Matching relationship verification method, device, equipment and machine-readable storage medium

Abstract

The present application provides a matching relationship verification method, apparatus, device and machine-readable storage medium, the method comprising: for each sample matching pair in a plurality of sample matching pairs, obtaining a collision feature of the sample matching pair, determining a fitness value corresponding to the collision feature, and determining a comprehensive confidence of the sample matching pair based on the fitness value; based on the comprehensive confidence of each sample matching pair in the plurality of sample matching pairs, dividing the plurality of sample matching pairs into at least one confidence interval; for each confidence interval, selecting some sample matching pairs from the sample matching pairs corresponding to the confidence interval, and determining the selected sample matching pairs as target sample matching pairs; training an initial network model based on the calibration information of the target sample matching pair to obtain a target network model, the target network model being used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested. Through the technical solution of the present application, the calibration workload and the amount of data to be manually verified can be reduced.

Description

Technical Field

[0001] The present application relates to the field of intelligent management technology, and in particular to a matching relationship verification method, apparatus, device, and machine-readable storage medium. Background Art

[0002] Cameras can be deployed in a target scene to capture images of the target scene and identify objects (such as vehicles) from the images. By collecting a large number of images of an object, its movement trajectory can be determined. However, due to the fixed angle, small coverage area, and limited coverage of the camera, it is impossible to capture images of the object in all areas of the target scene, and thus the complete trajectory of the object cannot be obtained.

[0003] Typically, objects carry terminal devices with unique device identifiers, such as MAC (Media Access Control) addresses and IMSI (International Mobile Subscriber Identity). Therefore, a collector can be deployed in the target scene to collect the device identifiers of the target scene. By collecting a large number of device identifiers of terminal devices, the movement trajectory of the terminal device can be determined, which is also the movement trajectory of the object carrying the terminal device.

[0004] In summary, to obtain the complete trajectory of an object, it is necessary to obtain the matching relationship between the object identifier of the object and the device identifier of the terminal device held by the object. By combining the matching relationship between the object identifier and the device identifier, the movement trajectory of the object corresponding to the object identifier, and the movement trajectory of the terminal device corresponding to the device identifier, the complete trajectory of the object can be obtained. However, due to the large number of objects and terminal devices in the target scene, there is currently no effective way to accurately obtain the matching relationship between the object identifier of the object and the device identifier of the terminal device. Summary of the Invention

[0005] This application provides a matching relationship verification method, the method comprising:

[0006] For each of the plurality of sample matching pairs, a collision feature of the sample matching pair is obtained, a fitness value corresponding to the collision feature is determined, and a comprehensive confidence of the sample matching pair is determined based on the fitness value; wherein the sample matching pair includes a sample object identifier and a sample device identifier, the collision feature represents a feature of a degree of matching between the sample object identifier and the sample device identifier, and the fitness value represents a degree of matching value corresponding to the feature of the degree of matching;

[0007] Based on the comprehensive confidence of each sample matching pair in the multiple sample matching pairs, the multiple sample matching pairs are divided into at least one confidence interval; wherein, for the sample matching pairs divided into the confidence interval, the comprehensive confidence of the sample matching pairs is within the confidence interval;

[0008] For each confidence interval, select some sample matching pairs from all sample matching pairs corresponding to the confidence interval, and determine the selected sample matching pairs as target sample matching pairs;

[0009] The initial network model is trained based on the calibration information of the target sample matching pair to obtain a target network model, which is used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested.

[0010] In one possible implementation, the collision feature includes at least one of the following: an initial confidence level, the initial confidence level being used to verify the matching relationship between the sample object identifier and the sample device identifier; a collision cycle number, the collision cycle number representing the number of unit cycles in which the sample object identifier and the sample device identifier collide; and a collision average number, the collision average number representing the average number of times the sample object identifier and the sample device identifier collide within a unit cycle.

[0011] Determining the fitness value corresponding to the collision feature includes:

[0012] If the collision feature includes an initial confidence level, and the absolute value of the difference between the initial confidence level and the associated confidence level is greater than a confidence threshold, then determining the fitness level corresponding to the initial confidence level to be a first value; if the absolute value of the difference between the initial confidence level and the associated confidence level is not greater than the confidence threshold, then determining the fitness level corresponding to the initial confidence level to be a second value; wherein the associated confidence level is used to verify a matching relationship between the sample object identifier and another sample device identifier other than the sample device identifier;

[0013] If the collision feature includes a collision cycle number, and the collision cycle number is greater than a cycle number threshold, determining that the fitness value corresponding to the collision cycle number is a first value; if the collision cycle number is not greater than the cycle number threshold, determining that the fitness value corresponding to the collision cycle number is less than the first value;

[0014] If the collision feature includes a collision average, and the collision average is greater than the average threshold, the fitness value corresponding to the collision average is determined to be a first value; if the collision average is not greater than the average threshold, the fitness value corresponding to the collision average is determined to be less than the first value.

[0015] In a possible implementation, selecting some sample matching pairs from all sample matching pairs corresponding to the confidence interval includes: determining the number of sample examples corresponding to the confidence interval;

[0016] Sorting the sample matching pairs corresponding to the confidence interval to obtain sorted sample matching pairs;

[0017] Determine a segmentation interval based on the number of sample instances and the total number of sample matching pairs corresponding to the confidence interval;

[0018] The sample matching pairs of the extracted samples are selected from the sorted sample matching pairs based on the segmentation interval.

[0019] Exemplarily, determining the number of sample samples corresponding to the confidence interval includes: determining the number of sample samples corresponding to the confidence interval based on the total number of configured sample samples, the total number of sample matching pairs corresponding to all confidence intervals, and the total number of sample matching pairs corresponding to the confidence interval.

[0020] In one possible implementation, the selecting the number of sample matching pairs extracted from the sorted sample matching pairs based on the segmentation interval includes: selecting the Lth sample matching pair from the sorted sample matching pairs, and selecting (K-1) sample matching pairs from the sorted sample matching pairs based on the Lth sample matching pair and the segmentation interval; wherein, the L is less than or equal to the segmentation interval, the K is the number of extracted samples, and the interval between two adjacent sample matching pairs selected is the segmentation interval.

[0021] In a possible implementation, the training of the initial network model based on the calibration information of the target sample matching pair to obtain the target network model includes:

[0022] Obtaining calibration information of the target sample matching pair, the calibration information including an initial confidence of the target sample matching pair, a fitness value corresponding to a collision feature of the target sample matching pair, and a label value of the target sample matching pair; wherein the initial confidence is used to verify a matching relationship between a sample object identifier and a sample device identifier in the target sample matching pair; and the label value is used to indicate whether the matching relationship between the sample object identifier and the sample device identifier in the target sample matching pair is true or false;

[0023] The calibration information of the target sample matching pair is input into the initial network model, and the initial network model is trained according to the calibration information of the target sample matching pair to obtain a target network model.

[0024] In a possible implementation, after training the initial network model based on the calibration information of the target sample matching pairs to obtain the target network model, the method further includes:

[0025] Obtaining collision features of a matching pair to be tested and the matching pair to be tested, and determining a fitness value corresponding to the collision features of the matching pair to be tested; wherein the matching pair to be tested includes a to-be-tested object identifier and a to-be-tested device identifier, the collision features include at least an initial confidence level, and the initial confidence level is used to verify a matching relationship between the to-be-tested object identifier and the to-be-tested device identifier;

[0026] Inputting the initial confidence of the matching pair to be tested and the fitness value corresponding to the collision feature of the matching pair to be tested into the target network model, and the target network model outputting the label value of the matching pair to be tested and the target confidence of the matching pair to be tested;

[0027] If the label value of the matching pair to be tested is used to indicate that the matching relationship between the object identifier to be tested and the device identifier to be tested is true, and the target confidence of the matching pair to be tested is greater than a preset threshold, then the matching relationship between the object identifier to be tested and the device identifier to be tested is determined to be accurate.

[0028] The present application provides a device for verifying a matching relationship, the device comprising: a determination module for obtaining, for each sample matching pair in a plurality of sample matching pairs, a collision feature of the sample matching pair, determining a fitness value corresponding to the collision feature, and determining a comprehensive confidence of the sample matching pair based on the fitness value; the sample matching pair comprises a sample object identifier and a sample device identifier, the collision feature represents a feature of a matching degree between the sample object identifier and the sample device identifier, and the fitness value represents a matching degree value corresponding to the feature of the matching degree; a division module for dividing, based on each sample matching pair in the plurality of sample matching pairs, The comprehensive confidence of the sample matching pairs is divided into at least one confidence interval; wherein, for the sample matching pairs divided into the confidence interval, the comprehensive confidence of the sample matching pairs is within the confidence interval; a selection module is used to select some sample matching pairs from all sample matching pairs corresponding to each confidence interval, and determine the selected sample matching pairs as target sample matching pairs; a training module is used to train the initial network model based on the calibration information of the target sample matching pairs to obtain a target network model; wherein, the target network model is used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested.

[0029] The present application provides a matching relationship verification device, comprising: a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions that can be executed by the processor;

[0030] The processor is configured to execute machine-executable instructions to implement the following steps:

[0031] For each of the plurality of sample matching pairs, a collision feature of the sample matching pair is obtained, a fitness value corresponding to the collision feature is determined, and a comprehensive confidence of the sample matching pair is determined based on the fitness value; wherein the sample matching pair includes a sample object identifier and a sample device identifier, the collision feature represents a feature of a degree of matching between the sample object identifier and the sample device identifier, and the fitness value represents a degree of matching value corresponding to the feature of the degree of matching;

[0032] Based on the comprehensive confidence of each sample matching pair in the multiple sample matching pairs, the multiple sample matching pairs are divided into at least one confidence interval; wherein, for the sample matching pairs divided into the confidence interval, the comprehensive confidence of the sample matching pairs is within the confidence interval;

[0033] For each confidence interval, select some sample matching pairs from all sample matching pairs corresponding to the confidence interval, and determine the selected sample matching pairs as target sample matching pairs;

[0034] The initial network model is trained based on the calibration information of the target sample matching pair to obtain a target network model, which is used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested.

[0035] The present application provides a machine-readable storage medium, on which a number of computer instructions are stored. When the computer instructions are executed by a processor, the above-mentioned method steps are implemented.

[0036] As can be seen from the above technical solutions, in the embodiment of the present application, by training the target network model and verifying the matching relationship between the identification of the object to be tested and the identification of the device to be tested through the target network model, that is, verifying whether the matching relationship between the identification of the object to be tested and the identification of the device to be tested is accurate, thereby accurately obtaining the matching relationship between the identification of the object to be tested and the identification of the device to be tested, and eliminating the wrong matching relationship. When training the target network model, some sample matching pairs are selected from a large number of sample matching pairs, and the target network model is trained by some sample matching pairs, so that only some sample matching pairs need to be calibrated, avoiding the calibration work of a large number of sample matching pairs, reducing the calibration workload, reducing the amount of data for manual verification, and solving the problems of large amount of manual verification data and difficulty of manual verification. When selecting some sample matching pairs from a large number of sample matching pairs, some sample matching pairs are selected based on the comprehensive confidence of the sample matching pairs. Representative sample matching pairs can be selected from a large number of sample matching pairs. When these sample matching pairs are used to train the target network model, the reliability of the target network model is higher, and the verification quality of the target network model can be guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of an application scenario in one embodiment of the present application;

[0038] Figure 2 This is a flowchart of a matching relationship verification method in one embodiment of the present application;

[0039] Figure 3 This is a flowchart of a matching relationship verification method in one embodiment of the present application;

[0040] Figure 4 This is a data processing flow chart in one embodiment of the present application;

[0041] Figure 5 It is a structural diagram of a matching relationship verification device in one embodiment of the present application;

[0042] Figure 6 This is a hardware structure diagram of a matching relationship verification device in one embodiment of the present application. DETAILED DESCRIPTION

[0043] The terms used in the embodiments of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The singular forms "a," "the," and "the" used in this application and claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to any or all possible combinations of one or more associated listed items.

[0044] It should be understood that although the terms first, second, third, etc. may be used to describe various information in the embodiments of the present application, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" used may also be interpreted as "at the time of" or "when" or "in response to determining".

[0045] See also Figure 1 As shown, it is a schematic diagram of the application scenario of an embodiment of the present application. A target scene (i.e., the area to be managed, where information about the object in the area to be managed needs to be determined) can deploy multiple cameras and multiple collectors.

[0046] Each camera can capture an image of a target scene and identify an object (such as a vehicle) from the image. It then sends data such as the object's identifier (which can be represented by object features (such as license plate features) or other types of object identifiers, without limitation), the capture time (indicating the image was captured at that capture time), and the physical location (such as latitude and longitude coordinates, indicating the object's location at that capture time) to a management device. The management device records the mapping between the object identifier, the capture time, and the physical location in a first database, as shown in Table 1.

[0047] Table 1

[0048]

[0049] Each collector can collect the device identifier (such as MAC address, IMSI, etc.) of the terminal device in the target scene and send data such as the device identifier, collection time (indicating that the device identifier was collected at the collection time), and physical location (such as latitude and longitude coordinates, indicating that the terminal device was at the physical location at the collection time) to the management device. The management device records the mapping relationship between the device identifier, the collection time, and the physical location in the second database, as shown in Table 2.

[0050] Table 2

[0051]

[0052] As can be seen from Table 1, based on all the collection times and physical locations corresponding to object identifier a1, the object's movement trajectory can be determined, namely, the trajectory consisting of physical locations c11, c12, and c13. As can be seen from Table 2, based on all the collection times and physical locations corresponding to device identifier d1, the movement trajectory of the terminal device (i.e., the movement trajectory of the object holding the terminal device) can be determined, namely, the trajectory consisting of physical locations e11, e12, and e13.

[0053] For example, in order to associate the movement trajectory of an object (determined based on data in the first database) with the movement trajectory of a terminal device (determined based on data in the second database), it is necessary to determine the matching relationship between the object identifier in Table 1 and the device identifier in Table 2. The matching relationship indicates that the object identifier matches the device identifier, that is, the terminal device corresponding to the device identifier is held by the object corresponding to the object identifier.

[0054] For example, if collection time e11 is the same as or similar to collection time b11 (i.e., the difference between the two is less than the duration threshold), and physical location f11 is the same as or similar to physical location c11 (i.e., the difference between the two is less than the distance threshold), then device ID d1 is determined to match object ID a1. Similarly, if collection time e21 is the same as or similar to collection time b21, and physical location f21 is the same as or similar to physical location c21, then device ID d2 is determined to match object ID a2, and so on.

[0055] In a possible implementation, in order to obtain the matching relationship between the object identifier and the device identifier, a collision model may be pre-trained, and the matching relationship between the object identifier and the device identifier may be determined using the collision model.

[0056] For example, the data in the first database (such as object identification, collection time, physical location, etc.) is input into the collision model, and the data in the second database (such as device identification, collection time, physical location, etc.) is input into the collision model. The collision model processes these input data to obtain the matching relationship between the object identification and the device identification, and the confidence of the matching relationship (for the convenience of distinction, the confidence will be recorded as the initial confidence later). The initial confidence is used to verify the matching relationship between the object identification and the device identification. In other words, the initial confidence is a probability value, which indicates the probability value that the object identification and the device identification have a matching relationship. For example, if the initial confidence of the matching relationship between the object identification and the device identification is 80%, it means that the probability value that the object identification and the device identification have a matching relationship is 80%.

[0057] Regarding the training process of the collision model, there is no restriction in this embodiment. The collision model can be trained using training data to obtain a trained collision model, and the trained collision model has the following functions: after the data in the first database and the data in the second database are input into the trained collision model, the collision model can output the matching relationship between the object identifier and the device identifier, and the initial confidence of the matching relationship.

[0058] Regarding the process of the collision model outputting the matching relationship between the object identifier and the device identifier, and the initial confidence of the matching relationship, the collision model can determine the matching relationship between the object identifier and the device identifier based on the collection time and physical location corresponding to the object identifier, and the collection time and physical location corresponding to the device identifier, and determine the initial confidence of the matching relationship. The implementation process will not be described in detail.

[0059] The collision model is an example of a collision algorithm. In this embodiment, the data in the first database and the data in the second database can be input into the collision algorithm, which processes the input data to obtain a matching relationship between the object identifier and the device identifier, as well as the initial confidence level of the matching relationship. There are no restrictions on the implementation of this collision algorithm. The collision algorithm can be a machine learning algorithm (such as a deep learning algorithm). The collision model described above is one implementation of the machine learning algorithm. In this embodiment, the collision model is used as an example.

[0060] In one possible implementation, after the data in the first database and the data in the second database are input into the collision algorithm, the collision algorithm can process the input data and output the content shown in Table 3. Of course, Table 3 is only an example, and the collision algorithm can output part or all of the content.

[0061] Table 3

[0062] Field Name Field Type Field Description rowkey string Unique identifier account string Account ID account_type string Account Type relation_account string Associated account ID relation_account_type string Linked account type relation_type string Associated Types relation_count bigint (bits) Number of associations relation_days bigint Associated days relation_devs bigint Number of associated devices account_count bigint Total number of account collections relation_account_count bigint Total number of linked account collections relation_similarity float Initial confidence calculate_cycle int (integer type) Calculation cycle calcul_start bigint Statistics start time calcul_end bigint Statistics end time create_time bigint Data generation time

[0063] In Table 3, rowkey represents the unique identifier of the data, that is, the unique identifier of the matching relationship between the object identifier and the device identifier. account represents the object identifier. relation_account represents the device identifier, that is, the device identifier of the terminal device associated with the object identifier. relation_type represents the matching type between the object and the device, for example, the vehicle MAC type (indicating that the object corresponding to the object identifier is a vehicle and the device identifier is a MAC address). relation_count represents the number of associations between the object identifier and the device identifier, that is, the total number of times the object identifier and the device identifier are associated. relation_days represents the number of days the object identifier and the device identifier are associated, that is, the number of days the object identifier and the device identifier are associated. relation_devs represents the number of associated devices for the object identifier, that is, the number of device identifiers associated with the object identifier. account_count represents the total number of times the object identifier has been collected, that is, the total number of times the object identifier has been collected. relation_account_count represents the total number of times the device identifier has been collected, that is, the total number of times the device identifier has been collected. relation_similarity represents the initial confidence level that the object identifier and the device identifier have a matching relationship.

[0064] As can be seen from Table 3, the collision algorithm outputs the matching relationship between the object ID (account) and the device ID (relation_account), as well as the initial confidence level (relation_similarity) of the matching relationship. This initial confidence level is used to verify the matching relationship between the object ID and the device ID. For example, a high initial confidence level indicates a high probability that the object ID and the device ID are a match; a low initial confidence level indicates a low probability that the object ID and the device ID are a match.

[0065] If the reliability of the collision algorithm is high, the initial confidence level output by the collision algorithm is credible, and it is possible to determine whether the matching relationship between the object identifier and the device identifier is accurate based on the initial confidence level. However, if the reliability of the collision algorithm is low, the initial confidence level output by the collision algorithm is unreliable, and it is impossible to determine whether the matching relationship between the object identifier and the device identifier is accurate based on the initial confidence level. For example, the object identifier a1 and the device identifier d1 actually have a matching relationship, but the initial confidence level of the matching relationship is low, resulting in the final identification that the object identifier a1 and the device identifier d1 do not have a matching relationship, and an incorrect matching relationship is obtained. For another example, the object identifier a1 and the device identifier d1 actually do not have a matching relationship, but the initial confidence level of the matching relationship is high, resulting in the final identification that the object identifier a1 and the device identifier d1 have a matching relationship, and an incorrect matching relationship is obtained.

[0066] In summary, since a large number of objects and terminal devices may appear in the target scene, an erroneous matching relationship may be obtained, and the matching relationship between the object identifier and the device identifier cannot be accurately obtained.

[0067] In response to the above findings, an embodiment of the present application proposes a matching relationship verification method, which trains a target network model and verifies the matching relationship between the identification of the object to be tested and the identification of the device to be tested through the target network model, that is, verifies whether the matching relationship between the identification of the object to be tested and the identification of the device to be tested is accurate, thereby accurately obtaining the matching relationship between the identification of the object to be tested and the identification of the device to be tested, and eliminating erroneous matching relationships.

[0068] The following describes the matching relationship verification method of the embodiment of the present application in conjunction with specific embodiments.

[0069] In order to verify the matching relationship, in an embodiment of the present application, a target network model is additionally trained on the basis of the collision model. This target network model is used to verify the matching relationship output by the collision model. Therefore, this embodiment involves a training process of the target network model and a testing process of the target network model. During the training process of the target network model, the target network model can be trained, and during the testing process of the target network model, the matching relationship output by the collision model can be verified by the target network model.

[0070] See also Figure 2 FIG. 1 is a flow chart of a matching relationship verification method in this embodiment. The method can be applied to a management device and used to train a target network model. The method may include:

[0071] Step 201: For each sample matching pair among multiple sample matching pairs, obtain a collision feature of the sample matching pair, determine a fitness value corresponding to the collision feature, and determine a comprehensive confidence of the sample matching pair based on the fitness value. The sample matching pair may include a sample object identifier and a sample device identifier. The collision feature represents a feature of the degree of matching between the sample object identifier and the sample device identifier, and the fitness value represents a matching degree value corresponding to the feature of the matching degree.

[0072] Illustratively, the degree of match between the sample object identifier and the sample device identifier represents the degree of match between the sample object corresponding to the sample object identifier and the terminal device corresponding to the sample device identifier, i.e., the degree of match between the sample object and the terminal device, and the collision feature is a feature representing the degree of match. The fitness value represents the degree of match corresponding to the feature of the degree of match. For example, a larger fitness value indicates a greater degree of match between the sample object and the terminal device, while a smaller fitness value indicates a smaller degree of match between the sample object and the terminal device.

[0073] Exemplarily, after the data in the first database and the data in the second database are input into the collision algorithm, the collision algorithm can output the matching relationship between the object identifier and the device identifier, as well as the initial confidence of the matching relationship. See Table 3, which is an example of data output by the collision algorithm.

[0074] In this embodiment, for the sake of convenience in distinction, the object identifier of the training process can be called a sample object identifier, and the device identifier of the training process can be called a sample device identifier. The sample object identifier and the sample device identifier constitute a matching pair, and the matching pair of the training process can be called a sample matching pair. That is, the sample matching pair is used to represent the matching relationship between the sample object identifier and the sample device identifier.

[0075] In summary, the collision algorithm can output multiple sample matching pairs, as well as the initial confidence, number of associations, number of association days, number of associated devices, etc. corresponding to each sample matching pair, as shown in Table 3.

[0076] Based on the content output by the collision algorithm, the management device can obtain multiple sample matching pairs, each sample matching pair can include a sample object identifier and a sample device identifier, and the management device can obtain the initial confidence, number of associations, number of association days, number of associated devices, and other content corresponding to each sample matching pair.

[0077] In step 201, it is necessary to determine the comprehensive confidence of each sample matching pair. Considering that the implementation process of each sample matching pair is the same, taking the implementation process of one sample matching pair as an example, for each sample matching pair in multiple sample matching pairs, the comprehensive confidence of the sample matching pair is determined using the following steps:

[0078] Step 2011: Obtain the collision features of the sample matching pair.

[0079] In a possible implementation, the collision feature may include but is not limited to at least one of the following:

[0080] Initial confidence: The initial confidence is used to verify the matching relationship between the sample object identifier and the sample device identifier. As shown in Table 3, since the collision algorithm can output the initial confidence of the sample matching pair, the initial confidence of the sample matching pair can be directly obtained from the output of the collision algorithm.

[0081] Collision Cycle Count: The collision cycle count represents the number of unit cycles during which a sample object identifier and a sample device identifier collide. For example, if the unit cycle is "day," the collision cycle count can represent the number of days during which a sample object identifier and a sample device identifier collide. This collision cycle count can also be called the collision day count. As shown in Table 3, since the collision algorithm outputs the associated days (i.e., the collision cycle count) for a sample matching pair, the collision cycle count for that sample matching pair can be directly obtained from the output of the collision algorithm.

[0082] Collision Average: The collision average represents the average number of collisions between the sample object identifiers and the sample device identifiers within a unit period. For example, if the unit period is a day, the collision average can represent the average number of collisions between the sample object identifiers and the sample device identifiers per day. For example, the collision average can be the quotient of the total number of collisions and the total number of collision days. As shown in Table 3, since the collision algorithm can output the number of associations (i.e., the total number of collisions) and the number of association days (i.e., the total number of collisions) for a sample matching pair, the total number of collisions and the total number of collision days for the sample matching pair can be obtained from the collision algorithm output, and the quotient of the total number of collisions and the total number of collision days can be used as the collision average.

[0083] Of course, the above are just a few examples of collision features, and there is no limitation to the collision features.

[0084] Step 2012: Determine the fitness value corresponding to the collision feature.

[0085] For example, if the collision feature includes an initial confidence level, the fitness value corresponding to the initial confidence level is determined, which can be subsequently recorded as fitness value s1. If the collision feature includes the number of collision cycles, the fitness value corresponding to the number of collision cycles is determined, which can be subsequently recorded as fitness value s2. If the collision feature includes the average number of collisions, the fitness value corresponding to the average number of collisions is determined, which can be subsequently recorded as fitness value s3.

[0086] For another example, if the collision feature includes an initial confidence level and a number of collision cycles, fitness values ​​s1 and s2 can be determined. If the collision feature includes an initial confidence level and an average number of collisions, fitness values ​​s1 and s3 can be determined. If the collision feature includes a number of collision cycles and an average number of collisions, fitness values ​​s2 and s3 can be determined. If the collision feature includes an initial confidence level, a number of collision cycles, and an average number of collisions, fitness values ​​s1, s2, and s3 can be determined.

[0087] For example, the fitness value corresponding to the collision feature may be determined using the following conditions:

[0088] Case 1: If the collision feature includes an initial confidence, the fitness value s1 corresponding to the initial confidence is determined based on the initial confidence, the associated confidence, and the confidence threshold. For example, if the absolute value of the difference between the initial confidence and the associated confidence is greater than the confidence threshold, the fitness value s1 corresponding to the initial confidence is determined to be a first value (such as 1). If the absolute value of the difference between the initial confidence and the associated confidence is not greater than the confidence threshold, the fitness value s1 corresponding to the initial confidence is determined to be a second value (such as 0). The associated confidence is used to verify the matching relationship between the sample object identifier and another sample device identifier other than the sample device identifier.

[0089] For example, the sample matching pair g1 includes the sample object identifier a1 and the sample device identifier d1, and the sample matching pair g2 includes the sample object identifier a1 and the sample device identifier d2. Obviously, the sample matching pair g1 and the sample matching pair g2 correspond to the same sample object identifier a1, but different sample device identifiers.

[0090] If step 201 is performed for the sample matching pair g1, the initial confidence is the confidence corresponding to the sample matching pair g1, and the associated confidence is the confidence corresponding to the sample matching pair g2. That is, the initial confidence is used to verify the matching relationship between the sample object identifier a1 and the sample device identifier d1, while the associated confidence is used to verify the matching relationship between the sample object identifier a1 and another sample device identifier d2 other than the sample device identifier d1.

[0091] Similarly, if step 201 is performed for the sample matching pair g2, the initial confidence may be the confidence corresponding to the sample matching pair g2, and the associated confidence may be the confidence corresponding to the sample matching pair g1.

[0092] To summarize, if step 201 is executed for the sample matching pair g1, when the absolute value of the difference between the initial confidence and the associated confidence is greater than the confidence threshold, the fitness value s1 is determined to be 1; when the absolute value of the difference between the initial confidence and the associated confidence is not greater than the confidence threshold, the fitness value s1 is determined to be 0.

[0093] In one possible implementation, the sample object identifier a1 may have a matching relationship with multiple sample device identifiers. However, the number of terminal devices carried by an object does not exceed k (e.g., 5). Therefore, for the sample object identifier a1, based on the initial confidence of the sample object identifier a1 and each sample device identifier, k initial confidences with the largest initial confidences may be retained and sorted. For example, the sorting result may be initial confidence h1, initial confidence h2, initial confidence h3, initial confidence h4, and initial confidence h5. Initial confidence h1 is the initial confidence of the sample matching pair g1, initial confidence h2 is the initial confidence of the matching relationship of the sample matching pair g2, and so on.

[0094] When executing step 201 for each sample matching pair corresponding to sample object identifier a1 (e.g., sample matching pair g1, sample matching pair g2, etc.), if the absolute value of the difference between any two initial confidences (e.g., the difference between initial confidence h1 and initial confidence h2, or the difference between initial confidence h2 and initial confidence h3, etc.) is greater than the confidence threshold, the fitness value s1 corresponding to the initial confidence of the sample matching pair is determined to be the first value. If the absolute value of the difference between any two initial confidences is not greater than the confidence threshold, the fitness value s1 corresponding to the initial confidence of the sample matching pair is determined to be the second value.

[0095] For example, if the initial confidence h1 is 0.9, the initial confidence h2 is 0.6, and the confidence threshold is 0.2, then the difference between the initial confidence h1 and the initial confidence h2 is greater than the confidence threshold, and the fitness value s1 is determined to be the first value. The difference between the initial confidence h1 and the initial confidence h2 is greater than the confidence threshold, indicating that there is a hierarchical difference in the confidence results. For example, the object most frequently uses one terminal device, and the number of times the sample device identifier of this terminal device appears should be greater than the number of times the sample device identifiers of other terminal devices appear. Therefore, there should be a hierarchical difference in the confidence results. It is for this reason that when the difference between the initial confidence h1 and the initial confidence h2 is greater than the confidence threshold, the fitness value s1 is determined to be the first value (such as 1).

[0096] Case 2: If the collision feature includes the number of collision cycles, the fitness value s2 corresponding to the number of collision cycles is determined based on the number of collision cycles and the cycle threshold. For example, if the number of collision cycles is greater than the cycle threshold, the fitness value s2 corresponding to the number of collision cycles is determined to be a first value (e.g., 1). If the number of collision cycles is not greater than the cycle threshold, the fitness value s2 corresponding to the number of collision cycles is determined to be less than the first value. For example, the fitness value s2 corresponding to the number of collision cycles can be the quotient of the number of collision cycles and the cycle threshold.

[0097] Because the collision algorithm prioritizes longitudinal comparisons of individuals, the initial confidence level for some individuals with poor data quality, such as those with only one collision, may be higher, for example, above 0.9. Therefore, the number of collision cycles can be used as one of the decision criteria for the overall confidence level. For example, if the number of collision cycles is greater than a cycle threshold (configured based on experience, such as 3 or 5 days), the fitness value s2 is set to 1. If the number of collision cycles is less than the cycle threshold, the fitness value s2 is determined to be the number of collision cycles divided by the cycle threshold.

[0098] Case 3: If the collision feature includes a collision average, the fitness value s3 corresponding to the collision average is determined based on the collision average and the average threshold. For example, if the collision average is greater than the average threshold, the fitness value s3 corresponding to the collision average is determined to be a first value (e.g., 1). If the collision average is not greater than the average threshold, the fitness value s3 corresponding to the collision average is determined to be less than the first value. For example, the fitness value s3 corresponding to the collision average can be the quotient of the collision average and the average threshold.

[0099] Because the collision average indicates that the matching relationship shifts simultaneously in space and time, it better reflects the dynamic nature of the terminal device (such as the sample device identifier) ​​moving with the object. Therefore, the collision average can be used as one of the decision criteria for the comprehensive confidence. For example, if the collision average is greater than the average threshold (configured based on experience, such as 5, 6, 7, etc.), the fitness value s3 is determined to be 1. If the collision average is not greater than the average threshold, the fitness value s3 is determined to be the collision average divided by the average threshold.

[0100] Step 2013: Determine the comprehensive confidence of the sample matching pair based on the fitness value.

[0101] In a possible implementation, the comprehensive confidence of the sample matching pair can be recorded as s all , based on this, if the collision feature includes the initial confidence, then s all =s1. If the collision characteristics include the number of collision cycles, then s all =s2. If the collision characteristics include the collision average, then s all=s3. If the collision characteristics include the initial confidence and the number of collision cycles, then s all =s1+s2. If the collision characteristics include the initial confidence and the collision average, then s all =s1+s3. If the collision characteristics include the number of collision cycles and the average number of collisions, then s all =s2+s3. If the collision characteristics include initial confidence, number of collision cycles and average number of collisions, then s all =s1+s2+s3.

[0102] At this point, for each sample matching pair, the comprehensive confidence of the sample matching pair can be obtained. The comprehensive confidence is the screening condition factor value obtained by different fitness values ​​based on the collision result data.

[0103] Step 202: Based on the comprehensive confidence of each sample matching pair in the multiple sample matching pairs, the multiple sample matching pairs are divided into at least one confidence interval; illustratively, for the sample matching pairs divided into the confidence interval, the comprehensive confidence of the sample matching pairs is within the confidence interval.

[0104] For example, the total confidence range can be divided into multiple confidence intervals, such as confidence interval i1, confidence interval i2, confidence interval i3, etc. For example, when the value range of the comprehensive confidence is 0-3, the total confidence range can be [0,3]. Assuming that the confidence interval is divided according to 0.5, the total confidence range can be divided into 6 confidence intervals, namely [0,0.5), [0.5,1), [1,1.5), [1.5,2), [2,2.5), and [2.5,3].

[0105] For example, after obtaining the comprehensive confidence of each sample matching pair in multiple sample matching pairs, for each sample matching pair, the confidence interval to which the sample matching pair belongs can be determined based on the comprehensive confidence of the sample matching pair, and the sample matching pair can be divided into the confidence interval.

[0106] For example, confidence interval i1 corresponds to [0, 0.5), confidence interval i2 corresponds to [0.5, 1), confidence interval i3 corresponds to [1, 1.5), and so on. If the overall confidence of sample matching pair g1 is 0.25, 0.25 is in [0, 0.5), then sample matching pair g1 is classified into confidence interval i1. If the overall confidence of sample matching pair g2 is 0.8, 0.8 is in [0.5, 1), then sample matching pair g2 is classified into confidence interval i2, and so on.

[0107] Step 203 : For each confidence interval, select some sample matching pairs from all sample matching pairs corresponding to the confidence interval, and determine the selected sample matching pairs as target sample matching pairs.

[0108] In order to train the target network model, in this embodiment, some sample matching pairs are selected from a large number of sample matching pairs based on the comprehensive confidence of the sample matching pairs, and some representative sample matching pairs can be selected from a large number of sample matching pairs, so that the target network model is trained by these selected sample matching pairs, making the target network model more reliable and ensuring the verification quality of the target network model.

[0109] In order to select some representative sample matching pairs from a large number of sample matching pairs as target sample matching pairs, in this embodiment, a combination of stratified sampling and systematic sampling can be adopted.

[0110] For example, since the purpose of sampling is to represent the overall data through a small sample, it is necessary to ensure sample representativeness, that is, to maintain consistency between the sample data structure and the overall data structure. However, in the result data of the collision algorithm, the amount of data in each interval varies greatly. Therefore, stratified sampling can be used to ensure the consistency of the sampling structure. Stratified sampling is to randomly select individuals from different strata according to a specified proportion from a population that can be divided into different sub-populations (strata). It has the advantages of good sample representativeness, probability sampling, and small sampling error.

[0111] Systematic sampling is to extract samples from the population based on a certain sampling distance, divide the population into several balanced parts, and extract an individual from each part. It is a variant of pure random sampling. The samples extracted are evenly distributed in the population. It has the advantages of simple operation, short time consumption, and low error rate.

[0112] When a combination of stratified sampling and systematic sampling is used, for each confidence interval, the following steps can be used to select some sample matching pairs from all sample matching pairs corresponding to the confidence interval:

[0113] Step 2031: Determine the number of samples corresponding to the confidence interval.

[0114] For example, the number of samples drawn corresponding to different confidence intervals may be the same or different.

[0115] For example, for each confidence interval, such as confidence interval i1, confidence interval i2, etc., the number of sample examples corresponding to the confidence interval can be determined, that is, only the number of sample matching pairs corresponding to the confidence interval needs to be selected to achieve the purpose of stratified sampling.

[0116] For example, the following steps may be used to determine the number of samples corresponding to the confidence interval:

[0117] Step 20311: Get the total number of configured extraction samples P1.

[0118] For example, a total number of extracted samples P1 can be pre-configured, indicating that P1 sample matching pairs need to be selected from all sample matching pairs corresponding to all confidence intervals, that is, the total number of target sample matching pairs is P1. Based on this, the pre-configured total number of extracted samples P1 can be obtained.

[0119] Step 20312: Obtain the total number P2 of sample matching pairs corresponding to all confidence intervals.

[0120] For example, in step 201, the comprehensive confidence of P2 sample matching pairs can be obtained. The total number of sample matching pairs corresponding to all confidence intervals is P2. P2 sample matching pairs need to be divided into each confidence interval. Based on this, in step 20312, the total number of sample matching pairs P2 can be obtained.

[0121] Step 20313: Obtain the total number P3 of sample matching pairs corresponding to the confidence interval.

[0122] For example, in step 202, P2 sample matching pairs have been divided into various confidence intervals. Therefore, in step 20313, the total number P3 of sample matching pairs corresponding to each confidence interval can be obtained.

[0123] For example, when determining the number of sample examples corresponding to the confidence interval i1, the total number of sample matching pairs P3-1 corresponding to the confidence interval i1 can be obtained. When determining the number of sample examples corresponding to the confidence interval i2, the total number of sample matching pairs P3-2 corresponding to the confidence interval i2 can be obtained, and so on.

[0124] Step 20314: For each confidence interval, determine the number of sample samples K corresponding to the confidence interval based on the configured total number of sample samples P1, the total number of sample matching pairs corresponding to all confidence intervals P2, and the total number of sample matching pairs corresponding to the confidence interval P3.

[0125] For example, the ratio (i.e., the quotient) between the total number of sample matching pairs P3 corresponding to the confidence interval and the total number of sample matching pairs P2 corresponding to all confidence intervals can be determined. Then, the product of this ratio and the total number of sample examples P1 can be used as the number of sample examples K corresponding to the confidence interval.

[0126] For example, when determining the number of samples K1 corresponding to the confidence interval i1, the following formula can be used: K1 = (P3-1 / P2)*P1; when determining the number of samples K2 corresponding to the confidence interval i2, the following formula can be used: K2 = (P3-2 / P2)*P1, and so on.

[0127] For example, assuming that the total number of sample examples P1 is 50, the total number of sample matching pairs corresponding to all confidence intervals P2 is 1000, and the total number of sample matching pairs corresponding to confidence interval i1 P3-1 is 100, then the number of sample examples K1 corresponding to confidence interval i1 is 5. Suppose that the total number of sample matching pairs corresponding to confidence interval i2 P3-2 is 200, then the number of sample examples K2 corresponding to confidence interval i2 is 10, and so on.

[0128] At this point, for each confidence interval, the number of samples corresponding to the confidence interval can be determined.

[0129] Step 2032: Sort all sample matching pairs corresponding to the confidence interval to obtain sorted sample matching pairs. For example, all sample matching pairs corresponding to the confidence interval are sorted according to their collision time, and the sorted sequence values ​​are used as the sample matching pair numbers.

[0130] For example, all sample matching pairs corresponding to the confidence interval can be sorted in order of collision time from front to back (i.e., from earliest to latest), and the sorting results are sample matching pair g1, sample matching pair g2, sample matching pair g3, ..., with sample matching pair g1 numbered 1, sample matching pair g2 numbered 2, and so on. Of course, all sample matching pairs corresponding to the confidence interval can also be sorted in order of collision time from back to front (i.e., from latest to earliest), and this sorting process will not be described in detail.

[0131] Step 2033: For each confidence interval, a segmentation interval is determined based on the number of samples drawn corresponding to the confidence interval and the total number of sample matching pairs corresponding to the confidence interval. For example, the quotient of the total number of sample matching pairs and the number of samples drawn is rounded up or down, and the rounded value is used as the segmentation interval corresponding to the confidence interval.

[0132] For example, for confidence interval i1, the number of samples drawn is K1, and the total number of sample matching pairs is P3-1. Therefore, the quotient of P3-1 and K1 is rounded, and the rounded value is used as the segment interval for confidence interval i1. For example, when P3-1 is 100 and K1 is 5, the segment interval is 20. For confidence interval i2, the number of samples drawn is K2, and the total number of sample matching pairs is P3-2. Therefore, the quotient of P3-2 and K2 is rounded, and the rounded value is used as the segment interval for confidence interval i2, and so on.

[0133] Step 2034: For each confidence interval, select the sample matching pairs of the extracted samples from the sorted sample matching pairs corresponding to the confidence interval based on the segmentation interval corresponding to the confidence interval.

[0134] For example, for the confidence interval i1, K1 sample matching pairs are selected from the sorted sample matching pairs corresponding to the confidence interval i1 based on the segmentation interval corresponding to the confidence interval i1, and so on.

[0135] In one possible implementation, the Lth sample matching pair can be first selected from the sorted sample matching pairs, and (K-1) sample matching pairs can be selected from the sorted sample matching pairs based on the Lth sample matching pair and the segmentation interval; illustratively, L is less than or equal to the segmentation interval, K is the number of sample instances, and the interval between two adjacent sample matching pairs selected is the segmentation interval.

[0136] For example, if the segment interval is denoted as M, then the Lth sample matching pair is first selected from the first M sample matching pairs, such as randomly selecting a sample matching pair from the first M sample matching pairs, and the number of the selected sample matching pair is denoted as L. Then, the L+Mth sample matching pair, the L+2Mth sample matching pair, ..., the L+(K-1)Mth sample matching pair are sequentially selected from the sorted sample matching pairs.

[0137] In summary, a total of K sample matching pairs are selected from the sorted sample matching pairs.

[0138] For confidence interval i1, the segmentation interval is M1, the number of samples extracted is K1, and the first sample matching pair is numbered L1. For confidence interval i2, the segmentation interval is M2, the number of samples extracted is K2, and the first sample matching pair is numbered L2, and so on. In this way, for each confidence interval, the above method can be used to select sample matching pairs equal to the number of samples extracted from all sample matching pairs corresponding to the confidence interval, and the selected sample matching pairs are used as the target sample matching pairs corresponding to the confidence interval.

[0139] Based on steps 2032 to 2034, representative data within each confidence interval can be determined, and a systematic sampling method is used to determine the representative data within each confidence interval, thereby achieving the purpose of systematic sampling.

[0140] At this point, in step 203 , for each confidence interval, some sample matching pairs may be selected from all sample matching pairs corresponding to the confidence interval as target sample matching pairs.

[0141] Step 204: train the initial network model based on the calibration information of the target sample matching pair to obtain a target network model. The target network model is used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested. For the verification process of the target network model, please refer to the subsequent embodiments.

[0142] In one possible implementation, the following steps may be used to train the initial network model:

[0143] Step 2041: Obtain calibration information for the target sample matching pair. This calibration information includes the initial confidence level of the target sample matching pair, the fitness value corresponding to the collision feature of the target sample matching pair, and the label value of the target sample matching pair. Exemplarily, the initial confidence level is used to verify the matching relationship between the sample object identifier and the sample device identifier in the target sample matching pair; the label value is used to indicate whether the matching relationship between the sample object identifier and the sample device identifier in the target sample matching pair is true or false.

[0144] After selecting a target sample matching pair from all sample matching pairs, the initial confidence of each target sample matching pair can be obtained. Referring to the above embodiment, the data output by the collision algorithm includes the initial confidence, so the initial confidence of the target sample matching pair can be directly obtained.

[0145] For each target sample matching pair, the fitness value corresponding to the collision feature of the target sample matching pair can be obtained. The acquisition method refers to step 201. The fitness value corresponding to the collision feature of the target sample matching pair can be obtained, such as at least one of the fitness value s1, the fitness value s2 and the fitness value s3. In subsequent embodiments, the fitness value s1, the fitness value s2 and the fitness value s3 are used as examples for explanation.

[0146] For each target sample matching pair, the label value of the target sample matching pair can also be obtained. The label value can be a first value (such as 1) or a second value (such as 0). The first value is used to indicate that the matching relationship between the sample object identifier and the sample device identifier is true, that is, the sample object identifier and the sample device identifier have a matching relationship. The second value is used to indicate that the matching relationship between the sample object identifier and the sample device identifier is false, that is, the sample object identifier and the sample device identifier do not have a matching relationship.

[0147] In order to obtain the label value of the target sample matching pair, the following method can be adopted: the management device adds a label value to the target sample matching pair, or the user adds a label value to the target sample matching pair. There is no restriction on the method of obtaining this label value. Take the user adding a label value to the target sample matching pair as an example.

[0148] Obviously, in this embodiment, since some sample matching pairs have been selected from a large number of sample matching pairs as target sample matching pairs, and the target sample matching pairs are representative data selected through comprehensive confidence, the user only needs to add label values ​​for the target sample matching pairs, so that the user only needs to add label values ​​for a small number of target sample matching pairs, avoiding the calibration work of a large number of sample matching pairs, reducing the calibration workload, reducing the amount of data to be manually verified, and solving problems such as the large amount of data to be manually verified and the difficulty of manual verification.

[0149] In a possible embodiment, in order to add a label value to the target sample matching pair, for the sample object identifier and the sample device identifier in the target sample matching pair, the movement trajectory corresponding to the sample object identifier (hereinafter referred to as the first movement trajectory) can be displayed on the map, and the movement trajectory corresponding to the sample device identifier (hereinafter referred to as the second movement trajectory) can be displayed on the map. Obviously, by checking whether the first movement trajectory and the second movement trajectory overlap on the map, it can be known whether the sample object identifier and the sample device identifier have a matching relationship. For example, if the first movement trajectory and the second movement trajectory overlap or approximately overlap, then the sample object identifier and the sample device identifier have a matching relationship, and a label value is added to the target sample matching pair, and the label value is the first value. If the first movement trajectory and the second movement trajectory do not overlap (that is, the two trajectories deviate greatly), then the sample object identifier and the sample device identifier do not have a matching relationship, and a label value is added to the target sample matching pair, and the label value is the second value.

[0150] When displaying the first movement trajectory corresponding to the sample object identifier on a map, based on the collection time and physical location corresponding to the sample object identifier, the collection time within the specified time interval is determined, and the physical locations corresponding to these collection time intervals are displayed on the map. The trajectory formed by these physical locations is the first movement trajectory. When displaying the second movement trajectory corresponding to the sample device identifier on a map, based on the collection time and physical location corresponding to the sample device identifier, the collection time within the specified time interval is determined, and the physical locations corresponding to these collection time intervals are displayed on the map. The trajectory formed by these physical locations is the second movement trajectory.

[0151] Obviously, by checking whether the first movement trajectory and the second movement trajectory overlap on the map, it is possible to know whether the sample object identifier and the sample device identifier have a matching relationship, so that the accuracy of the matching relationship can be intuitively verified. Based on the visual interface, the accuracy of the matching relationship can be intuitively felt.

[0152] Step 2042: Input the calibration information of the target sample matching pair into the initial network model, so as to train the initial network model with the calibration information of the target sample matching pair to obtain the target network model.

[0153] For example, after obtaining calibration information for multiple target sample matching pairs, this calibration information can be used as training data and input into an initial network model, which is a pre-configured network model. After inputting this training data into the initial network model, the initial network model can be trained based on this training data to obtain a trained target network model. This embodiment does not limit the training process of the initial network model.

[0154] For example, the initial confidence, fitness value s1, fitness value s2, and fitness value s3 in the calibration information can form a 4*1-dimensional feature vector, a 2*2-dimensional feature vector, or a 1*4-dimensional feature vector, and the label value in the calibration information serves as the label value corresponding to the feature vector. After inputting a large number of feature vectors and the label value corresponding to each feature vector into the initial network model, a target network model can be trained, and the target network model can output the label value corresponding to the feature vector and the target confidence corresponding to the label value. For the processing process of the target network model, please refer to the subsequent embodiments.

[0155] In one possible implementation, the initial network model and the target network model can be any type of machine learning model. For example, both the initial network model and the target network model are decision tree models. That is, the initial decision tree model is trained using calibration information from multiple target sample matching pairs to obtain a trained target decision tree model. Of course, the decision tree model is only an example, and there is no limitation on this model type.

[0156] When the initial decision tree model is trained using the calibration information of multiple target sample matching pairs to obtain a trained target decision tree model, a Cross-Validation algorithm can be used to train the initial decision tree model to obtain a target decision tree model. There is no restriction on this training algorithm.

[0157] Referring to the above embodiment, when the Cross-Validation algorithm is used to train the initial decision tree model, the input data of the initial decision tree model includes a feature vector and a label value corresponding to the feature vector. The feature vector consists of an initial confidence, a fitness value s1, a fitness value s2, and a fitness value s3.

[0158] The initial confidence, fitness value s1, fitness value s2 and fitness value s3 are all used as attribute values ​​of the initial decision tree model, and the attribute value a (such as initial confidence, or fitness value s1, or fitness value s2, or fitness value s3) is a continuous value a i , you can use the continuous value a of the attribute value a i Perform discretization processing.

[0159] When performing discretization processing, the discretization strategy can be a dichotomy method. The formula for the candidate partition point set of the dichotomy discretization strategy is: That is to say, the attributes a of all training data are sorted into intervals [a i +a i +1) median point As candidate partition points, the continuous variables are discretized. After discretization, the optimal partition point is determined according to the information gain formula:

[0160]

[0161] In the above formula, D represents the entire data set, and Ent is the function for calculating information entropy.

[0162] To sum up, the information gain value of each attribute can be calculated in turn, and the attribute with the largest information gain value can be selected in turn as the attribute for decision tree model division, so as to use the decision tree model for classification training.

[0163] Of course, the above is just an example of training the initial decision tree model and there is no limitation to this.

[0164] For example, the above execution order is only for the convenience of describing the examples given. In actual applications, the execution order between the steps can also be changed, and this execution order is not limited. Moreover, in other embodiments, the steps of the corresponding method are not necessarily executed in the order shown and described in this specification, and the steps included in the method may be more or less than those described in this specification. In addition, a single step described in this specification may be decomposed into multiple steps for description in other embodiments; multiple steps described in this specification may also be combined into a single step for description in other embodiments.

[0165] As can be seen from the above technical solutions, in the embodiment of the present application, by training the target network model and verifying the matching relationship between the identification of the object to be tested and the identification of the device to be tested through the target network model, that is, verifying whether the matching relationship between the identification of the object to be tested and the identification of the device to be tested is accurate, thereby accurately obtaining the matching relationship between the identification of the object to be tested and the identification of the device to be tested, and eliminating the wrong matching relationship. When training the target network model, some sample matching pairs are selected from a large number of sample matching pairs, and the target network model is trained by some sample matching pairs, so that only some sample matching pairs need to be calibrated, avoiding the calibration work of a large number of sample matching pairs, reducing the calibration workload, reducing the amount of data for manual verification, and solving the problems of large amount of manual verification data and difficulty of manual verification. When selecting some sample matching pairs from a large number of sample matching pairs, some sample matching pairs are selected based on the comprehensive confidence of the sample matching pairs. Representative sample matching pairs can be selected from a large number of sample matching pairs. When these sample matching pairs are used to train the target network model, the reliability of the target network model is higher, and the verification quality of the target network model can be guaranteed.

[0166] See also Figure 3 FIG. 1 is a flow chart of a method for verifying a matching relationship in an embodiment of the present application. The method can be applied to a management device. During the test of a target network model, the method is used to verify the matching relationship output by the collision model using the target network model. The method may include:

[0167] Step 301: Obtain a collision signature of a matching pair to be tested and the matching pair to be tested, and determine a fitness value corresponding to the collision signature of the matching pair to be tested. Exemplarily, the matching pair to be tested may include a matching object identifier and a matching device identifier to be tested, and the collision signature includes at least an initial confidence level, which is used to verify the matching relationship between the matching object identifier and the matching device identifier to be tested.

[0168] In this embodiment, for the sake of convenience in distinction, the object identifier of the test process is called the object identifier to be tested, and the device identifier of the test process is called the device identifier to be tested. The object identifier to be tested and the device identifier to be tested constitute a matching pair. The matching pair of the test process can be called the matching pair to be tested. That is, the matching pair to be tested is used to represent the matching relationship between the object identifier to be tested and the device identifier to be tested.

[0169] Exemplarily, the collision characteristics of the matching pair to be tested can be determined based on the data output by the collision algorithm. The collision characteristics include at least an initial confidence level. On this basis, the collision characteristics may also include but are not limited to at least one of the following: the number of collision cycles (indicating the number of unit cycles in which the identification of the object to be tested and the identification of the device to be tested collide), the average number of collisions (indicating the average number of times the identification of the object to be tested and the identification of the device to be tested collide within a unit period). This process is similar to step 201 and will not be repeated here.

[0170] After obtaining the collision features of the matching pair to be tested, the fitness values ​​corresponding to the collision features of the matching pair to be tested can be determined, such as the fitness value s1 corresponding to the initial confidence, the fitness value s2 corresponding to the number of collision cycles, and the fitness value s3 corresponding to the average number of collisions. This process is similar to step 201 and will not be repeated here.

[0171] In step 302, the initial confidence of the matching pair to be tested and the fitness value corresponding to the collision feature of the matching pair to be tested are input into the target network model, so that the target network model outputs the label value (such as the first value or the second value) of the matching pair to be tested and the target confidence of the matching pair to be tested.

[0172] For example, the initial confidence, fitness value s1, fitness value s2 and fitness value s3 of the matching pair to be tested can be combined into a 4*1-dimensional feature vector, or a 2*2-dimensional feature vector, or a 1*4-dimensional feature vector, and the feature vector (subsequently recorded as the target feature vector) is input into the target network model.

[0173] Referring to step 204, when training the target network model, the target network model can output a label value corresponding to the feature vector and a target confidence corresponding to the label value. Therefore, after the target feature vector is input to the target network model, the target network model can process the target feature vector without any restriction on this processing process, and obtain a label value corresponding to the target feature vector and a target confidence corresponding to the label value (for the sake of convenience, the confidence output by the target network model is recorded as the target confidence).

[0174] Step 303: If the label value of the matching pair to be tested is used to indicate that the matching relationship between the identification of the object to be tested and the identification of the device to be tested is true, and the target confidence of the matching pair to be tested is greater than a preset threshold (which can be configured based on experience), then it is determined that the matching relationship between the identification of the object to be tested and the identification of the device to be tested is accurate.

[0175] For example, if the label value output by the target network model is a first value (e.g., 1), which indicates that the matching relationship is true, and the target confidence level output by the target network model is greater than a preset threshold, then the matching relationship between the object identifier and the device identifier under test is determined to be accurate, i.e., the two have a matching relationship. For example, if the target confidence level is 90%, then the probability value of the two having a matching relationship is 90%.

[0176] Exemplarily, if the label value of the matching pair to be tested is used to indicate that the matching relationship between the object identifier to be tested and the device identifier to be tested is false, and the target confidence of the matching pair to be tested is greater than a preset threshold, then the matching relationship between the object identifier to be tested and the device identifier to be tested is determined to be incorrect.

[0177] For example, if the label value output by the target network model is a second value (e.g., 0), which indicates that the matching relationship is false, and the target confidence level output by the target network model is greater than a preset threshold, then the matching relationship between the object identifier and the device identifier under test is determined to be false, i.e., the two do not have a matching relationship. For example, if the target confidence level is 90%, then the probability value that the two do not have a matching relationship is 90%.

[0178] Exemplarily, if the label value of the matching pair to be tested is used to indicate that the matching relationship between the identification of the object to be tested and the identification of the device to be tested is true, and the target confidence of the matching pair to be tested is not greater than a preset threshold, then the matching relationship between the identification of the object to be tested and the identification of the device to be tested may be correct or incorrect, that is, it is impossible to accurately know whether the matching relationship is true based on the target confidence. If the label value of the matching pair to be tested is used to indicate that the matching relationship between the identification of the object to be tested and the identification of the device to be tested is false, and the target confidence of the matching pair to be tested is not greater than a preset threshold, then the matching relationship between the identification of the object to be tested and the identification of the device to be tested may be correct or incorrect, that is, it is impossible to accurately know whether the matching relationship is false based on the target confidence.

[0179] In summary, the matching relationship between the identification of the object to be tested and the identification of the device to be tested can be verified based on the target network model, and then it can be determined whether the matching relationship is accurate or incorrect.

[0180] The following combination Figure 4 The data processing flow chart shown is used to illustrate the above embodiment.

[0181] During the training of the target network model, the basic data output by the collision algorithm (see Table 3) is first obtained, such as vehicle MAC collision data and vehicle IMSI collision data. Data processing is then performed based on this basic data. For example, the comprehensive confidence of sample matching pairs is obtained. Based on the comprehensive confidence of each sample matching pair, some sample matching pairs are selected from all sample matching pairs as target sample matching pairs, and calibration information (such as feature vectors and label values) of the target sample matching pairs is obtained. The target network model is then trained based on the calibration information of the target sample matching pairs. This target network model can include a vehicle MAC classifier and a vehicle IMSI classifier. This completes the training process of the target network model.

[0182] During the target network model-based testing process, collision data to be tested (such as the initial confidence level of the matching pairs to be tested and the fitness values ​​corresponding to the collision features of the matching pairs to be tested) is obtained. This data is then fed into the target network model to obtain a classification result, which indicates whether the matching relationship is accurate or incorrect. This classification result can be a vehicle MAC address classification result, a vehicle IMSI classification result, etc. Based on the classification result, IoT data applications such as fusion trajectory splicing and identity relationship association are performed. This process will not be further described.

[0183] Based on the same application concept as the above method, the embodiment of the present application proposes a matching relationship verification device, see Figure 5 FIG. 5 is a schematic diagram of the structure of the device, which includes: a determination module 51, for obtaining, for each sample matching pair in a plurality of sample matching pairs, a collision feature of the sample matching pair, determining a fitness value corresponding to the collision feature, and determining a comprehensive confidence of the sample matching pair based on the fitness value; wherein the sample matching pair includes a sample object identifier and a sample device identifier, the collision feature represents a feature of the degree of matching between the sample object identifier and the sample device identifier, and the fitness value represents a matching degree value corresponding to the feature of the matching degree; a division module 52, for dividing the sample matching pairs based on each sample matching pair in the plurality of sample matching pairs; The comprehensive confidence of the pairing is used to divide the multiple sample matching pairs into at least one confidence interval; wherein, for the sample matching pairs divided into the confidence interval, the comprehensive confidence of the sample matching pairs is in the confidence interval; a selection module 53 is used to select some sample matching pairs from all sample matching pairs corresponding to the confidence interval for each confidence interval, and determine the selected sample matching pairs as target sample matching pairs; a training module 54 is used to train the initial network model based on the calibration information of the target sample matching pairs to obtain a target network model; the target network model is used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested.

[0184] Exemplarily, the collision feature includes at least one of the following: initial confidence, number of collision cycles, and average number of collisions. The determination module 51 determines the fitness value corresponding to the collision feature by:

[0185] If the collision feature includes an initial confidence level, and the absolute value of the difference between the initial confidence level and the associated confidence level is greater than a confidence threshold, then determining the fitness level corresponding to the initial confidence level to be a first value; if the absolute value of the difference between the initial confidence level and the associated confidence level is not greater than the confidence threshold, then determining the fitness level corresponding to the initial confidence level to be a second value; wherein the associated confidence level is used to verify a matching relationship between the sample object identifier and another sample device identifier other than the sample device identifier;

[0186] If the collision feature includes a collision cycle number, and the collision cycle number is greater than a cycle number threshold, determining that the fitness value corresponding to the collision cycle number is a first value; if the collision cycle number is not greater than the cycle number threshold, determining that the fitness value corresponding to the collision cycle number is less than the first value;

[0187] If the collision feature includes a collision average, and the collision average is greater than the average threshold, the fitness value corresponding to the collision average is determined to be a first value; if the collision average is not greater than the average threshold, the fitness value corresponding to the collision average is determined to be less than the first value.

[0188] Exemplarily, when the selection module 53 selects some sample matching pairs from all sample matching pairs corresponding to the confidence interval, it is specifically used to: determine the number of extracted samples corresponding to the confidence interval; sort the sample matching pairs corresponding to the confidence interval to obtain sorted sample matching pairs; determine the segmentation interval based on the number of extracted samples and the total number of sample matching pairs corresponding to the confidence interval; and select the number of sample matching pairs of the extracted samples from the sorted sample matching pairs based on the segmentation interval.

[0189] Based on the same application concept as the above method, the embodiment of the present application proposes a matching relationship verification device (i.e., management device), see Figure 6 As shown, the management device includes: a processor 61 and a machine-readable storage medium 62, wherein the machine-readable storage medium 62 stores machine-executable instructions that can be executed by the processor 61; the processor 61 is configured to execute the machine-executable instructions to implement the following steps:

[0190] For each of the plurality of sample matching pairs, a collision feature of the sample matching pair is obtained, a fitness value corresponding to the collision feature is determined, and a comprehensive confidence of the sample matching pair is determined based on the fitness value; wherein the sample matching pair includes a sample object identifier and a sample device identifier, the collision feature represents a feature of a degree of matching between the sample object identifier and the sample device identifier, and the fitness value represents a degree of matching value corresponding to the feature of the degree of matching;

[0191] Based on the comprehensive confidence of each sample matching pair in the multiple sample matching pairs, the multiple sample matching pairs are divided into at least one confidence interval; wherein, for the sample matching pairs divided into the confidence interval, the comprehensive confidence of the sample matching pairs is within the confidence interval;

[0192] For each confidence interval, select some sample matching pairs from all sample matching pairs corresponding to the confidence interval, and determine the selected sample matching pairs as target sample matching pairs;

[0193] The initial network model is trained based on the calibration information of the target sample matching pair to obtain a target network model, which is used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested.

[0194] Based on the same application concept as the above method, an embodiment of the present application also provides a machine-readable storage medium, on which a number of computer instructions are stored. When the computer instructions are executed by a processor, the matching relationship verification method disclosed in the above example of the present application can be implemented.

[0195] The machine-readable storage medium may be any electronic, magnetic, optical, or other physical storage device that may contain or store information, such as executable instructions, data, and the like. For example, the machine-readable storage medium may be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, a storage drive (such as a hard disk drive), a solid-state drive, any type of storage disk (such as a CD, DVD, etc.), or similar storage media, or a combination thereof.

[0196] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer, which may be in the form of a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email transceiver, game console, tablet computer, wearable device, or any combination of these devices.

[0197] For the convenience of description, the above devices are described as being divided into various units according to their functions. Of course, when implementing this application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0198] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the embodiments of the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0199] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0200] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable device to implement the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0201] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A matching relationship verification method, characterized in that: The method comprises: For each sample matching pair in a plurality of sample matching pairs, a collision feature of the sample matching pair is obtained, a fitness value corresponding to the collision feature is determined, and a comprehensive confidence of the sample matching pair is determined based on the fitness value; wherein, the sample matching pair includes a sample object identifier and a sample device identifier, the collision feature represents a feature of the degree of matching between the sample object identifier and the sample device identifier, and the fitness value represents a matching degree value corresponding to the feature of the matching degree; wherein, the object identifier of the object identified from the image, the acquisition time and physical location corresponding to the object identifier are input into the collision model, and the device identifier of the terminal device, the acquisition time and physical location corresponding to the device identifier are input into the collision model, and the collision model is inputted through the collision model. A matching relationship between a sample object identifier and a sample device identifier is obtained, where the sample object identifier and the sample device identifier form a sample matching pair, and the matching relationship indicates that the sample object identifier matches the sample device identifier, and the terminal device corresponding to the sample device identifier is held by the object corresponding to the sample object identifier; wherein the collision feature includes at least one of the following: an initial confidence level, where the initial confidence level is used to verify the matching relationship between the sample object identifier and the sample device identifier; a collision cycle number, where the collision cycle number indicates the number of unit cycles in which the sample object identifier and the sample device identifier collide; and a collision average number, where the collision average number indicates the average number of times the sample object identifier and the sample device identifier collide within a unit cycle; Based on the comprehensive confidence of each sample matching pair in the multiple sample matching pairs, the multiple sample matching pairs are divided into at least one confidence interval; wherein, for the sample matching pairs divided into the confidence interval, the comprehensive confidence of the sample matching pairs is within the confidence interval; For each confidence interval, select some sample matching pairs from all sample matching pairs corresponding to the confidence interval, and determine the selected sample matching pairs as target sample matching pairs; The initial network model is trained based on the calibration information of the target sample matching pair to obtain a target network model, which is used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested.

2. The method according to claim 1, characterized in that Determining the fitness value corresponding to the collision feature includes: If the collision feature includes an initial confidence level, and the absolute value of the difference between the initial confidence level and the associated confidence level is greater than a confidence threshold, then determining the fitness level corresponding to the initial confidence level to be a first value; if the absolute value of the difference between the initial confidence level and the associated confidence level is not greater than the confidence threshold, then determining the fitness level corresponding to the initial confidence level to be a second value; wherein the associated confidence level is used to verify a matching relationship between the sample object identifier and another sample device identifier other than the sample device identifier; If the collision feature includes a collision cycle number, and the collision cycle number is greater than a cycle number threshold, determining that the fitness value corresponding to the collision cycle number is a first value; if the collision cycle number is not greater than the cycle number threshold, determining that the fitness value corresponding to the collision cycle number is less than the first value; If the collision feature includes a collision average, and the collision average is greater than the average threshold, the fitness value corresponding to the collision average is determined to be a first value; if the collision average is not greater than the average threshold, the fitness value corresponding to the collision average is determined to be less than the first value.

3. The method according to claim 1, characterized in that The selecting of some sample matching pairs from all sample matching pairs corresponding to the confidence interval includes: Determine the number of samples corresponding to the confidence interval; Sorting the sample matching pairs corresponding to the confidence interval to obtain sorted sample matching pairs; Determine a segmentation interval based on the number of sample instances and the total number of sample matching pairs corresponding to the confidence interval; The sample matching pairs of the extracted samples are selected from the sorted sample matching pairs based on the segmentation interval.

4. The method according to claim 3, characterized in that Determining the number of sample examples corresponding to the confidence interval includes: determining the number of sample examples corresponding to the confidence interval based on the total number of configured sample examples, the total number of sample matching pairs corresponding to all confidence intervals, and the total number of sample matching pairs corresponding to the confidence interval.

5. The method according to claim 3, characterized in that The step of selecting the sample matching pairs of the extracted samples from the sorted sample matching pairs based on the segmented interval includes: Selecting an Lth sample matching pair from the sorted sample matching pairs, and selecting K-1 sample matching pairs from the sorted sample matching pairs based on the Lth sample matching pair and the segmentation interval; Wherein, L is less than or equal to the segmentation interval, K is the number of sample instances, and the interval between two adjacent sample matching pairs selected is the segmentation interval.

6. The method according to claim 1, characterized in that The initial network model is trained based on the calibration information of the target sample matching pair to obtain the target network model, including: Obtaining calibration information of the target sample matching pair, the calibration information including an initial confidence of the target sample matching pair, a fitness value corresponding to a collision feature of the target sample matching pair, and a label value of the target sample matching pair; wherein the initial confidence is used to verify a matching relationship between a sample object identifier and a sample device identifier in the target sample matching pair; and the label value is used to indicate whether the matching relationship between the sample object identifier and the sample device identifier in the target sample matching pair is true or false; The calibration information of the target sample matching pair is input into the initial network model, and the initial network model is trained according to the calibration information of the target sample matching pair to obtain a target network model.

7. The method according to claim 1 or 6, characterized in that After the initial network model is trained based on the calibration information of the target sample matching pair to obtain the target network model, the method further includes: Obtaining collision features of a matching pair to be tested and the matching pair to be tested, and determining a fitness value corresponding to the collision features of the matching pair to be tested; wherein the matching pair to be tested includes a to-be-tested object identifier and a to-be-tested device identifier, the collision features include at least an initial confidence level, and the initial confidence level is used to verify a matching relationship between the to-be-tested object identifier and the to-be-tested device identifier; Inputting the initial confidence of the matching pair to be tested and the fitness value corresponding to the collision feature of the matching pair to be tested into the target network model, and the target network model outputting the label value of the matching pair to be tested and the target confidence of the matching pair to be tested; If the label value of the matching pair to be tested is used to indicate that the matching relationship between the object identifier to be tested and the device identifier to be tested is true, and the target confidence of the matching pair to be tested is greater than a preset threshold, then the matching relationship between the object identifier to be tested and the device identifier to be tested is determined to be accurate.

8. A matching relationship verification device, characterized in that: The device comprises: A determination module is used to obtain, for each sample matching pair in a plurality of sample matching pairs, a collision feature of the sample matching pair, determine a fitness value corresponding to the collision feature, and determine a comprehensive confidence of the sample matching pair based on the fitness value; wherein, the sample matching pair includes a sample object identifier and a sample device identifier, the collision feature represents a feature of the degree of matching between the sample object identifier and the sample device identifier, and the fitness value represents a matching degree value corresponding to the feature of the matching degree; wherein, the object identifier of the object identified from the image, the acquisition time and physical location corresponding to the object identifier are input into the collision model, and the device identifier of the terminal device, the acquisition time and physical location corresponding to the device identifier are input into the collision model, and through the collision The model outputs a matching relationship between a sample object identifier and a sample device identifier, where the sample object identifier and the sample device identifier form a sample matching pair, and the matching relationship indicates that the sample object identifier matches the sample device identifier, and the terminal device corresponding to the sample device identifier is held by the object corresponding to the sample object identifier; wherein the collision feature includes at least one of the following: an initial confidence level, which is used to verify the matching relationship between the sample object identifier and the sample device identifier; a collision cycle number, which indicates the number of unit cycles in which the sample object identifier and the sample device identifier collide; and a collision average number, which indicates the average number of times the sample object identifier and the sample device identifier collide within a unit cycle. a partitioning module, configured to partition the plurality of sample matching pairs into at least one confidence interval based on the comprehensive confidence of each sample matching pair in the plurality of sample matching pairs; wherein, for the sample matching pairs partitioned into the confidence interval, the comprehensive confidence of the sample matching pair is within the confidence interval; A selection module is used to select, for each confidence interval, some sample matching pairs from all sample matching pairs corresponding to the confidence interval, and determine the selected sample matching pairs as target sample matching pairs; The training module is used to train the initial network model based on the calibration information of the target sample matching pair to obtain a target network model; wherein the target network model is used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested.

9. A matching relationship verification device, characterized in that: include: a processor and a machine-readable storage medium storing machine-executable instructions capable of being executed by the processor; The processor is configured to execute machine-executable instructions to implement the following steps: For each sample matching pair in a plurality of sample matching pairs, a collision feature of the sample matching pair is obtained, a fitness value corresponding to the collision feature is determined, and a comprehensive confidence of the sample matching pair is determined based on the fitness value; wherein, the sample matching pair includes a sample object identifier and a sample device identifier, the collision feature represents a feature of the degree of matching between the sample object identifier and the sample device identifier, and the fitness value represents a matching degree value corresponding to the feature of the matching degree; wherein, the object identifier of the object identified from the image, the acquisition time and physical location corresponding to the object identifier are input into the collision model, and the device identifier of the terminal device, the acquisition time and physical location corresponding to the device identifier are input into the collision model, and the collision model is inputted through the collision model. A matching relationship between a sample object identifier and a sample device identifier is obtained, where the sample object identifier and the sample device identifier form a sample matching pair, and the matching relationship indicates that the sample object identifier matches the sample device identifier, and the terminal device corresponding to the sample device identifier is held by the object corresponding to the sample object identifier; wherein the collision feature includes at least one of the following: an initial confidence level, where the initial confidence level is used to verify the matching relationship between the sample object identifier and the sample device identifier; a collision cycle number, where the collision cycle number indicates the number of unit cycles in which the sample object identifier and the sample device identifier collide; and a collision average number, where the collision average number indicates the average number of times the sample object identifier and the sample device identifier collide within a unit cycle; Based on the comprehensive confidence of each sample matching pair in the multiple sample matching pairs, the multiple sample matching pairs are divided into at least one confidence interval; wherein, for the sample matching pairs divided into the confidence interval, the comprehensive confidence of the sample matching pairs is within the confidence interval; For each confidence interval, select some sample matching pairs from all sample matching pairs corresponding to the confidence interval, and determine the selected sample matching pairs as target sample matching pairs; The initial network model is trained based on the calibration information of the target sample matching pair to obtain a target network model, which is used to verify the matching relationship between the identification of the object to be tested and the identification of the device to be tested.

10. A machine-readable storage medium, characterized in that The machine-readable storage medium stores a number of computer instructions. When the computer instructions are executed by the processor, Implement the method steps as described in any one of claims 1 to 7.

Images