiment, the server can obtain target device feature data for each trading device under multiple first outlier dimensions. For each trading device, the target device feature data of the trading device under each first outlier dimension is compared with the corresponding outlier range. If the target device feature data is within the outlier range, the trading device is determined to be outlier under the corresponding outlier dimension. Thus, the outlier judgment result of the trading device under each first outlier dimension can be obtained.

[0109] The following example illustrates this:

[0110] Taking Table 1 as an example, assuming the server identifies the following outlier dimensions from multiple first dimensions: daily transaction count per device, number of transaction objects per device, and average transaction amount per device, the outlier thresholds for the "daily transaction count per device" dimension include μ1-3σ1 and μ1+3σ1. The outlier thresholds for the "number of transaction objects per device" dimension include μ2-3σ2 and μ2+3σ2. The outlier thresholds for the "average transaction amount per device" dimension include μ3-3σ3 and μ3+3σ3. The server can consider (-∞, μ1-3σ1) and (μ1+3σ1, +∞) as the outlier range for the "daily transaction count per device" dimension. It can also consider (-∞, μ2-3σ2) and (μ2+3σ2, +∞) as the outlier range for the "number of transaction objects per device" dimension. Finally, it can consider (-∞, μ3-3σ3) and (μ3+3σ3, +∞) as the outlier range for the "average transaction amount per device" dimension. For each transaction device, taking abc2 in Table 1 as an example, the target device characteristic data for abc2 in the dimension of "number of transactions per day on the device" is 2000, the target device characteristic data in the dimension of "number of transaction objects on the device" is 300, and the target device characteristic data in the dimension of "average transaction amount per device" is 0.1. The server can determine whether 2000 is within the range of (-∞, μ1-3σ1) or (μ1+3σ1, +∞). If it is, then abc2 is determined to be abnormal in the dimension of "number of transactions per day on the device"; if it is not, then abc2 is determined to be normal in the dimension of "number of transactions per day on the device". The server can also determine whether 300 is within the range of (-∞, μ2-3σ2) or (μ2+3σ2, +∞). Within the specified range, if it falls within the range of "Number of Transaction Objects on Device," then abc2 is determined to be abnormal in the dimension of "Number of Transaction Objects on Device." If it does not fall within the range of "Number of Transaction Objects on Device," then abc2 is determined to be normal in the dimension of "Number of Transaction Objects on Device." Next, determine if 0.1 falls within the range of (-∞, μ3-3σ3) or (μ3+3σ3, +∞). If it does, then abc2 is determined to be abnormal in the dimension of "Average Transaction Amount on Device." If it does not fall within the range of "Average Transaction Amount on Device," then abc2 is determined to be normal in the dimension of "Average Transaction Amount on Device." Using the same method, the abnormality judgment results for abc1, abc3, abc4, and abc5 under "Daily Transaction Count on Device," "Number of Transaction Objects on Device," and "Average Transaction Amount on Device," respectively, can be obtained.

[0111] S503. Based on the anomaly judgment results of each trading device under each first outlier dimension, determine the anomaly judgment result of each trading device.

[0112] S504. Based on the anomaly judgment results of each trading device, determine the set of abnormal trading devices from multiple trading devices.

[0113] In one embodiment, after obtaining the anomaly judgment results of each trading device under each first outlier dimension, the server counts the number of anomaly results from the anomaly judgment results of each trading device under each first outlier dimension for each trading device, and then calculates the ratio of the number of anomaly results to the number of first outliers. If the ratio is greater than a preset threshold, the trading device is determined to be abnormal; if the ratio is less than or equal to the preset threshold, the trading device is determined to be normal. Using the same method, the anomaly judgment results of all trading devices can be obtained. Based on the anomaly judgment results of each trading device, a set of abnormal trading devices is obtained.

[0114] The following example illustrates this:

[0115] Using the example of S502, taking abc2 as an example, assuming the preset threshold is 0.8, after judgment, abc2 is abnormal in the dimension of "number of transactions per day on the device", abnormal in the dimension of "number of transaction objects on the device", and normal in the dimension of "average transaction amount per device". The number of abnormal results is 2, the number of the first outlier dimension is 3, and the ratio of the number of abnormal results to the number of the first outlier dimension is 2 / 3. Since 2 / 3 is less than 0.8, abc2 is determined to be abnormal. Using the same method, we can obtain the abnormal judgment results for abc1, abc3, abc4, and abc5. Assuming that the abnormal judgment result of abc1 is normal, the abnormal judgment result of abc2 is abnormal, the abnormal judgment result of abc3 is normal, the abnormal judgment result of abc4 is abnormal, and the abnormal judgment result of abc5 is normal, then the set consisting of abc2 and abc4 can be regarded as the abnormal transaction device set.

[0116] In the above embodiments, for each trading device, the anomaly judgment result of the trading device under each first outlier dimension is obtained, and the proportion of the anomaly results is counted. When the proportion exceeds a preset threshold, the trading device is determined to be an abnormal trading device. This method determines whether a trading device is an abnormal trading device from multiple first outlier dimensions, which is more accurate than the method in some embodiments that uses only one dimension to determine whether a trading device is an abnormal trading device.

[0117] In one embodiment, there are multiple first outlier dimensions identified from multiple first dimensions, where "multiple" refers to two or more, see [link to relevant documentation]. Figure 6 As shown, based on the target device characteristic data, a set of abnormal trading devices is determined from multiple trading devices, including:

[0118] S601. Obtain the pre-built first decision tree model, which includes the first root node, multiple first intermediate nodes, and multiple first leaf nodes.

[0119] S602. For each trading device, based on the target device feature data under each first outlier dimension corresponding to the corresponding trading device, starting from the first root node of the first decision tree model, continuously search from top to bottom in the first decision tree model for the first target intermediate node that matches the trading device, until the first target leaf node is reached.

[0120] S603. Based on the classification label corresponding to the first target leaf node of the first decision tree model, determine the anomaly judgment result corresponding to the corresponding transaction device.

[0121] The classification labels are used to indicate whether the trading equipment is normal or abnormal.

[0122] The first root node and the first intermediate node are used to perform branch judgments on the target device feature data of the trading device under each first outlier dimension, and the first leaf node is used to store classification labels. Specifically, a pre-constructed first decision tree model can be obtained, and then the target device feature data of the trading device under each first outlier dimension can be input into the first decision tree model. Starting from the first root node of the first decision tree model, the model is traversed, and the target device feature data of the trading device under each first outlier dimension is compared with the feature data of the corresponding node in the first decision tree model. Based on the comparison result, a subtree branch is selected, which is the first target intermediate node that is matched. The iteration continues until a leaf node is reached. It can be understood that the reached leaf node is the first target leaf node. Based on the classification label stored in the first target leaf node, the anomaly judgment result corresponding to the trading device is determined.

[0123] In one embodiment, for each trading device, based on the target device feature data under each first outlier dimension corresponding to the corresponding trading device, starting from the first root node of the first decision tree model, the system continuously searches for the first target intermediate node matching the trading device from top to bottom in the first decision tree model until the first target leaf node is reached. This includes: starting from the first root node of the first decision tree model, determining the first target intermediate node of the next layer based on the target device feature data under the first outlier dimension corresponding to the first root node and the feature data corresponding to the first root node; determining the first target intermediate node of the next layer based on the target device feature data under the first outlier dimension corresponding to the first target internal node of the next layer and the feature data corresponding to the first target intermediate node of the next layer, and continuously searching for the first target intermediate node matching the trading device until the first target leaf node is reached.

[0124] In one embodiment, the feature data corresponding to the first root node may include: the anomaly range corresponding to the first outlier dimension corresponding to the first root node, and the calculation of the anomaly range can refer to the foregoing embodiments. Similarly, the feature data corresponding to the first target intermediate node of the next layer may include: the anomaly range corresponding to the first outlier dimension corresponding to the first target intermediate node of the next layer.

[0125] In one embodiment, when classifying the target device feature data of a trading device under each first outlier dimension using a first decision tree model, the process starts from the first root node of the first decision tree model. Based on the target device feature data in the dimension corresponding to the first root node in each first outlier dimension, the next branch path is determined, which is to determine the first target intermediate node of the next layer. Then, based on the target device feature data in the dimension corresponding to the first target intermediate node of the next layer in each first outlier dimension, the subsequent branch paths are determined. This branching judgment is performed layer by layer until the first target leaf node is reached. Based on the classification label stored in the reached first target leaf node, the anomaly judgment result corresponding to the trading device is determined.

[0126] For example, see Figure 7 The first decision tree model includes a root node a, first intermediate nodes bd, and first leaf nodes P1-P5. The first leaf nodes P1, P2, P3, and P4 store the classification label "abnormal," while the first leaf node P5 stores the classification label "normal." The root node a corresponds to the dimension of "number of transaction objects on the device sharing a common social group." The first intermediate node b corresponds to the dimension of "average transaction amount." The first intermediate node c corresponds to the dimension of "number of transactions per day on the device." The first intermediate node d corresponds to the dimension of "number of transaction objects on the device." The feature data corresponding to the first root node a includes the abnormal range corresponding to the dimension of "number of transaction objects on the device sharing a common social group." Figure 7 (Used as (-∞, μ1-3σ1) and (μ1+3σ1, +∞)). See the feature data corresponding to the first intermediate node bd. Figure 7As shown, this will not be elaborated further. After the server obtains the target device feature data of the transaction device under each first outlier dimension, it obtains the target device feature data under the first outlier dimension corresponding to the first root node a. That is, it obtains the target device feature data of the transaction device under the dimension of "the number of common social groups of transaction objects on the device". It determines whether the target device feature data is in the range of (-∞, μ1-3σ1) or (μ1+3σ1, +∞). If it is, it reaches the first target leaf node P1. Since the classification label stored in the first target leaf node P1 is abnormal, the current transaction device is determined to be an abnormal transaction device. If the target device feature data is not in the range of (-∞, μ1-3σ1) and (μ1+3σ1, +∞), it reaches the first target intermediate node b. Similar processing is performed at the first target node b until the first target leaf node is reached. Based on the classification label stored in the first target leaf node, the abnormal judgment result corresponding to the transaction device is determined.

[0127] It should be noted that: Figure 7 This is just one example of the first decision tree model; the first decision tree model can also have other structures. Figure 7 The structure shown does not constitute a limitation on the embodiments of this application.

[0128] S604. Based on the anomaly judgment results of each trading device, determine the set of abnormal trading devices from multiple trading devices.

[0129] In one embodiment, after the server obtains the anomaly judgment results of each trading device, it filters out a set of abnormal trading devices based on the anomaly judgment results of each trading device.

[0130] The following example illustrates this:

[0131] Based on Table 1, assuming that the first decision tree model is used to perform branch judgment on the target device feature data of abc1 under multiple first outlier dimensions, and abc1 is determined to be normal, and abc3 and abc5 are determined to be normal, while abc2 and abc4 are determined to be abnormal, then the set consisting of abc2 and abc4 can be regarded as the set of abnormal transaction devices.

[0132] In the above embodiments, a pre-constructed first decision tree model is used to determine whether each transaction device is an abnormal transaction device. Since the first decision tree model includes a first root node, a first intermediate node, and a first leaf node, the first decision tree model makes judgments on the transaction device from multiple dimensions. Compared with some embodiments that use only one dimension to determine whether a transaction device is an abnormal transaction device, the result is more accurate.

[0133] In one embodiment, the construction steps of the first decision tree model include: obtaining a first sample set and a first classification label corresponding to each device sample, wherein the first sample set includes device feature sample data of each device sample under each first outlier dimension; determining the information gain corresponding to each first outlier dimension based on the device feature sample data of each device sample under each first outlier dimension and the first classification label of each device sample; selecting a target first outlier dimension as a classification feature from multiple first outlier dimensions based on the information gain corresponding to each first outlier dimension to create a first root node; creating a first intermediate node and a first leaf node based on the device feature sample data under the first root node and the first classification label corresponding to each device feature sample data; and determining the first decision tree model based on the created first root node, the first intermediate node under the first root node, and the first leaf node.

[0134] In one embodiment, determining the information gain corresponding to each first outlier dimension based on the device feature sample data of each device sample under each first outlier dimension and the first classification label of each device sample includes: statistically analyzing the device feature sample data under each first outlier dimension based on the device feature sample data of each device sample under each first outlier dimension; for each first outlier dimension, determining a first proportion of device feature sample data with the first classification label being normal and a second proportion of device feature sample data with the first classification label being abnormal based on the first and second proportions; determining the information entropy of the first sample set based on the first and second proportions; and determining the information gain corresponding to each first outlier dimension based on the information entropy of the first sample set.

[0135] In one embodiment, determining the information entropy of the first sample set based on a first ratio and a second ratio includes:

[0136] The information entropy of the first sample set is determined using the following formula:

[0137]

[0138] Where |y| represents the type of the first category label. In this embodiment, the first category label has two types: normal and abnormal. Therefore, |y| = 2, Ent(D) = -p1log2p1 - p2log2p2, where p1 and p2 are the first ratio and the second ratio, respectively.

[0139] In one embodiment, based on the information entropy of the first sample set, determining the information gain corresponding to each first outlier dimension includes:

[0140] For each first outlier dimension, based on the device feature sample data under the first outlier dimension, determine the category corresponding to each device feature sample data, obtain the number of device feature sample data under each category and the information entropy corresponding to each category, and determine the information gain corresponding to the first outlier dimension based on the information entropy of the first sample set, the number of device feature sample data under each category, and the information entropy corresponding to each category. For example, based on all device feature sample data under the first outlier dimension, calculate the mean and variance of all device feature sample data, and based on the mean and variance, divide all device feature sample data under the first outlier dimension into two categories, where device feature sample data belonging to the range (μ-3σ, μ+3σ) is one category, and device feature sample data not belonging to the range (μ-3σ, μ+3σ) is another category.

[0141] The process of obtaining the information entropy corresponding to each category is similar to the process of obtaining the information entropy of the first sample set in the previous embodiment, and will not be described again here.

[0142] In one embodiment, the information gain corresponding to the first outlier dimension is determined based on the information entropy of the first sample set, the number of device feature sample data under each category, and the information entropy corresponding to each category, including:

[0143] The information gain corresponding to the first outlier dimension is determined using the following formula:

[0144]

[0145] Where V represents the number of categories corresponding to each device feature sample data under the first outlier dimension, Ent(D) represents the information entropy of the first sample set, and D v Ent(D) represents the number of device feature sample data in each category. v ) represents the information entropy corresponding to each category, and D represents the number of device samples.

[0146] In one embodiment, based on the information gain corresponding to each first outlier dimension, a target first outlier dimension is selected from multiple first outlier dimensions as a classification feature to create the first root node, including:

[0147] The first outlier dimension with the largest information gain among multiple first outlier dimensions is taken as the target first outlier dimension. The target first outlier dimension is used as the classification feature to create the first root node. In other words, the first root node corresponds to the target first outlier dimension.

[0148] In one embodiment, creating a first intermediate node and a first leaf node based on the device feature sample data under the first root node and the first classification label corresponding to each device feature sample data includes: classifying the device feature sample data under the first root node according to the values ​​of the device feature sample data under the first root node; for each category, obtaining the device sample corresponding to the category, and obtaining the device feature sample data of the device sample corresponding to the category under other first outlier dimensions besides the target first outlier dimension; obtaining the first classification label corresponding to the device sample corresponding to the category from the first classification label corresponding to each device feature sample data; calculating the information gain of each first outlier dimension in other first outlier dimensions based on the device feature sample data of the device sample corresponding to the category in other first outlier dimensions and the first classification label corresponding to the device sample corresponding to the category; determining the intermediate node (first intermediate node) corresponding to the category based on the information gain of each first outlier dimension in other first outlier dimensions, and so on, until the first leaf node is obtained. The process of calculating the information gain of each of the other first outlier dimensions, and the process of determining intermediate nodes based on the information gain of each of the other first outlier dimensions, can be found in the process of creating the first root node in the above embodiments. The method of classifying the device feature sample data under the first root node can be found in the method of dividing all device feature sample data under the first outlier dimension into two categories in the aforementioned embodiments. It will not be elaborated further here.

[0149] Since a greater information gain means a greater "purity improvement" when using the first outlier dimension for partitioning, in the above embodiment, based on screening multiple first outlier dimensions from multiple first dimensions, the information gain of each first outlier dimension is calculated, and a first decision tree model is constructed based on the information gain, resulting in a first decision tree model with high classification accuracy.

[0150] In one embodiment, see Figure 8 As shown, based on the object feature data of each object under multiple second dimensions, the second outlier dimension is determined from the multiple second dimensions, including:

[0151] S801. Based on the object feature data of each object in multiple second dimensions, determine the distribution of object feature data in each second dimension, and determine the abnormal data in each second dimension based on the distribution.

[0152] For each second dimension, the server can determine the distribution in the following ways: the distribution can be obtained by calculating statistical parameters (such as mean, variance, median, mode, etc.) using statistical methods; it can also be obtained by plotting points and drawing the corresponding distribution map by establishing data axes or coordinate systems; or it can be obtained by drawing heat maps based on device feature data and obtaining the corresponding distribution based on the heat maps.

[0153] S802. Based on the number of outliers in each second dimension, select the second dimension from multiple second dimensions whose number of outliers meets the second preset condition, and use it as the second outlier dimension.

[0154] In one embodiment, after obtaining the number of outliers in each second dimension, the server can filter out the second dimension from multiple second dimensions whose number of outliers is greater than a preset value, and use it as the first outlier dimension. The preset value can be flexibly set according to the actual situation, and this embodiment does not limit it.

[0155] In one embodiment, after obtaining the number of outliers in each second dimension, the server can sort the multiple second dimensions in descending order of the number of outliers, and take the top N second dimensions as the second outlier dimensions. N can be flexibly set according to the actual situation, and this embodiment does not limit it.

[0156] In the above embodiments, the distribution of object feature data in the second dimension can reflect the degree of deviation of a certain object feature data from other object feature data. The abnormal data determined based on this distribution is more accurate. Furthermore, the second outlier dimension selected based on the number of abnormal data is also more accurate.

[0157] In one embodiment, based on the object feature data of each object under multiple second dimensions, the distribution of object feature data under each second dimension is determined, and abnormal data under each second dimension is determined based on the distribution. This includes: determining the mean and standard deviation corresponding to each second dimension based on the object feature data of each object under multiple second dimensions; determining the abnormal threshold corresponding to each second dimension based on the mean and standard deviation corresponding to each second dimension; and determining the abnormal data under each second dimension based on the object feature data under each second dimension and the abnormal threshold corresponding to each second dimension.

[0158] In one embodiment, the server can determine the mean and standard deviation for each second dimension based on device feature data of each object under multiple second dimensions. After obtaining the mean and standard deviation for each second dimension, μ-3σ and μ+3σ can be calculated for each second dimension. μ-3σ and μ+3σ are used as the outlier thresholds for that second dimension. For each object feature data under that second dimension, it is determined whether the object feature data is within the interval (μ-3σ, μ+3σ). If it is, the object feature data is determined to be normal data; if it is not, the object feature data is determined to be outlier data. Thus, the outlier data under that second dimension can be obtained.

[0159] The following example illustrates this:

[0160] Referring to Table 2, which shows five second dimensions, taking the dimension of "number of transactions per day for an object" as an example, the object feature data for each object in this dimension in Table 2 are 100, 200, 150, 10, and 5. The server calculates the mean μ and standard deviation σ of these object feature data, and uses μ-3σ and μ+3σ as outlier thresholds. For each object feature data in 100, 200, 150, 10, and 5, it is determined whether the object feature data is within the range of (μ-3σ, μ+3σ). If not, the object feature data is determined to be outlier data. Thus, the outlier data in the dimension of "number of transactions per day for an object" can be obtained.

[0161] In the above embodiments, in statistical work, the mean and standard deviation are the two most important measures to describe the central tendency and dispersion of data. The outlier threshold determined based on the mean and standard deviation satisfies the basic assumption of low probability events and has high accuracy.

[0162] In one embodiment, there are multiple second outlier dimensions identified from a plurality of second dimensions, where "multiple" refers to two or more, see [link to relevant documentation]. Figure 9 As shown, based on the target object's characteristic data, a set of abnormal risk objects is determined from multiple objects, including:

[0163] S901. Obtain the outlier threshold corresponding to each second outlier dimension.

[0164] In one embodiment, the server can determine the mean and standard deviation of each second dimension based on the object feature data of each object under multiple second dimensions; determine the outlier threshold of each second dimension based on the mean and standard deviation of each second dimension; determine the outlier data of each second dimension based on the object feature data of each second dimension and the outlier threshold of each second dimension; and select the second dimension from the multiple second dimensions whose number of outlier data meets a second preset condition as the second outlier dimension based on the number of outlier data in each second dimension. The outlier thresholds corresponding to each second dimension have already been obtained in the above process. After determining the second outlier dimension, the server searches for the outlier threshold corresponding to each second outlier dimension from the outlier thresholds corresponding to each second dimension.

[0165] S902. Based on the target object feature data under each second outlier dimension and the outlier threshold corresponding to each second outlier dimension, determine the outlier judgment result of each object under each second outlier dimension.

[0166] In one embodiment, for each second outlier dimension, the server can determine the outlier range based on the outlier threshold corresponding to the second outlier dimension, and can determine the outlier judgment result of the object in each second outlier dimension based on the target object feature data of the object in each second outlier dimension and the outlier range corresponding to each second outlier dimension.

[0167] In one embodiment, the outlier threshold corresponding to the second outlier dimension may include: μ-3σ and μ+3σ, where μ is the mean of the device feature data under the second outlier dimension, σ is the standard deviation of the device feature data under the second outlier dimension, and the outlier range determined based on the outlier threshold may include: (-∞, μ-3σ) and (μ+3σ, +∞).

[0168] In one embodiment, the server can obtain target object feature data for each object under multiple second outlier dimensions. For each object, the target object feature data of the object under each second outlier dimension is compared with the anomaly range corresponding to the corresponding dimension. If the target object feature data is within the anomaly range, the object is determined to be anomaly under the corresponding outlier dimension. Thus, the anomaly judgment result of the object under each first outlier dimension can be obtained.

[0169] The following example illustrates this:

[0170] Taking Table 2 as an example, assuming the server identifies the following outlier dimensions from multiple second dimensions: daily transaction count, daily transaction device count, and number of device locations, the outlier thresholds for the "daily transaction count" dimension are μ1-3σ1 and μ1+3σ1. The outlier thresholds for the "daily transaction device count" dimension are μ2-3σ2 and μ2+3σ2. The outlier thresholds for the "number of device locations" dimension are μ3-3σ3 and μ3+3σ3. The server can consider (-∞, μ1-3σ1) and (μ1+3σ1, +∞) as the outlier range for the "daily transaction count" dimension. It can also consider (-∞, μ2-3σ2) and (μ2+3σ2, +∞) as the outlier range for the "daily transaction device count" dimension. Finally, it can consider (-∞, μ3-3σ3) and (μ3+3σ3, +∞) as the outlier range for the "number of device locations" dimension. For each object, taking 10002 in Table 2 as an example, the target object feature data for 10002 under the dimension of "Daily Transaction Count" is 200, the target object feature data under the dimension of "Daily Transaction Device Count" is 20, and the target object feature data under the dimension of "Number of Device Locations" is 1. The server can determine whether 200 is within the range of (-∞, μ1-3σ1) or (μ1+3σ1, +∞). If it is, then 10002 is determined to be abnormal under the dimension of "Daily Transaction Count"; if it is not, then 10002 is determined to be normal under the dimension of "Daily Transaction Count". It can also determine whether 20 is within the range of (-∞, μ2-3σ2) or (μ2+3σ2, +∞). If it is, then 10002 is determined to be normal. If 1 is an anomaly in the "Daily Transaction Device Quantity" dimension, and not in the "Daily Transaction Device Quantity" dimension, then 10002 is considered normal in this dimension. Next, check if 1 is within the range of (-∞, μ3-3σ3) or (μ3+3σ3, +∞). If it is, then 10002 is an anomaly in the "Number of Device Locations" dimension; otherwise, it is considered normal. Using the same method, we can obtain the anomaly detection results for 10001, 10003, 10004, and 10005 in the "Daily Transaction Count," "Daily Transaction Device Quantity," and "Number of Device Locations" dimensions, respectively.

[0171] S903. Determine the anomaly judgment result for each object based on the anomaly judgment results for each object under each second outlier dimension.

[0172] S904. Based on the anomaly judgment results of each object, determine the set of anomaly risk objects from multiple objects.

[0173] In one embodiment, after obtaining the anomaly judgment results for each object under each second outlier dimension, the server counts the number of anomaly results for each object from the anomaly judgment results under each second outlier dimension, and then calculates the ratio of the number of anomaly results to the number of the second outlier dimension. If the ratio is greater than a preset threshold, the object is determined to be abnormal; if the ratio is less than or equal to the preset threshold, the object is determined to be normal. Using the same method, the anomaly judgment results for all objects can be obtained. Based on the anomaly judgment results of each object, a set of objects with anomaly risk is obtained.

[0174] The following example illustrates this:

[0175] Using S902 as an example, let's take 10002 as an example. Assuming the preset threshold is 0.8, after judgment, 10002 is abnormal in the dimension of "number of transactions per day" and "number of devices used for transactions per day," but normal in the dimension of "number of device locations." The number of abnormal results is 2, and the number of the second outlier dimension is 3. The ratio of the number of abnormal results to the number of the second outlier dimension is 2 / 3. Since 2 / 3 is less than 0.8, 10002 is determined to be abnormal. Using the same method, we can obtain the anomaly judgment results for 10001, 10003, 10004, and 10005. Assuming the anomaly judgment result for 10001 is normal, the anomaly judgment result for 10002 is abnormal, the anomaly judgment result for 10003 is normal, the anomaly judgment result for 10004 is abnormal, and the anomaly judgment result for 10005 is normal, then the set consisting of 10002 and 10004 can be regarded as the set of abnormal risk objects.

[0176] In the above embodiments, for each object, the anomaly judgment result of the object under each second outlier dimension is obtained, and the proportion of the outlier results is counted. When the proportion exceeds a preset threshold, the object is determined to be an anomalous risk object. This method determines whether an object is an anomalous risk object from multiple second outlier dimensions. Compared with some embodiments that use only one dimension to determine whether an object is an anomalous risk object, the results are more accurate.

[0177] In one embodiment, there are multiple second outlier dimensions identified from a plurality of second dimensions, where "multiple" refers to two or more, see [link to relevant documentation]. Figure 10 As shown, based on the target object's characteristic data, a set of abnormal risk objects is determined from multiple objects, including:

[0178] S1001. Obtain a pre-built second decision tree model, which includes a second root node, multiple second intermediate nodes, and multiple second leaf nodes.

[0179] S1002. For each object, based on the target device feature data under each second outlier dimension corresponding to the object, starting from the second root node of the second decision tree model, continuously search from top to bottom for the second target intermediate node that matches the object in the second decision tree model until the second target leaf node is reached.

[0180] S1003. Based on the classification label corresponding to the second target leaf node of the second decision tree model, determine the anomaly judgment result corresponding to the corresponding object.

[0181] The category label is used to indicate whether an object is normal or abnormal.

[0182] The second root node and the second intermediate node are used to branch and determine the anomaly of the object based on its feature data in each of the second outlier dimensions. The second leaf node stores the classification label. Specifically, a pre-constructed second decision tree model is obtained, and the feature data of the object in each of the second outlier dimensions is input into the second decision tree model. The model is traversed starting from the second root node. The feature data of the object in each of the second outlier dimensions is compared with the feature data of the corresponding node in the second decision tree model. Based on the comparison result, a subtree branch is selected, which is the matching second target intermediate node. This iteration continues until a leaf node is reached. The reached leaf node is the second target leaf node. Based on the classification label stored in this second target leaf node, the anomaly judgment result for the object is determined.

[0183] In one embodiment, for each object, based on the target object feature data under each second outlier dimension corresponding to the corresponding object, starting from the second root node of the second decision tree model, the system continuously searches for the second target intermediate node matching the object from top to bottom in the second decision tree model until the second target leaf node is reached. This includes: starting from the second root node of the second decision tree model, determining the second target intermediate node of the next layer based on the target object feature data under the second outlier dimension corresponding to the second root node and the feature data corresponding to the second root node; determining the second target intermediate node of the next layer based on the target object feature data under the second outlier dimension corresponding to the second target intermediate node of the next layer and the feature data corresponding to the second target intermediate node of the next layer, and continuously searching for the second target intermediate node matching the object until the second target leaf node is reached.

[0184] In one embodiment, the feature data corresponding to the second root node may include: the anomaly range corresponding to the second outlier dimension corresponding to the second root node, and the calculation of the anomaly range can refer to the aforementioned embodiment. Similarly, the feature data corresponding to the second target intermediate node in the next layer may include: the anomaly range corresponding to the second outlier dimension corresponding to the second target intermediate node in the next layer.

[0185] In one embodiment, when classifying target object feature data under each second outlier dimension using a second decision tree model, the process starts from the second root node of the second decision tree model. Based on the target object feature data in the dimension corresponding to the second root node in each second outlier dimension, the next branch path is determined, i.e., the intermediate node of the second target at the next level is determined. Then, based on the target object feature data in the dimension corresponding to the intermediate node of the second target at the next level in each second outlier dimension, subsequent branch paths are determined. This branching process is repeated layer by layer until the leaf node of the second target is reached. Based on the classification label stored in the reached leaf node of the second target, the anomaly judgment result corresponding to the object is determined.

[0186] For example, see Figure 11 The second decision tree model includes a second root node a1, second intermediate nodes b1-d1, and second leaf nodes L1-L5. Second leaf nodes L1, L2, L3, and L4 store the classification label "abnormal," while second leaf node L5 stores the classification label "normal." The second root node a corresponds to the dimension of "number of device locations," the second intermediate node b corresponds to the dimension of "average transaction interval," the second intermediate node c corresponds to the dimension of "number of transactions per day," and the second intermediate node d corresponds to the dimension of "number of devices transacted per day." The feature data corresponding to the second root node a includes the abnormal range corresponding to the dimension of "number of device locations" (…). Figure 11 (Used as (-∞, μ1-3σ1) and (μ1+3σ1, +∞)). See the feature data corresponding to the second intermediate node bd. Figure 11As shown, this will not be elaborated further. After obtaining the target object feature data of the object in each of the second outlier dimensions, the target object feature data of the second outlier dimension corresponding to the second root node a is obtained, that is, the target object feature data of the object in the dimension of "number of device locations" is obtained. It is determined whether the target object feature data is in the range of (-∞, μ1-3σ1) or (μ1+3σ1, +∞). If it is, the second target leaf node L1 is reached. Since the classification label stored in the second target leaf node L1 is anomaly, the current object is determined to be an anomalous risk object. If the target object feature data is not in the range of (-∞, μ1-3σ1) and (μ1+3σ1, +∞), the second target intermediate node b1 is reached. Similar processing is performed in the second target node b1 until the second target leaf node is reached. Based on the classification label stored in the second target leaf node, the anomaly judgment result corresponding to the object is determined.

[0187] It should be noted that: Figure 11 This is just one example of a second decision tree model; other structures are also possible. Figure 11 The structure shown does not constitute a limitation on the embodiments of this application.

[0188] S1004. Based on the anomaly judgment results of each object, determine the set of anomaly risk objects from multiple objects.

[0189] In one embodiment, after the server obtains the anomaly assessment results for each object, it filters out a set of objects with anomaly risks based on the anomaly assessment results for each object.

[0190] The following example illustrates this:

[0191] Based on Table 2, assuming that the second decision tree model is used to perform branch judgment on the target object feature data of 10001 under multiple second outlier dimensions, and 10001 is determined to be normal, and in the same way, 10003 and 10005 are determined to be normal, and 10002 and 10004 are determined to be abnormal, then the set consisting of 10002 and 10004 can be regarded as the set of abnormal risk objects.

[0192] In the above embodiments, a pre-constructed second decision tree model is used to determine whether each transaction device is an abnormal risk object. Since the second decision tree model includes a second root node, a second intermediate node, and a second leaf node, the second decision tree model makes judgments on the object from multiple dimensions. Compared with some embodiments that use only one dimension to determine whether an object is an abnormal risk object, the result is more accurate.

[0193] In one embodiment, the construction steps of the second decision tree model include: obtaining a second sample set and a second classification label corresponding to each object sample, wherein the second sample set includes object feature sample data of each device sample under each second outlier dimension; determining the information gain corresponding to each second outlier dimension based on the object feature sample data of each object sample under each second outlier dimension and the second classification label of each object sample; selecting a target second outlier dimension as a classification feature from multiple second outlier dimensions based on the information gain corresponding to each second outlier dimension to create a second root node; creating a second intermediate node and a second leaf node based on the object feature sample data under the second root node and the second classification label corresponding to each object feature sample data; and determining the second decision tree model based on the created second root node, the second intermediate node under the second root node, and the second leaf node.

[0194] In one embodiment, determining the information gain corresponding to each second outlier dimension based on the object feature sample data of each object sample under each second outlier dimension and the second classification label of each object sample includes: statistically analyzing the object feature sample data under each second outlier dimension based on the object feature sample data of each object sample under each second outlier dimension; for each second outlier dimension, determining a third proportion of object feature sample data with the second classification label being normal and a fourth proportion of object feature sample data with the second classification label being abnormal based on the second classification label of each object feature sample data under the second outlier dimension; determining the information entropy of the second sample set based on the third and fourth proportions; and determining the information gain corresponding to each second outlier dimension based on the information entropy of the second sample set.

[0195] In one embodiment, determining the information entropy of the second sample set based on a third ratio and a fourth ratio includes:

[0196] The information entropy of the second sample set is determined using the following formula:

[0197]

[0198] Where |y| represents the type of the second category label. In this embodiment, there are two types of category labels: normal and abnormal. Therefore, |y| = 2, Ent(D) = -p1log2p1 - p2log2p2, where p1 and p2 are the third and fourth proportions, respectively.

[0199] In one embodiment, based on the information entropy of the second sample set, the information gain corresponding to each second outlier dimension is determined, including:

[0200] For each second outlier dimension, based on the object feature sample data under the second outlier dimension, determine the category corresponding to each object feature sample data, obtain the number of object feature sample data under each category and the information entropy corresponding to each category, and determine the information gain corresponding to the second outlier dimension based on the information entropy of the second sample set, the number of object feature sample data under each category, and the information entropy corresponding to each category. For example, based on all object feature sample data under the second outlier dimension, calculate the mean and variance of all object feature sample data, and based on the mean and variance, divide all object feature sample data under the second outlier dimension into two categories, where object feature sample data belonging to the range (μ-3σ, μ+3σ) is one category, and object feature sample data not belonging to the range (μ-3σ, μ+3σ) is the other category.

[0201] The process of obtaining the information entropy corresponding to each category is similar to the process of obtaining the information entropy of the second sample set in the previous embodiment, and will not be described again here.

[0202] In one embodiment, the information gain corresponding to the second outlier dimension is determined based on the information entropy of the second sample set, the number of object feature sample data under each category, and the information entropy corresponding to each category, including:

[0203] The information gain corresponding to the second outlier dimension is determined using the following formula:

[0204]

[0205] Where V represents the number of categories corresponding to each object feature sample data under the second outlier dimension, Ent(D) represents the information entropy of the second sample set, and D v Ent(D) represents the number of object feature sample data in each category. v ) represents the information entropy corresponding to each category, and D represents the number of object samples.

[0206] In one embodiment, based on the information gain corresponding to each second outlier dimension, a target second outlier dimension is selected from multiple second outlier dimensions as a classification feature to create a second root node, including:

[0207] The second outlier dimension with the largest information gain among multiple second outlier dimensions is taken as the target second outlier dimension. The target second outlier dimension is used as the classification feature to create the second root node. In other words, the root node corresponds to the target second outlier dimension.

[0208] In one embodiment, creating a second intermediate node and a second leaf node based on the object feature sample data under the second root node and the second classification label corresponding to each object feature sample data includes: classifying the object feature sample data under the second root node according to the values ​​of the object feature sample data under the second root node; for each category, obtaining the object sample corresponding to that category, and obtaining the object feature sample data of the object sample corresponding to that category under other second outlier dimensions besides the target second outlier dimension; obtaining the second classification label corresponding to the object sample corresponding to that category from each object feature sample data; calculating the information gain of each second outlier dimension in other second outlier dimensions based on the object feature sample data of the object sample corresponding to that category in other second outlier dimensions and the second classification label corresponding to the object sample corresponding to that category; determining the intermediate node (second intermediate node) corresponding to that category based on the information gain of each second outlier dimension in other second outlier dimensions, and so on, until the second leaf node is obtained. The process of calculating the information gain of each second outlier dimension in other second outlier dimensions and the process of determining the intermediate node based on the information gain of each second outlier dimension in other second outlier dimensions can be found in the process of creating the second root node in the above embodiment. The method for classifying the object feature sample data under the second root node can be found in the aforementioned embodiment, which divides all object feature sample data under the second outlier dimension into two categories. This will not be elaborated upon further here.

[0209] Since a greater information gain means a greater "purity improvement" when using the second outlier dimension for partitioning, in the above embodiment, based on the selection of multiple second outlier dimensions from multiple second dimensions, the information gain of each second outlier dimension is calculated, and a second decision tree model is constructed based on the information gain, resulting in a second decision tree model with high classification accuracy.

[0210] In one embodiment, historical transaction data refers to transaction data generated by multiple historical transaction operations within a preset historical time period. Based on the set of abnormal transaction devices and the set of abnormal risk objects, abnormal transaction operations are determined, including: determining the target transaction device and target object associated with each historical transaction operation that generated the historical transaction data; and determining that any historical transaction operation is an abnormal transaction operation if the target transaction device associated with any historical transaction operation is in the set of abnormal transaction devices and the associated target object is in the set of abnormal risk objects.

[0211] In one embodiment, after obtaining the set of abnormal transaction devices and the set of abnormal risk objects, the server, for any transaction data in the historical transaction data, obtains the transaction device and object contained in the transaction data, takes the transaction device as the target transaction device associated with the transaction operation that generated the transaction data, takes the object as the target object associated with the transaction operation that generated the transaction data, and determines whether the target transaction device is in the set of abnormal transaction devices and whether the target object is in the set of abnormal risk objects. If the target transaction device is in the set of abnormal transaction devices and the target object is in the set of abnormal risk objects, then the transaction operation that generated the transaction data is determined to be an abnormal transaction operation.

[0212] The following example illustrates this:

[0213] Assume the set of abnormal transaction devices includes abc2 and abc4 in Table 1; and the set of abnormal risk objects includes 10002 and 10004 in Table 2. If a transaction in a certain transaction data is recorded as transaction device abc2 and object 10002, since abc2 is in the set of abnormal transaction devices and 10002 is in the set of abnormal risk objects, then the transaction operation that generated this transaction data can be determined as an abnormal transaction operation.

[0214] In the above embodiments, the transaction operation is judged from two dimensions: transaction device and object. Only when the transaction device associated with the transaction operation is in the abnormal transaction device set and the object is in the abnormal risk object set, is the transaction operation classified as an abnormal transaction operation. Compared with some embodiments, where transaction operations occur frequently on a certain transaction device and these transactions are directly regarded as abnormal transaction operations, the solution of this application embodiment has higher accuracy.

[0215] In one embodiment, an abnormal transaction operation detection method is provided, applied to a server, including:

[0216] Based on historical transaction data, device characteristic data for each transaction device under multiple first dimensions and object characteristic data for each object under multiple second dimensions are determined.

[0217] Based on the device characteristic data of each trading device under multiple first dimensions, determine the mean and standard deviation corresponding to each first dimension; based on the mean and standard deviation corresponding to each first dimension, determine the anomaly threshold corresponding to each first dimension; based on the device characteristic data under each first dimension and the anomaly threshold corresponding to each first dimension, determine the abnormal data under each first dimension.

[0218] Based on the number of outliers in each first dimension, the first dimension whose number of outliers meets the first preset condition is selected from multiple first dimensions and used as the first outlier dimension.

[0219] Obtain the anomaly threshold corresponding to each first outlier dimension; based on the target device feature data under each first outlier dimension and the anomaly threshold corresponding to each first outlier dimension, determine the anomaly judgment result of each trading device under each first outlier dimension; based on the anomaly judgment result of each trading device under each first outlier dimension, determine the anomaly judgment result of each trading device; based on the anomaly judgment result of each trading device, determine the set of abnormal trading devices from multiple trading devices.

[0220] or,

[0221] Obtain a first sample set and a first classification label corresponding to each device sample. The first sample set includes device feature sample data of each device sample under each first outlier dimension. Based on the device feature sample data of each device sample under each first outlier dimension and the first classification label of each device sample, determine the information gain corresponding to each first outlier dimension. Based on the information gain corresponding to each first outlier dimension, select a target first outlier dimension as a classification feature from multiple first outlier dimensions to create a first root node. Based on the device feature sample data under the first root node and the first classification label corresponding to each device feature sample data, create a first intermediate node and a first leaf node. Based on the created first root node, the first intermediate node under the first node, and the first leaf node, determine the first decision tree model. Obtain a pre-constructed first decision tree model, which includes a first root node, multiple first intermediate nodes, and multiple first leaf nodes. For each trading device, based on the target device feature data under each first outlier dimension corresponding to the corresponding trading device, starting from the first root node of the first decision tree model, continuously search from top to bottom for the first target intermediate node that matches the trading device until the first target leaf node is reached. Based on the classification label corresponding to the first target leaf node of the first decision tree model, determine the anomaly judgment result corresponding to the corresponding trading device. Based on the anomaly judgment results of each trading device, determine the set of anomalous trading devices from multiple trading devices.

[0222] Based on the object feature data of each object under multiple second dimensions, determine the mean and standard deviation corresponding to each second dimension; based on the mean and standard deviation corresponding to each second dimension, determine the outlier threshold corresponding to each second dimension; based on the object feature data under each second dimension and the outlier threshold corresponding to each second dimension, determine the outlier data under each second dimension.

[0223] Based on the number of outliers in each second dimension, the second dimension whose number of outliers meets the second preset condition is selected from multiple second dimensions and used as the second outlier dimension.

[0224] Obtain the anomaly threshold corresponding to each second outlier dimension; based on the target object feature data under each second outlier dimension and the anomaly threshold corresponding to each second outlier dimension, determine the anomaly judgment result of each object under each second outlier dimension; based on the anomaly judgment result of each object under each second outlier dimension, determine the anomaly judgment result of each object; based on the anomaly judgment result of each object, determine the set of anomaly risk objects from multiple objects.

[0225] or,

[0226] Obtain a second sample set and a second classification label corresponding to each object sample. The second sample set includes object feature sample data of each device sample under each second outlier dimension. Based on the object feature sample data of each object sample under each second outlier dimension and the second classification label of each object sample, determine the information gain corresponding to each second outlier dimension. Based on the information gain corresponding to each second outlier dimension, select a target second outlier dimension as a classification feature from multiple second outlier dimensions to create a second root node. Based on the object feature sample data under the second root node and the second classification label corresponding to each object feature sample data, create a second intermediate node and a second leaf node. Based on the created second root node, the second intermediate node under the second root node, and the second leaf node, determine the second decision tree model. Obtain a pre-constructed second decision tree model, which includes a second root node, multiple second intermediate nodes, and multiple second leaf nodes. For each object, based on the target device feature data under each second outlier dimension corresponding to the object, starting from the second root node of the second decision tree model, continuously search from top to bottom for the second target intermediate node that matches the object until the second target leaf node is reached. Determine the anomaly judgment result corresponding to the object based on the classification label corresponding to the second target leaf node of the second decision tree model. Based on the anomaly judgment result of each object, determine the set of abnormal risk objects from multiple objects.

[0227] Identify the target trading devices and target objects associated with each historical trading operation that generated the historical trading data; if the target trading device associated with any historical trading operation is in the abnormal trading device set and the associated target object is in the abnormal risk object set, then determine any historical trading operation as an abnormal trading operation.

[0228] In the above embodiments, outlier detection is performed based on the device feature data of each trading device under multiple first dimensions, which can initially extract the first outlier dimension from the multiple first dimensions. Correspondingly, outlier detection is performed based on the object feature data of each object under multiple second dimensions, which can initially extract the second outlier dimension from the multiple second dimensions. In this way, for trading devices, the set of abnormal trading devices can be quickly and accurately determined from multiple trading devices based on the target device feature data under the first outlier dimension. For trading objects, the set of abnormal risk objects can be quickly and accurately determined from multiple objects based on the target object feature data under the second outlier dimension. Thus, trading operations completed by abnormal risk objects on abnormal trading devices can be identified as abnormal trading operations, reducing the risk of normal trading operations being misjudged as abnormal trading operations and greatly improving the accuracy of the detection results.

[0229] In one possible scenario, the transaction device is a facial recognition device. Service providers can deploy multiple facial recognition devices in different stores. After selecting goods at a store, consumers can choose any facial recognition device in that store to complete the transaction at checkout. Once the facial recognition device detects the transaction, it generates transaction data and sends it to the server. The server can retrieve historical transaction data from the received data and, based on this, determine the device characteristics of each facial recognition device across multiple first dimensions, and the object characteristics of each object across multiple second dimensions. The multiple first dimensions include: the number of transactions per device per day, the daily transaction amount per device, the number of transacting objects on the device, the number of shared social groups among transacting objects on the device, and the average transaction amount per device. The multiple second dimensions include: the number of transactions per object per day, the number of facial recognition devices per day, the number of device locations, the average transaction interval, and the average transaction time per object. The server determines the first outlier dimension from the device feature data of each facial recognition device across multiple first dimensions, obtains the target device feature data of each facial recognition device under the first outlier dimension, and identifies an abnormal set of facial recognition devices based on the target device feature data. Similarly, based on the object feature data of each object across multiple second dimensions, the server determines the second outlier dimension, obtains the target object feature data of each object under the second outlier dimension, and identifies an abnormal risk object set from the multiple objects. Based on the abnormal set of facial recognition devices and the abnormal risk object set, abnormal transaction operations are identified. This reduces the risk of normal transactions on facial recognition devices being misjudged as abnormal transactions, significantly improving the accuracy of the detection results.

[0230] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0231] Based on the same inventive concept, this application also provides an abnormal transaction operation detection device for implementing the abnormal transaction operation detection method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the abnormal transaction operation detection device provided below can be found in the limitations of the abnormal transaction operation detection method described above, and will not be repeated here.

[0232] In one embodiment, such as Figure 12 As shown, an abnormal transaction operation detection device 12 is provided. This device can be a software module, a hardware module, or a combination of both as part of a computer device. Specifically, the device includes:

[0233] The first determining module 121 is used to determine, based on historical transaction data, the device characteristic data of each transaction device under multiple first dimensions, and the object characteristic data of each object under multiple second dimensions.

[0234] The second determining module 122 is used to determine a first outlier dimension from multiple first dimensions based on the device feature data of each transaction device under multiple first dimensions; and to determine a second outlier dimension from multiple second dimensions based on the object feature data of each object under multiple second dimensions.

[0235] The third determination module 123 is used to obtain the target device feature data of each trading device under the first outlier dimension, and determine the abnormal trading device set from multiple trading devices based on the target device feature data; and to obtain the target object feature data of each object under the second outlier dimension, and determine the abnormal risk object set from multiple objects based on the target object feature data.

[0236] The fourth determination module 124 is used to determine abnormal transaction operations based on the abnormal transaction device set and the abnormal risk object set.

[0237] In one embodiment, the second determining module 122 is specifically used to determine the distribution of device feature data in each first dimension based on the device feature data of each transaction device in multiple first dimensions, and to determine the abnormal data in each first dimension based on the distribution; and to select the first dimension from multiple first dimensions whose number of abnormal data meets the first preset condition based on the number of abnormal data in each first dimension, as the first outlier dimension.

[0238] In one embodiment, the second determining module 122 is specifically used to determine the mean and standard deviation corresponding to each first dimension based on the device feature data of each trading device under multiple first dimensions; determine the abnormal threshold corresponding to each first dimension based on the mean and standard deviation corresponding to each first dimension; and determine the abnormal data under each first dimension based on the device feature data under each first dimension and the abnormal threshold corresponding to each first dimension.

[0239] In one embodiment, there are multiple first outlier dimensions. The third determining module 123 is specifically used to obtain the anomaly threshold corresponding to each first outlier dimension; determine the anomaly judgment result of each trading device under each first outlier dimension based on the target device feature data under each first outlier dimension and the anomaly threshold corresponding to each first outlier dimension; determine the anomaly judgment result of each trading device based on the anomaly judgment result of each trading device under each first outlier dimension; and determine the set of abnormal trading devices from multiple trading devices based on the anomaly judgment result of each trading device.

[0240] In one embodiment, there are multiple first outlier dimensions. The third determining module 123 is specifically used to obtain a pre-constructed first decision tree model, which includes a root node, multiple intermediate nodes, and multiple leaf nodes. For each trading device, based on the target device feature data under each first outlier dimension corresponding to the corresponding trading device, starting from the root node of the first decision tree model, the target intermediate node matching the trading device is continuously searched from top to bottom in the first decision tree model until the target leaf node is reached. Based on the classification label corresponding to the target leaf node of the first decision tree model, the anomaly judgment result corresponding to the corresponding trading device is determined. Based on the anomaly judgment result of each trading device, the set of abnormal trading devices is determined from multiple trading devices.

[0241] In one embodiment, the third determining module 123 is further configured to obtain a first sample set and classification labels corresponding to each device sample, wherein the first sample set includes device feature sample data of each device sample under each first outlier dimension; determine the information gain corresponding to each first outlier dimension based on the device feature sample data of each device sample under each first outlier dimension and the classification labels of each device sample; select a target first outlier dimension as a classification feature from multiple first outlier dimensions to create a root node based on the information gain corresponding to each first outlier dimension; split the first sample set into multiple training subsets based on the selected classification feature; continuously select classification features to create intermediate nodes in the training subsets generated by the splitting; and perform data splitting based on the selected classification features to generate new training subsets until each training subset obtained by the final splitting is classified onto the corresponding classification label; create corresponding leaf nodes based on each classification label; and determine a first decision tree model based on the created root node, intermediate nodes under the root node, and leaf nodes.

[0242] In one embodiment, the second determining module 122 is specifically used to determine the distribution of object feature data in each second dimension based on object feature data of each object in multiple second dimensions, and to determine the outlier data in each second dimension based on the distribution; and to select the second dimension from multiple second dimensions whose number of outlier data satisfies the second preset condition based on the number of outlier data in each second dimension, as the second outlier dimension.

[0243] In one embodiment, the second determining module 122 is specifically used to determine the mean and standard deviation corresponding to each second dimension based on the object feature data of each object under multiple second dimensions; determine the anomaly threshold corresponding to each second dimension based on the mean and standard deviation corresponding to each second dimension; and determine the abnormal data under each second dimension based on the object feature data under each second dimension and the anomaly threshold corresponding to each second dimension.

[0244] In one embodiment, there are multiple second outlier dimensions. The third determining module 123 is specifically used to obtain the anomaly threshold corresponding to each second outlier dimension; determine the anomaly judgment result of each object under each second outlier dimension based on the target object feature data under each second outlier dimension and the anomaly threshold corresponding to each second outlier dimension; determine the anomaly judgment result of each object based on the anomaly judgment result of each object under each second outlier dimension; and determine the set of anomaly risk objects from multiple objects based on the anomaly judgment result of each object.

[0245] In one embodiment, there are multiple second outlier dimensions. The third determining module 123 is specifically used to obtain a pre-constructed second decision tree model, which includes a root node, multiple intermediate nodes, and multiple leaf nodes. For each object, based on the target device feature data under each second outlier dimension corresponding to the object, starting from the root node of the second decision tree model, the target intermediate node matching the object is continuously searched from top to bottom in the second decision tree model until the target leaf node is reached. Based on the classification label corresponding to the target leaf node of the second decision tree model, the anomaly judgment result corresponding to the corresponding object is determined. Based on the anomaly judgment result of each object, an abnormal risk object set is determined from multiple objects.

[0246] In one embodiment, the third determining module 123 is further configured to obtain a second sample set and classification labels corresponding to each object sample, wherein the second sample set includes object feature sample data of each device sample under each second outlier dimension; determine the information gain corresponding to each second outlier dimension based on the object feature sample data of each object sample under each second outlier dimension and the classification labels of each object sample; select a target second outlier dimension as a classification feature from multiple second outlier dimensions based on the information gain corresponding to each second outlier dimension to create a root node, and split the second sample set into multiple training subsets according to the selected classification feature; continuously select classification features to create intermediate nodes in the training subsets generated by the splitting, and perform data splitting according to the selected classification features to generate new training subsets, until each training subset obtained by the final splitting is classified onto the corresponding classification label; create corresponding leaf nodes according to each classification label; and determine the second decision tree model based on the created root node, intermediate nodes under the root node, and leaf nodes.

[0247] In one embodiment, the fourth determining module 124 is specifically used to determine the target trading device and target object associated with each historical trading operation that generates historical trading data; if the target trading device associated with any historical trading operation is in the abnormal trading device set and the associated target object is in the abnormal risk object set, then any historical trading operation is determined to be an abnormal trading operation.

[0248] Each module in the aforementioned abnormal transaction operation detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0249] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 13 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores historical transaction data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When executed by the processor, the computer program implements a method for detecting abnormal transaction operations.

[0250] Those skilled in the art will understand that Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0251] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0252] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0253] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0254] It should be noted that the object information (including but not limited to object identifiers, object social group information, object's region, object's daily transaction count, object's daily transaction device count, number of locations of devices used by the object, average transaction interval duration, average transaction time, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data, etc.) involved in this application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0255] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0256] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0257] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting abnormal transaction operations, characterized in that, The method includes: Based on historical transaction data, device feature data of each transaction device under multiple first dimensions and object feature data of each object under multiple second dimensions are determined. The historical transaction data includes data generated by multiple objects performing transaction operations on different transaction devices. The transaction devices include at least one of the following: smartphones, computers, facial recognition devices, card swiping devices, palm swiping devices, IoT devices, and portable wearable devices installed in stores. The objects include at least one of users, enterprises, or transaction accounts. The historical transaction data includes transaction data generated by multiple historical transaction operations within a preset historical time period. Based on the device characteristic data of each transaction device under multiple first dimensions, a first outlier dimension is determined from the multiple first dimensions. The multiple first dimensions include at least one of the following: transaction device identifier, transaction device location, transaction device model, transaction device source, number of transactions per day per device, transaction amount per day per device, number of transaction objects on the device, number of transaction objects on the device that share a common social group, and average transaction amount. Obtain the target device feature data of each trading device under the first outlier dimension, and determine the set of abnormal trading devices from multiple trading devices based on the target device feature data; Based on the object feature data of each object under multiple second dimensions, a second outlier dimension is determined from the multiple second dimensions. The multiple second dimensions include at least one of the following: object identifier, object social group information, object region, number of transactions per day, number of devices used for transactions per day, number of device locations, average transaction interval duration, and average transaction time of the object. Obtain the target object feature data of each object under the second outlier dimension, and determine the set of abnormal risk objects from multiple objects based on the target object feature data; Based on the set of abnormal transaction devices and the set of abnormal risk objects, abnormal transaction operations are determined, including: determining the target transaction device and target object associated with each historical transaction operation that generated the historical transaction data; if the target transaction device associated with any historical transaction operation is in the set of abnormal transaction devices and the associated target object is in the set of abnormal risk objects, the any historical transaction operation is determined to be an abnormal transaction operation, and the abnormal transaction operation includes order brushing operations.

2. The method according to claim 1, characterized in that, The step of determining the first outlier dimension from the multiple first dimensions based on the device characteristic data of each transaction device under multiple first dimensions includes: Based on the device feature data of each trading device under multiple first dimensions, determine the distribution of device feature data under each first dimension, and determine the abnormal data under each first dimension based on the distribution. Based on the number of outliers in each first dimension, the first dimension whose number of outliers meets the first preset condition is selected from the multiple first dimensions and used as the first outlier dimension.

3. The method according to claim 2, characterized in that, The step of determining the distribution of device feature data under each first dimension based on device feature data of each transaction device under multiple first dimensions, and determining abnormal data under each first dimension based on the distribution, includes: Based on the device characteristic data of each trading device under multiple first dimensions, determine the mean and standard deviation corresponding to each first dimension; Based on the mean and standard deviation of each of the first dimensions, determine the anomaly threshold corresponding to each of the first dimensions; Based on the device feature data under each first dimension and the corresponding anomaly threshold for each first dimension, the abnormal data under each first dimension is determined.

4. The method according to claim 1, characterized in that, The first outlier dimension has multiple aspects. The step of determining the set of abnormal trading devices from multiple trading devices based on the target device feature data includes: Obtain the outlier threshold for each first outlier dimension; Based on the target device feature data under each first outlier dimension and the anomaly threshold corresponding to each first outlier dimension, determine the anomaly judgment result of each trading device under each first outlier dimension. Based on the anomaly judgment results of each trading device under each first outlier dimension, determine the anomaly judgment result of each trading device; Based on the anomaly judgment results of each of the aforementioned trading devices, a set of abnormal trading devices is determined from multiple trading devices.

5. The method according to claim 1, characterized in that, The first outlier dimension has multiple aspects. The step of determining the set of abnormal trading devices from multiple trading devices based on the target device feature data includes: Obtain a pre-constructed first decision tree model, which includes a first root node, multiple first intermediate nodes, and multiple first leaf nodes; For each trading device, based on the target device feature data under each first outlier dimension corresponding to the corresponding trading device, starting from the first root node of the first decision tree model, continuously search from top to bottom in the first decision tree model for the first target intermediate node that matches the trading device until the first target leaf node is reached. Based on the classification label corresponding to the first target leaf node of the first decision tree model, determine the anomaly judgment result corresponding to the corresponding transaction device; Based on the anomaly judgment results of each of the aforementioned trading devices, the set of abnormal trading devices is determined from multiple trading devices.

6. The method according to claim 5, characterized in that, The construction steps of the first decision tree model include: Obtain a first sample set and a first classification label corresponding to each device sample. The first sample set includes device feature sample data of each device sample under each first outlier dimension. Based on the device feature sample data of each device sample under each first outlier dimension and the first classification label of each device sample, determine the information gain corresponding to each first outlier dimension. Based on the information gain corresponding to each of the first outlier dimensions, a target first outlier dimension is selected from multiple first outlier dimensions as a classification feature to create the first root node; Based on the device feature sample data under the first root node and the first classification label corresponding to each device feature sample data, create the first intermediate node and the first leaf node; The first decision tree model is determined based on the first root node, the first intermediate node under the first root node, and the first leaf node.

7. The method according to claim 1, characterized in that, The step of determining the second outlier dimension from the multiple second dimensions based on the object feature data of each object under multiple second dimensions includes: Based on the object feature data of each object under multiple second dimensions, determine the distribution of object feature data under each second dimension, and determine the abnormal data under each second dimension based on the distribution. Based on the number of outliers in each second dimension, the second dimension whose number of outliers meets the second preset condition is selected from the plurality of second dimensions and used as the second outlier dimension.

8. The method according to claim 7, characterized in that, The step of determining the distribution of object feature data in each second dimension based on object feature data of each object in multiple second dimensions, and determining abnormal data in each second dimension based on the distribution, includes: Based on the object feature data of each object under multiple second dimensions, determine the mean and standard deviation corresponding to each second dimension; Based on the mean and standard deviation of each of the second dimensions, determine the anomaly threshold corresponding to each of the second dimensions; Based on the object feature data under each second dimension and the corresponding anomaly threshold for each second dimension, the abnormal data under each second dimension is determined.

9. The method according to claim 1, characterized in that, The second outlier dimension has multiple aspects. The step of determining the set of anomalous risk objects from multiple objects based on the target object feature data includes: Obtain the outlier threshold corresponding to each second outlier dimension; Based on the target object feature data under each second outlier dimension and the outlier threshold corresponding to each second outlier dimension, determine the outlier judgment result of each object under each second outlier dimension. Based on the anomaly assessment results of each object under each second outlier dimension, determine the anomaly assessment result of each object; Based on the anomaly assessment results of each object, a set of objects with anomaly risk is determined from multiple objects.

10. The method according to claim 1, characterized in that, The second outlier dimension has multiple aspects. The step of determining the set of anomalous risk objects from multiple objects based on the target object feature data includes: Obtain a pre-constructed second decision tree model, which includes a second root node, multiple second intermediate nodes, and multiple second leaf nodes; For each object, based on the target object feature data under each second outlier dimension corresponding to the corresponding object, starting from the second root node of the second decision tree model, continuously search from top to bottom in the second decision tree model for the second target intermediate node that matches the object, until the second target leaf node is reached; Based on the classification label corresponding to the second target leaf node of the second decision tree model, determine the anomaly judgment result corresponding to the corresponding object; Based on the anomaly assessment results of each object, a set of objects with anomaly risk is determined from multiple objects.

11. The method according to claim 10, characterized in that, The construction steps of the second decision tree model include: Obtain a second sample set and a second classification label corresponding to each object sample. The second sample set includes object feature sample data of each object sample under each second outlier dimension. Based on the object feature sample data of each object sample under each second outlier dimension and the second classification label of each object sample, determine the information gain corresponding to each second outlier dimension; Based on the information gain corresponding to each of the second outlier dimensions, a target second outlier dimension is selected from multiple second outlier dimensions as a classification feature to create a second root node; Based on the object feature sample data under the second root node and the second classification label corresponding to each object feature sample data, create the second intermediate node and the second leaf node; The second decision tree model is determined based on the created second root node, the second intermediate node under the second root node, and the second leaf node.

12. The method according to any one of claims 1 to 11, characterized in that, The method further includes: Each transaction device sends transaction data to the server after it generates transaction data each time. The server uses all the received transaction data as historical transaction data, or extracts transaction data generated by multiple historical transaction operations within a first time period based on the transaction time point, and uses the extracted transaction data as the historical transaction data.

13. An abnormal transaction operation detection device, characterized in that, The device includes: The first determining module is used to determine, based on historical transaction data, the device characteristic data of each transaction device under multiple first dimensions, and the object characteristic data of each object under multiple second dimensions. The historical transaction data includes data generated by multiple objects performing transaction operations on different transaction devices. The transaction devices include at least one of the following: smartphones, computers, facial recognition devices, card swiping devices, palm swiping devices, IoT devices, and portable wearable devices installed in stores. The objects include at least one of users, enterprises, or transaction accounts. The historical transaction data includes transaction data generated by multiple historical transaction operations within a preset historical time period. The second determining module is used to determine a first outlier dimension from the multiple first dimensions based on the device characteristic data of each transaction device under multiple first dimensions. The multiple first dimensions include at least one of the following: transaction device identifier, transaction device location, transaction device model, transaction device source, number of transactions per day per device, transaction amount per day per device, number of transaction objects on the device, number of common social groups among transaction objects on the device, and average transaction amount; and to determine a second outlier dimension from the multiple second dimensions based on the object characteristic data of each object under multiple second dimensions. The multiple second dimensions include at least one of the following: object identifier, object social group information, object region, number of transactions per day per object, number of transaction devices per day, number of device locations, average transaction interval duration, and average transaction time of the object. The third determination module is used to acquire target device feature data of each trading device under the first outlier dimension, and determine an abnormal trading device set from multiple trading devices based on the target device feature data; acquire target object feature data of each object under the second outlier dimension, and determine an abnormal risk object set from multiple objects based on the target object feature data; The fourth determining module is used to determine abnormal transaction operations based on the abnormal transaction device set and the abnormal risk object set, including: determining the target transaction device and target object associated with each historical transaction operation that generated the historical transaction data; and determining that any historical transaction operation is an abnormal transaction operation if the target transaction device associated with any historical transaction operation is in the abnormal transaction device set and the associated target object is in the abnormal risk object set, wherein the abnormal transaction operation includes order brushing.

14. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 12.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 12.

16. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 12.

Citation Information

Patent Citations

  • Abnormal transaction identification method and apparatus, and server

    CN107918905A

  • Click farming detecting method based on user behavior data characteristic and system thereof

    CN108550052A