Data source quality assessment method, device, equipment and computer-readable storage medium

Abstract

The present application provides a data source quality assessment method, apparatus, device and computer-readable storage medium, which are applied to a first participant device; the method includes: obtaining first feature data held by the first participant device and a second encrypted feature sent by the second participant device, the second encrypted feature being calculated by the second participant device based on the second feature data in the data source to be assessed; training a preset first model based on the first feature data and the second encrypted feature to obtain a trained first model; obtaining test data held by the first participant device, and determining a test result based on the test data and the trained first model; and determining an assessment result of the data source to be assessed based on the test result. Through the present application, the quality of the passive party data source can be assessed through joint modeling, thereby improving the model effect of the joint modeling.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to a data source quality assessment method, device, electronic device, computer-readable storage medium, and computer program product. Background Art

[0002] As industries increasingly strengthen data privacy protection, federated learning—a technology that leverages multi-party data to build machine learning while preserving data privacy—has become a key enabler for collaboration across businesses and industries. Vertical federated learning, when participants have minimal overlap in data features but significant user overlap, leverages the shared user data to perform joint modeling. This improves model performance, enabling participants to provide better services to customers.

[0003] In a vertical federation scenario, when an active party (with labeled data and feature data X1) performs a joint modeling exercise with a passive party (with feature data X2 (X2 is not identical to X1), the quality of the passive party's data affects the effectiveness of the active party's model during the joint modeling exercise. Therefore, assessing the passive party's data quality is a crucial step in the federated modeling process. However, in real-world scenarios, the active party sometimes has less labeled data, resulting in even less labeled data to match the active and passive parties. This limited data makes it impossible to perform data quality assessments, or the assessment results may not reflect the actual data quality, leading to poor joint modeling performance. Summary of the Invention

[0004] The embodiments of the present application provide a data source quality assessment method, device, electronic device, computer-readable storage medium, and computer program product, which can assess the quality of passive party data sources without the need for labeled data, thereby improving the model effect of joint modeling.

[0005] The technical solution of the embodiment of the present application is implemented as follows:

[0006] This embodiment of the present application provides a data source quality assessment method based on a federated learning system. The federated learning system includes a first participant device and at least one second participant device to be assessed. The method is applied to the first participant device and includes:

[0007] Obtaining first feature data held by the first participant device and an encrypted feature sent by the second participant device, where the encrypted feature is obtained by the second participant device encrypting second feature data in the data source to be evaluated;

[0008] Training a preset first model based on the first feature data and the encrypted feature to obtain a trained first model;

[0009] Obtaining test data held by the first participant device, and determining a test result based on the test data and the trained first model;

[0010] Based on the test result, an evaluation result of the data source to be evaluated is determined.

[0011] The present invention provides a data source quality assessment device, comprising:

[0012] a first acquisition module, configured to acquire first feature data held by the first participant device and an encrypted feature sent by the second participant device, wherein the encrypted feature is obtained by encrypting second feature data in the data source to be evaluated by the second participant device;

[0013] A training module, configured to train a preset first model based on the first feature data and the encrypted feature to obtain a trained first model;

[0014] A second acquisition module, configured to acquire test data held by the first participant device;

[0015] A first determining module, configured to determine a test result based on the test data and the trained first model;

[0016] The second determining module is configured to determine an evaluation result of the data source to be evaluated based on the test result.

[0017] In the above solution, the training module is further used to:

[0018] Projecting the first feature data using a preset first model to obtain a projection feature;

[0019] Processing the projected features and the encrypted features using a preset loss function to determine a similarity loss sum corresponding to the first feature data;

[0020] The similarity loss is back-propagated to the preset first model to adjust the parameters of the preset first model to obtain a trained first model.

[0021] In the above solution, the training module is further used to:

[0022] Acquire a preset first model, where the preset first model includes an initial first sub-model and an initial second sub-model;

[0023] Processing the first feature data using the initial first sub-model to obtain a target feature corresponding to the first feature data;

[0024] The target feature is projected using the initial second sub-model to obtain a projection feature.

[0025] In the above solution, the first determining module is further configured to:

[0026] When the first participant device holds the label data, obtaining training data and label data corresponding to the training data;

[0027] Training an initial first classification model based on the training data and label data corresponding to the training data to obtain a trained first classification model, wherein the initial first classification model is constructed based on the preset first model and the preset first classifier;

[0028] Training an initial second classification model based on the training data and label data corresponding to the training data to obtain a trained second classification model, wherein the initial second classification model is constructed based on the trained first model and a preset second classifier;

[0029] A test result is determined based on the test data, the trained first classification model, and the trained second classification model.

[0030] In the above solution, the first determining module is further configured to:

[0031] Obtaining label data corresponding to the test data;

[0032] Testing the trained first classification model based on the test data and label data corresponding to the test data to obtain a reference value;

[0033] Testing the trained second classification model based on the test data and label data corresponding to the test data to obtain a test value;

[0034] The reference value and the test value are compared to obtain a test result.

[0035] In the above solution, the first determining module is further configured to:

[0036] Obtain training data and the number K of label categories, and cluster the training data based on the number K of label categories to obtain K cluster centers;

[0037] Inputting the training data into the trained first model to obtain training features, clustering the training features to obtain N cluster centers; wherein K and N are positive integers;

[0038] A test result is determined based on the test data, the K cluster centers, and the N cluster centers.

[0039] In the above solution, the first determining module is further configured to:

[0040] Determining a reference clustering error between the test data and the K cluster centers;

[0041] Determining a test clustering error between the test data and the N cluster centers;

[0042] The reference clustering error and the test clustering error are compared to obtain a test result.

[0043] In the above solution, the second determining module is further configured to:

[0044] When the test result indicates that the test value is greater than the reference value, or the test result indicates that the reference clustering error is greater than the test clustering error, determining the evaluation result of the data source to be evaluated as allowing access;

[0045] When the test result indicates that the test value is less than or equal to the reference value, or the test result indicates that the reference clustering error is less than or equal to the test clustering error, the evaluation result of the data source to be evaluated is determined to be access denied.

[0046] The present application also provides a data source quality assessment method based on a federated learning system. The federated learning system includes a first participant device and at least one second participant device. The method is applied to the second participant device and includes:

[0047] Acquire second feature data and a preset second model held by the second participant device, where the preset second model includes an initial third sub-model;

[0048] Processing the second feature data using the initial third sub-model to obtain features corresponding to the second feature data;

[0049] encrypting the feature corresponding to the second feature data to obtain an encrypted feature;

[0050] The encrypted feature is sent to the first participant device, so that the first participant device determines an evaluation result of the data source to be evaluated held by the second participant device based on the encrypted feature.

[0051] The present invention also provides a data source quality assessment device, including:

[0052] a third acquisition module, configured to acquire second feature data and a preset second model held by the second participant device, wherein the preset second model includes an initial third sub-model;

[0053] a processing module, configured to process the second feature data using the initial third sub-model to obtain features corresponding to the second feature data;

[0054] an encryption module, configured to encrypt the feature corresponding to the second feature data to obtain an encrypted feature;

[0055] A sending module is configured to send the encryption feature to the first participant device, so that the first participant device determines an evaluation result of the data source to be evaluated held by the second participant device based on the encryption feature.

[0056] An embodiment of the present application provides an electronic device, including:

[0057] a memory for storing executable instructions;

[0058] The processor is configured to implement the data source quality assessment method provided in the embodiment of the present application when executing the executable instructions stored in the memory.

[0059] An embodiment of the present application provides a computer-readable storage medium storing executable instructions for causing a processor to execute and implement the data source quality assessment method provided in the embodiment of the present application.

[0060] An embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the data source quality assessment method provided in the embodiment of the present application.

[0061] The embodiments of the present application have the following beneficial effects:

[0062] The data source quality assessment method provided in this application embodiment, by introducing a label-free joint modeling process, addresses the problem of limited labeled data for overlapping active and passive users in vertical scenarios, making it difficult to accurately assess the quality of passive data sources. This not only expands the scope of data sources, but also improves the accuracy of data source quality assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a schematic diagram of the architecture of the data source quality assessment system provided in an embodiment of the present application;

[0064] Figures 2A-2B is a structural diagram of an electronic device provided in an embodiment of the present application;

[0065] Figure 3 Schematic diagram of the data source quality assessment method provided in the embodiment of the present application;

[0066] Figure 4 A schematic diagram of a process for training a preset first model provided in an embodiment of the present application;

[0067] Figure 5 A schematic diagram of a process for projecting first feature data provided in an embodiment of the present application;

[0068] Figures 6A-6B A schematic diagram of a supervised method evaluation process provided in an embodiment of the present application;

[0069] Figure 7 This is another flow chart of the data source quality assessment method provided in an embodiment of the present application;

[0070] Figure 8 A schematic diagram of the vertical federated learning system architecture provided in an embodiment of the present application;

[0071] Figure 9 Schematic diagram of the joint modeling framework for longitudinal federated learning in unlabeled data scenarios provided in an embodiment of the present application;

[0072] Figures 10A-10B A schematic diagram of a supervised method evaluation process provided in an embodiment of the present application;

[0073] Figures 11A-11B A schematic diagram of the process of evaluating the unsupervised method provided in the embodiments of the present application. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0075] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0076] If similar descriptions of "first / second" appear in the application documents, the following explanation is added. In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific order for the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0077] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0078] Before further explaining the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0079] 1) Federated learning refers to a method of machine learning that unites different participants (participants, or parties, also known as data owners or clients). In federated learning, participants do not need to disclose their data to other participants or the coordinator (also known as a parameter server or aggregation server). Therefore, federated learning can effectively protect user privacy and ensure data security.

[0080] 2) Horizontal federated learning is a method of performing joint machine learning on the data with the same data features but different users when there is a lot of overlap in the data features of the participants but little overlap in the users.

[0081] 3) Vertical federated learning is a method in which the overlap of participants’ data features is small, but the overlap of users is large. The method extracts the users and data with the same participants but different user data features for joint machine learning training.

[0082] The embodiments of the present application provide a data source quality assessment method, device, electronic device, computer-readable storage medium, and computer program product, which can assess the quality of passive party data sources without the need for labeled data, thereby improving the model effect of joint modeling.

[0083] Based on the above explanations of the nouns and terms involved in the embodiments of the present application, the data source quality assessment system provided in the embodiments of the present application is first described. Figure 1 , Figure 1 This is an architectural diagram of a data source quality assessment system provided in an embodiment of the present application. In the data source quality assessment system 100, a first participant device 400 and a second participant device 410 (two second participant devices are shown as examples, respectively denoted as 410-1 and 410-2, for distinction) are interconnected via a network 300, and are also connected to a server device via the network 300. The network 300 may be a wide area network or a local area network, or a combination of the two, and data transmission is achieved using a wireless link.

[0084] In some embodiments, the first participant device 400 and the second participant device 410 are connected to each other through the network 300, and can also be connected to third-party devices (collaborators, servers, etc.) that may be involved in the federated learning system through the network 300.

[0085] In some embodiments, the first participant device 400 and the second participant device 410 can be laptops, tablets, desktop computers, smartphones, dedicated messaging devices, portable gaming devices, smart speakers, smart watches, etc. They can also be client terminals of federated learning participants, such as participant devices of banks or financial institutions that store user feature data, but are not limited to these. The third-party device can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, to assist the participating devices in performing federated learning to obtain a federated learning model. Network 300 can be a wide area network or a local area network, or a combination of the two. The first participant device 400 and the second participant device 410 can be connected directly or indirectly via wired or wireless communication, which is not limited in the embodiments of this application.

[0086] The first participant device 400 is used to obtain the first feature data held by the first participant device 400 and the encrypted feature sent by the second participant device 410, where the encrypted feature is obtained by encrypting the second feature data in the data source to be evaluated by the second participant device 410; then, a preset first model is trained based on the first feature data and the encrypted feature to obtain a trained first model; then, test data held by the first participant device 400 is obtained; finally, a test result is determined based on the test data and the trained first model; and based on the test result, an evaluation result of the data source to be evaluated is determined.

[0087] The second participant device 410 is used to obtain the second feature data held by the second participant device 410 and a preset second model, where the preset second model includes an initial third sub-model; then the second feature data is processed using the initial third sub-model to obtain features corresponding to the second feature data; the features corresponding to the second feature data are then encrypted to obtain encrypted features; and finally, the encrypted features are sent to the first participant device 400, so that the first participant device 400 determines the evaluation result of the data source to be evaluated held by the second participant device 410 based on the encrypted features.

[0088] See also Figures 2A-2B , Figures 2A-2B This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. In practical applications, the electronic device 500 can be implemented as Figure 1 The first participant device 400 or the second participant device 410 in the embodiment is used to illustrate an electronic device that implements the data source quality assessment method of an embodiment of the present application. Figures 2A-2B The electronic device 500 shown includes: at least one processor 510, a memory 550, at least one network interface 520 and a user interface 530. The various components in the electronic device 500 are coupled together via a bus system 540. It is understood that the bus system 540 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 540 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 540 is not shown in FIG. Figures 2A-2B Various buses are labeled as bus system 540 .

[0089] The processor 510 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0090] The user interface 530 includes one or more output devices 531 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 530 also includes one or more input devices 532, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0091] The memory 550 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 550 may optionally include one or more storage devices that are physically remote from the processor 510.

[0092] The memory 550 includes volatile memory or nonvolatile memory, or may include both volatile and nonvolatile memory. The nonvolatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 550 described in the embodiments of the present application is intended to include any suitable type of memory.

[0093] In some embodiments, the memory 550 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplified below.

[0094] Operating system 551, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0095] A network communication module 552 for reaching other computing devices via one or more (wired or wireless) network interfaces 520 , exemplary network interfaces 520 including Bluetooth, WiFi, and USB;

[0096] a presentation module 553 for enabling presentation of information via one or more output devices 531 (e.g., a display screen, a speaker, etc.) associated with the user interface 530 (e.g., a user interface for operating peripheral devices and displaying content and information);

[0097] The input processing module 554 is configured to detect one or more user inputs or interactions from one of the one or more input devices 532 and to translate the detected inputs or interactions.

[0098] In some embodiments, the data source quality assessment device provided in the embodiments of the present application can be implemented in software. Figure 2A The electronic device provided in an embodiment of the present application is a schematic diagram of the structure of the first participant device 400. A data source quality assessment device 555 stored in a memory 550 can be software in the form of a program or plug-in, and includes the following software modules: a first acquisition module 5551, a training module 5552, a second acquisition module 5553, a first determination module 5554, and a second determination module 5555. These modules are logical and can be arbitrarily combined or further separated according to the functions implemented. The functions of each module will be described below.

[0099] In some embodiments, as Figure 2B As shown, Figure 2B The figure shows a schematic diagram of the structure of the electronic device provided in an embodiment of the present application, which is the second participant device 410. The software modules stored in the data source quality assessment device 555 of the memory 550 may include: a third acquisition module 5556, a processing module 5557, an encryption module 5558, and a sending module 5559. These modules are logical and can be arbitrarily combined or further separated according to the functions implemented. The functions of each module will be described below.

[0100] In other embodiments, the data source quality assessment device provided in the embodiments of the present application can be implemented in hardware. As an example, the data source quality assessment device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the data source quality assessment method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0101] The data source quality assessment method provided by the embodiment of the present application will be described in conjunction with the exemplary application and implementation of the first participant device provided by the embodiment of the present application. The data source quality assessment method provided by the embodiment of the present application is based on a federated learning system, wherein the federated learning system includes a first participant device and at least one second participant device. Figure 3 , Figure 3 This is a flow chart of the data source quality assessment method provided in the embodiment of the present application, which will be combined with Figure 3 The steps shown are explained.

[0102] Step S301: Acquire first feature data held by a first participant device and encrypted features sent by a second participant device.

[0103] Here, the encrypted feature is obtained by the second participant device encrypting the second feature data in the data source to be evaluated.

[0104] In actual implementation, in the scenario of a vertical federation model, at least two participants are usually involved. The first participant is the participant who holds feature data and label data (also called the Guest party or active party), and the second participant is the participant who holds feature data (also called the Host party or passive party). The method provided in the embodiment of the present application can be applied to a vertical federation model involving one Guest party and at least one Host party. In this scenario, among the overlapping users held by the first participant and the second participant, the first participant has no labeled data or only a small amount of labeled data, and conventional joint modeling cannot be performed through labeled data. The embodiment of the present application provides a joint modeling method in a scenario with unlabeled data.

[0105] For example, in the embodiment of the present application, the active party A evaluates the data quality of a passive party B. The first participant is equivalent to the active party, and the second participant is equivalent to the passive party. In actual application, the same scheme can be extended to simultaneously evaluate multiple passive party data sources separately or jointly. The first feature data of the first participant is recorded as X A , the second characteristic data of the second participant is recorded as X B The second participant device maps and encrypts the second feature data to obtain the second encrypted feature f B (The second encrypted feature here is the encrypted feature sent by the second participant. To distinguish it from the encrypted feature processed by the first participant's device, the feature encrypted by the first participant's device is called the first encrypted feature, and the feature encrypted by the second participant's device is called the second encrypted feature.) The first participant's device obtains the first feature data X from its own storage space. A , obtain the second encryption feature f from the second participant device B .

[0106] Step S302: training a preset first model based on the first feature data and the encrypted feature to obtain a trained first model.

[0107] Utilize (X A , f B ) Train the preset first model to obtain a trained first model. The preset first model here is a private feature extraction model of the first participant.

[0108] Step S301 and step S302 implement joint modeling based on model pre-training of unlabeled data.

[0109] Figure 4 A schematic diagram of a process for training a preset first model provided in an embodiment of the present application is shown in FIG. Figure 4 As shown, in one implementation, step S302 can be performed by Figure 4 Steps S3021 to S3023 shown are implemented as follows:

[0110] Step S3021: Project the first feature data using a preset first model to obtain a projection feature.

[0111] The first feature data X A Input into the preset first model, and use the first model to analyze the first feature data X A Perform projection processing to obtain the projection feature, recorded as z A .

[0112] Step S3022: Process the projected features and the encrypted features using a preset loss function to determine the sum of similarity losses corresponding to the first feature data.

[0113] This step is the model update process, the purpose of which is to make the features of the same user as close as possible in the feature space and the features of different users as far away as possible in the feature space. In the embodiment of the present application, the classification goal can be achieved by using a preset loss function. Specifically, it can be implemented as follows: Party A can be based on (z A , f B ) Calculate the similarity loss function, the calculation formula is shown in the following formula (1):

[0114] (1);

[0115] Where i is a positive integer with a value range of [1, the number of second encrypted features], fi represents the i-th second encrypted feature; j is a positive integer with a value range of [1, the number of first feature data], and zj represents the j-th first feature data. i and j traverse the samples in the current training pool of Party A and Party B respectively. Fi and zj are the feature data or encrypted features of the same user. That is, if a certain sample i of Party A and a sample j of Party B are aligned, then the calculation method of its overall loss function is as shown in the above formula (1). The similarity loss is calculated for each aligned sample and then summed to obtain the final result, as shown in formula (2):

[0116] (2);

[0117] Step S3023: The similarity loss is back-propagated to the preset first model to adjust the parameters of the preset first model to obtain a trained first model.

[0118] Party A based on (z A , f B ) Calculate its loss function L A,sim , and update the first model by gradient descent method.

[0119] The above method is also applicable to Party B's quality assessment of Party A's data.

[0120] Step S303: Acquire test data held by the first participant's device.

[0121] The test data here is private data held by the first party.

[0122] Step S304: Determine a test result based on the test data and the trained first model.

[0123] Based on the private data of the first participant, the trained first model obtained by the joint modeling is tested to determine whether the joint model is worse than the independent modeling model. If it is better, it indicates that the data source held by the second participant's device can be accessed; if the joint modeling model is not as good as the independent modeling model, it indicates that the data source held by the second participant's device will cause the model effect to deteriorate, and access to the data source held by the second participant's device is denied.

[0124] In the embodiment of the present application, the quality of the data source can be evaluated by supervised and unsupervised methods to determine the test results. The two methods can be used simultaneously to complement each other, or they can be used separately.

[0125] Step S305: Determine the evaluation result of the data source to be evaluated based on the test result.

[0126] Based on the test results, the first participant device determines the evaluation result of the data source to be evaluated, which can be either qualified or unqualified. If the quality is qualified, the first participant device allows access to the second participant device for joint modeling. If the quality is unqualified, to ensure the effectiveness of the joint model, the first participant device does not allow access to the second participant device for joint modeling.

[0127] The data source quality assessment method provided in the embodiment of the present application is based on a federated learning system including a first participant device and at least one second participant device to be assessed. The method is applied to the first participant device and includes: obtaining first feature data held by the first participant device and encrypted features sent by the second participant device, wherein the encrypted features are obtained by encrypting the second feature data in the data source to be assessed by the second participant device; training a preset first model based on the first feature data and the encrypted features to obtain a trained first model; obtaining test data held by the first participant device, and determining a test result based on the test data and the trained first model. Through this method, the quality of the passive party data source can be assessed without the need for labeled data, which can improve the model effect of joint modeling.

[0128] Figure 5 A schematic diagram of a process for projecting the first feature data provided in an embodiment of the present application is shown in FIG. Figure 5 As shown, in some embodiments, the above step S3021 "projecting the first feature data using the preset first model to obtain the projection feature" can be implemented as the following steps:

[0129] Step S30211, obtain the preset first model.

[0130] The preset first model here includes the initial first sub-model and the initial second sub-model, and the initial first sub-model is recorded as MA , the initial second sub-model is recorded as H A .

[0131] Step S30212: Process the first feature data using the initial first sub-model to obtain target features corresponding to the first feature data.

[0132] The first feature data X A Input to the initial first sub-model M A In the first feature data X A Perform feature extraction to obtain the first feature data X A The corresponding target feature f A .

[0133] In the embodiment of the present application, in order to ensure that data privacy is not leaked, the f A Encrypt it, for example, using a differential privacy method.

[0134] Step S30213: Project the target feature using the initial second sub-model to obtain a projected feature.

[0135] The target feature f A Input to the initial first sub-model H A In the example, for the target feature f A Perform projection extraction to obtain the first feature data X A The corresponding projection feature z A .

[0136] In some embodiments, the above step S304 “determining a test result based on the test data and the trained first model” can be implemented as the following steps:

[0137] Step S304a1: When the first participant device holds label data, obtain training data and label data corresponding to the training data.

[0138] Step S304a2: training the initial first classification model based on the training data and the label data corresponding to the training data to obtain a trained first classification model.

[0139] The initial first classification model is constructed based on a preset first model and a preset first classifier.

[0140] Step S304a3: training the initial second classification model based on the training data and the label data corresponding to the training data to obtain a trained second classification model.

[0141] The initial second classification model is constructed based on the trained first model and the preset second classifier.

[0142] Step S304a4: Determine a test result based on the test data, the trained first classification model, and the trained second classification model.

[0143] Usually, Party A has one or more batches of labeled data (X A , Y A As mentioned above, because there is little or no overlap between Party A and Party B's data, effective joint modeling cannot be performed directly using labeled data. Here, Party A can train a model from scratch and then train it based on a pre-trained model based on this labeled data, and then evaluate the final results of both. The detailed implementation of step S3054 is as follows:

[0144] Step S4a41, obtaining label data corresponding to the test data.

[0145] Party A divides the data into (X train , Y train ) and (X test , Y test ).

[0146] Step S4a42: Testing the trained first classification model based on the test data and the label data corresponding to the test data to obtain a reference value.

[0147] Party A in the data (X train , Y train ) to train a randomly initialized model from scratch (M A , P), after training on the test set (X test , Yt est ) to test and get the result S ind In addition, if Party A has a set of benchmark data, for example, it can jointly model with a data source C to obtain model M C，pretrained , in (X train , Y train ) and trained on (X test , Y test ) The test result S baseline , and it has been verified that data source C works better, usually S baseline >S ind (Assuming that the larger the indicator, the better the model effect).

[0148] Step S4a43: Testing the trained second classification model based on the test data and the label data corresponding to the test data to obtain a test value.

[0149] Party A in the data (X train , Y train ) using the pre-trained model M A，pretrained and a randomly initialized classifier P(MA，pretrained ), after training on the test set (X test , Y test ) to test and get the result S fed .

[0150] Step S4a44: compare the reference value and the test value to obtain the test result.

[0151] Compare S ind (or S baseline ) and S fed If S fed >S ind (or S baseline ), it indicates that the data quality of data source B is good and can be considered for access; otherwise, access to this data source is not considered.

[0152] Figures 6A-6B A flow chart of the supervised method evaluation process provided in the embodiment of the present application is as follows: Figure 6A As shown, the random initialization model M is used A and pre-trained model M A，pretrained , in the training set (X train , Y train ) are trained separately to obtain the trained (fine-tuned) models; Figure 6B As shown, in (X test , Y test ) and tested the two trained models respectively to obtain the evaluation index S ind and S fed .

[0153] On this basis, the above-mentioned step S305 "determining the evaluation result of the data source to be evaluated based on the test result" can be implemented as follows: when the test result indicates that the test value is greater than the reference value, the evaluation result of the data source to be evaluated is determined to be allowing access; when the test result indicates that the test value is less than or equal to the reference value, the evaluation result of the data source to be evaluated is determined to be denying access.

[0154] When the first participant device holds labeled data, steps S304a1 to S304a4 can use supervised learning methods to determine the test results. When the first participant device holds only a small amount of labeled data or no labeled data, the first participant device can use the following unsupervised learning method to determine the test results. In one implementation, step S304 "determining the test results based on the test data and the trained first model" can also be implemented as the following steps:

[0155] Step S304b1: Obtain training data and the number K of label categories, and cluster the training data based on the number K of label categories to obtain K cluster centers.

[0156] Step S304b2: input the training data into the trained first model to obtain training features, cluster the training features, and obtain N cluster centers.

[0157] Step S304b3: Determine the test result based on the test data, the K cluster centers, and the N cluster centers.

[0158] This step can be implemented as follows: determining a reference clustering error between the test data and the K cluster centers; determining a test clustering error between the test data and the N cluster centers; and comparing the reference clustering error and the test clustering error to obtain a test result.

[0159] On this basis, the above-mentioned step S305 "determining the evaluation result of the data source to be evaluated based on the test results" can be implemented as follows: when the test result indicates that the reference clustering error is greater than the test clustering error, the evaluation result of the data source to be evaluated is determined to be allowing access; when the test result indicates that the reference clustering error is less than or equal to the test clustering error, the evaluation result of the data source to be evaluated is determined to be denying access.

[0160] Next, the data source quality assessment method provided by the embodiment of the present application will be described in conjunction with the exemplary application and implementation of the second participant device provided by the embodiment of the present application. Figure 7 , Figure 7 This is another flow chart of the data source quality assessment method provided in the embodiment of the present application, which is combined with Figure 7 The steps shown are explained.

[0161] Step S701: Acquire second feature data and a preset second model held by a second participant device.

[0162] Here, the preset second model includes an initial third sub-model, and the initial third sub-model is used to perform feature extraction on the second feature data to obtain a second feature corresponding to the second feature data.

[0163] Step S702: Process the second feature data using the initial third sub-model to obtain features corresponding to the second feature data.

[0164] Step S703: encrypt the feature corresponding to the second feature data to obtain an encrypted feature.

[0165] The encryption feature here is the second encryption feature in the above embodiment.

[0166] In an embodiment of the present application, in order to protect the data privacy of the second participant's device, the second feature can be encrypted to obtain a second encrypted feature.

[0167] Step S704: Send the encrypted feature to the first participant device, so that the first participant device determines the evaluation result of the data source to be evaluated held by the second participant device based on the encrypted feature.

[0168] In some embodiments, the preset second model may further include an initial fourth sub-model for projecting the second encrypted feature to obtain a second projected feature. The second participant device executes the data source quality assessment method provided in the above embodiment, and uses the second participant device to assess the data source quality of the active party or other passive parties through joint modeling, thereby improving the model effect of the joint modeling.

[0169] The following describes an exemplary application of the embodiments of the present application in a practical application scenario.

[0170] Vertical federated learning involves joint machine learning training, where participants have minimal overlap in data features but significant user overlap. This involves extracting the shared user data and data from participants with different data characteristics. For example, consider two participants, A and B, located in the same region: A is a bank, and B is an e-commerce platform. While participants A and B share many of the same users in the same region, their businesses differ, resulting in different user data characteristics. In particular, the user data characteristics recorded by A and B may be complementary. In such scenarios, vertical federated learning can help A and B build a joint machine learning prediction model, enabling them to provide better services to their customers.

[0171] Figure 8 The vertical federated learning system architecture diagram provided in this application embodiment requires the participation of coordinator C in order to help A and B jointly model. Part I: Participants A and B implement encrypted sample alignment, such as Figure 8 As an example, since the user groups of companies A and B do not completely overlap, the system uses encryption-based user sample alignment technology to identify common users between the two parties without revealing their respective data. This technology also prevents the exposure of non-overlapping users, allowing the combined features of these users to be used for modeling.

[0172] Part 2: Encrypted Model Training. After determining the shared user group, the data can be used to train the machine learning model. To ensure the confidentiality of the data during training, encrypted training is required with the help of a coordinator C. Taking the linear regression model as an example, the training process can be divided into the following four steps. Step 1: Coordinator C distributes the public key to A and B to encrypt the data that needs to be exchanged during training. Step 2: Participants A and B exchange intermediate results for calculating gradients in encrypted form. Step 3: Participants A and B each perform calculations based on the encrypted gradient values, while participant B calculates the loss function based on its labeled data and summarizes the results to coordinator C. Coordinator C calculates the total gradient value by summarizing the results and decrypts it. Step 4: Coordinator C transmits the decrypted gradient back to participants A and B respectively, and participants A and B update the parameters of their respective models based on the gradients.

[0173] Participants and coordinators iterate the above steps until the loss function converges or the model parameters converge or the maximum number of iterations or the maximum training time is reached, thus completing the entire model training process.

[0174] It should be noted that in horizontal federated learning and vertical federated learning, encryption operations and encrypted transmission are optional and need to be determined based on the specific application scenario. Not all application scenarios require encryption operations and encrypted transmission.

[0175] As mentioned above, in a vertical federation scenario, the active party, which possesses labeled data Y and feature data X1, conducts joint modeling with the passive party, which possesses feature data X2 (which is not identical to X1), to improve the performance of its own model. The data quality of the passive party (X2) determines the degree of improvement in the active party's model performance during joint modeling. Therefore, assessing the passive party's data quality is a key step in the federated modeling process. Data quality includes various evaluation metrics, such as user overlap and improvement in modeling performance relative to baseline data. Improvement in modeling performance is typically assessed based on one or more sets of labeled data jointly modeled with the passive party. If the model performance improves significantly, the passive party's data quality is considered high. However, in real-world scenarios, the active party sometimes has less labeled data, and the amount of labeled data that can be matched between the active and passive parties is even less. This limited data size makes data quality assessment impossible, or the assessment results may not reflect the actual data quality. Therefore, assessing the passive party's data quality through joint modeling based on a large amount of unlabeled data can expand the range of accessible data sources and improve the accuracy of data assessment.

[0176] The embodiment of the present application proposes a data source quality assessment method in a vertical federated learning scenario, which mainly consists of two parts. The first step is to pre-train the model based on unlabeled data, and the second step is to evaluate the modeling effect based on the pre-trained model. The two steps are explained below. This scheme only describes the scheme in which the active party A evaluates the data quality of a passive party B. Based on the same scheme, it can be expanded to simultaneously evaluate multiple passive party data sources separately and jointly.

[0177] In this scenario, among the overlapping users of party A and party B, party A has no labeled data or only a small amount of labeled data, and conventional joint modeling cannot be performed using labeled data.

[0178] The first step is joint modeling based on unlabeled data.

[0179] Figure 9 Schematic diagram of the joint modeling framework for longitudinal federated learning of unlabeled data scenarios provided in the embodiment of this application, such as Figure 9 As shown, Party A and Party B have their own Model A (M A ) and Model B (M B ) output features f respectively A and f B In addition, Party A has Head A , Party B has Head B Act on f respectively A and f B , generating the projection z A and z B .

[0180] Basic training process:

[0181] Initialization: Each participant initializes their own Model A, Model B, Head A, and Head B. Party A and Party B align their data using encryption.

[0182] 1. Party A and Party B obtain the data feature f through their respective models Model A and Model B A and f B , respectively through HeadA and HeadB to get z A and z B . Party B will f B Pass it to Party A, Party A will f A In addition, f can be encrypted in a variety of ways. A , f B , further increase the security of data, such as differential privacy methods.

[0183] 2. This step is the model update process, the goal of which is to make the features of the same user as close as possible in the feature space and the features of different users as far apart as possible. There are many ways to achieve this goal, such as by selecting a specific loss function. A typical method is as follows. Party A and Party B are based on (z A , f B ) and (z B , f A ) to calculate the similarity loss function.

[0184] For the two aligned samples fi and zj, the specific form of the loss function is as follows (3):

[0185] (3);

[0186] Here, i and j respectively traverse the samples in the current training batch of party A and party B. For a sample i of party A, which is aligned with sample j of party B, the overall loss function is calculated as shown above. The similarity loss is calculated for each aligned sample and then summed to obtain the final result:

[0187] (4).

[0188] 3. Party A and Party B respectively according to (z A , f B ) and (z B , f A ) Calculate the respective loss function L A,sim and L B,sim , and update their respective models (Model A, HeadA) and (Model B, HeadB) by gradient descent method.

[0189] In the second step, the active party conducts data source quality assessment based on the pre-trained model.

[0190] The quality of data sources can be evaluated by supervised and unsupervised methods. The two methods can be used simultaneously to complement each other or separately. Through the first step of joint modeling based on unlabeled data, Party A obtains a pre-trained model Model A (M A ), denoted as M A，pretrained In M A A classifier P can be added to form a complete classification model (M A , P).

[0191] 1. Supervised methods

[0192] Usually, Party A has one or more batches of labeled data (X A ,Y AAs mentioned above, since this data has little or no overlap with Party B, effective joint modeling cannot be performed directly using labeled data. Here, Party A can train a model from scratch and then train it based on a pre-trained model based on this labeled data, and then evaluate the final results of both. The detailed process is as follows:

[0193] Party A divides the data into (X train ,Y train ) and (X test ,Y test ).

[0194] a) Party A has data (X train ,Y train ) to train a randomly initialized model from scratch (M A , P), after training on the test set (X test ,Y test) Test and get result S ind In addition, if Party A has a set of benchmark data, for example, it can jointly model with a data source C to obtain model M C，pretrained , in (X train ,Y train ) and trained on (X test ,Y test ) The test result S baseline , and it has been verified that data source C works better, usually S baseline >S ind (Assuming that the larger the indicator, the better the model effect).

[0195] b) Party A has data (X train ,Y train ) using the pre-trained model M A，pretrained and a randomly initialized classifier P, (M A，pretrained ), after training on the test set (X test ,Y test ) to test and get the result S fed ;

[0196] c) Compare S ind (or S baseline ) and S fed If S fed >S ind (or S baseline ), it indicates that the data quality of data source B is good and can be considered for access; otherwise, access to this data source is not considered.

[0197] Figures 10A-10B For a flow chart of the supervised method evaluation provided in the embodiment of this application, see Figure 10A , using random initialization (MA ) and pre-trained models (M A，pretrained ) in the training set (X train ,Y train ) are trained separately to obtain the trained (fine-tuned) models; Figure 10B As shown, in (X test ,Y test ) and tested the two trained models respectively to obtain the evaluation index S ind and S fed .

[0198] 2. Unsupervised methods

[0199] Unsupervised methods can serve as a supplement to supervised methods; they can also be used alone. For example, in some scenarios, there is very little labeled data, and the evaluation indicators based on supervised methods have low discrimination, making it difficult to evaluate the quality of the data source.

[0200] The unsupervised method process is as follows:

[0201] Party A divides the data into (X traun ) and (X test ).

[0202] a) Party A is the label holder and knows the number of label categories K. An unsupervised clustering method, such as K-means, is used to cluster the labels in X. train Train the model on X to get K cluster centers; optionally, first adopt a general representation learning method, such as AutoEncoder, train Learn a representation model H, and then represent the output of the model M autoencoder (X train ) to perform unsupervised clustering training (such as K-means) and obtain K cluster centers on the test set (X test )(or M autoencoder (X test )) Test clustering error E ind In addition, if Party A has passed the modeling with the other Party C, the test result E can be obtained. baseline , and the effect of data source C has been verified. Common error calculation indicators such as mean squared error (MSE) are first obtained based on the distance from each cluster center to its category, then the distance from the corresponding cluster center is calculated, and finally the distance calculated for all samples is averaged.

[0203] b) Feature M of Party A in the pre-trained model C，pretrained (X train ) for unsupervised clustering training, and then C，pretrained (X test ) to calculate the clustering error E fed.

[0204] c) Compare E fed and E ind (E baseline ). If E fed <E ind (E baseline ), it means that data source B improves the regularity of data A and can be considered for connection to the data source; otherwise, it is not considered for connection to the data source.

[0205] Figures 11A-11B For a flow chart of the unsupervised method evaluation provided in the embodiment of the present application, see Figure 11A , on the training set X train First, train the representation learning model, then perform unsupervised clustering on its output to obtain the cluster center; calculate the clustering error E on the test set; or directly on the training set X train Clustering and testing are performed on (dashed line). Figure 11B As shown in Figure 1, the output of the training set on the pre-training model is first clustered to obtain the cluster center, and then the test set is used to calculate the clustering error E through the output of the pre-training model.

[0206] Based on the above method, after modeling and evaluating multiple data sources and conducting actual usage tests, a correspondence between actual usage effects and corresponding S / E indicators can be established. This can serve as a reference standard for testing new data sources, further improving the accuracy of the data source quality assessment of the solution.

[0207] The embodiment of the present application solves the problem of insufficient labeled data among overlapping users of the active and passive parties in vertical scenarios, making it difficult to accurately assess the quality of the passive party's data source, by introducing an unlabeled joint modeling process. On the one hand, it can expand the source range of data sources, and on the other hand, it can improve the accuracy of data source quality assessment. The embodiment of the present application proposes a method for assessing data source quality in a vertical federation scenario without the need for labeled data; because it is not limited to using labeled data to assess data source quality, it expands the range of testable and adopted data sources, while improving the accuracy of data source quality assessment in scenarios with less labeled data.

[0208] The following continues to describe the exemplary structure of the data source quality assessment device 555 provided in the embodiment of the present application implemented as a software module. In some embodiments, such as Figure 2A As shown, Figure 2A FIG. 4 is a schematic diagram of the structure of a first participant device 400 provided in an embodiment of the present application. The software modules stored in the data source quality assessment device 555 of the memory 550 may include:

[0209] A first acquisition module 5551 is configured to acquire first feature data held by the first participant device and an encrypted feature sent by the second participant device, wherein the encrypted feature is obtained by encrypting the second feature data in the data source to be evaluated by the second participant device;

[0210] A training module 5552 is configured to train a preset first model based on the first feature data and the encrypted feature to obtain a trained first model;

[0211] A second acquisition module 5553 is configured to acquire the test data held by the first participant device;

[0212] A first determination module 5554 is configured to determine a test result based on the test data and the trained first model;

[0213] The second determining module 5555 is configured to determine an evaluation result of the data source to be evaluated based on the test result.

[0214] In some embodiments, the training module is further configured to:

[0215] Projecting the first feature data using a preset first model to obtain a projection feature;

[0216] Processing the projected features and the encrypted features using a preset loss function to determine a similarity loss sum corresponding to the first feature data;

[0217] The similarity loss is back-propagated to the preset first model to adjust the parameters of the preset first model to obtain a trained first model.

[0218] In some embodiments, the training module is further configured to:

[0219] Acquire a preset first model, where the preset first model includes an initial first sub-model and an initial second sub-model;

[0220] Processing the first feature data using the initial first sub-model to obtain a target feature corresponding to the first feature data;

[0221] The target feature is projected using the initial second sub-model to obtain a projection feature.

[0222] In some embodiments, the first determining module is further configured to:

[0223] When the first participant device holds the label data, obtaining training data and label data corresponding to the training data;

[0224] Training an initial first classification model based on the training data and label data corresponding to the training data to obtain a trained first classification model, wherein the initial first classification model is constructed based on the preset first model and the preset first classifier;

[0225] Training an initial second classification model based on the training data and label data corresponding to the training data to obtain a trained second classification model, wherein the initial second classification model is constructed based on the trained first model and a preset second classifier;

[0226] A test result is determined based on the test data, the trained first classification model, and the trained second classification model.

[0227] In some embodiments, the first determining module is further configured to:

[0228] Obtaining label data corresponding to the test data;

[0229] Testing the trained first classification model based on the test data and label data corresponding to the test data to obtain a reference value;

[0230] Testing the trained second classification model based on the test data and label data corresponding to the test data to obtain a test value;

[0231] The reference value and the test value are compared to obtain a test result.

[0232] In some embodiments, the first determining module is further configured to:

[0233] Obtain training data and the number K of label categories, and cluster the training data based on the number K of label categories to obtain K cluster centers;

[0234] Inputting the training data into the trained first model to obtain training features, clustering the training features to obtain N cluster centers;

[0235] A test result is determined based on the test data, the K cluster centers, and the N cluster centers.

[0236] In some embodiments, the first determining module is further configured to:

[0237] Determining a reference clustering error between the test data and the K cluster centers;

[0238] Determining a test clustering error between the test data and the N cluster centers;

[0239] The reference clustering error and the test clustering error are compared to obtain a test result.

[0240] In some embodiments, the second determining module is further configured to:

[0241] When the test result indicates that the test value is greater than the reference value, or the test result indicates that the reference clustering error is greater than the test clustering error, determining the evaluation result of the data source to be evaluated as allowing access;

[0242] When the test result indicates that the test value is less than or equal to the reference value, or the test result indicates that the reference clustering error is less than or equal to the test clustering error, the evaluation result of the data source to be evaluated is determined to be access denied.

[0243] In some embodiments, as Figure 2B As shown, Figure 2B FIG. 4 is a schematic diagram of the structure of the second participant device 410 provided in an embodiment of the present application. The software modules stored in the data source quality assessment device 555 of the memory 550 may include:

[0244] A third acquisition module 5556 is configured to acquire second feature data and a preset second model held by the second participant device, wherein the preset second model includes an initial third sub-model;

[0245] a processing module 5557, configured to process the second feature data using the initial third sub-model to obtain features corresponding to the second feature data;

[0246] An encryption module 5558 is configured to encrypt the feature corresponding to the second feature data to obtain an encrypted feature;

[0247] The sending module 5559 is configured to send the encryption feature to the first participant device, so that the first participant device determines the evaluation result of the data source to be evaluated held by the second participant device based on the encryption feature.

[0248] It should be noted that the description of the device of the embodiment of the present application is similar to the description of the above-mentioned method embodiment, and has similar beneficial effects as the method embodiment, so it will not be repeated.

[0249] An embodiment of the present application provides a computer program product, including a computer program, characterized in that when the computer program is executed by a processor, the data source quality assessment method provided in the embodiment of the present application is implemented.

[0250] The embodiment of the present application provides a computer-readable storage medium storing executable instructions, wherein the executable instructions are stored. When the executable instructions are executed by a processor, the processor will execute the method provided by the embodiment of the present application, for example, Figure 3 The data source quality assessment method is shown.

[0251] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or various devices including one or any combination of the above memories.

[0252] In some embodiments, executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0253] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0254] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0255] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. A data source quality assessment method, characterized in that: Based on a federated learning system, the federated learning system includes a first participant device and at least one second participant device to be evaluated. The method is applied to the first participant device, and the method includes: Obtaining first feature data held by the first participant device and an encrypted feature sent by the second participant device, where the encrypted feature is obtained by the second participant device encrypting second feature data in the data source to be evaluated; Projecting the first feature data using a preset first model to obtain a projection feature; Processing the projected features and the encrypted features using a preset loss function to determine a similarity loss sum corresponding to the first feature data; Back-propagating the similarity loss to the preset first model to adjust the parameters of the preset first model to obtain a trained first model; Obtaining test data held by the first participant device, and determining a test result based on the test data and the trained first model; Determining an evaluation result of the data source to be evaluated based on the test result; The determining of a test result based on the test data and the trained first model includes: When the first participant device holds the label data, obtaining training data and label data corresponding to the training data; Training an initial first classification model based on the training data and label data corresponding to the training data to obtain a trained first classification model, wherein the initial first classification model is constructed based on the preset first model and the preset first classifier; Training an initial second classification model based on the training data and label data corresponding to the training data to obtain a trained second classification model, wherein the initial second classification model is constructed based on the trained first model and a preset second classifier; Determining a test result based on the test data, the trained first classification model, and the trained second classification model; Alternatively, determining a test result based on the test data and the trained first model includes: Obtain training data and the number K of label categories, and cluster the training data based on the number K of label categories to obtain K cluster centers; Inputting the training data into the trained first model to obtain training features, clustering the training features to obtain N cluster centers; wherein K and N are positive integers; A test result is determined based on the test data, the K cluster centers, and the N cluster centers.

2. The method according to claim 1, characterized in that The projecting process of the first feature data using a preset first model to obtain a projection feature includes: Acquire a preset first model, where the preset first model includes an initial first sub-model and an initial second sub-model; Processing the first feature data using the initial first sub-model to obtain a target feature corresponding to the first feature data; The target feature is projected using the initial second sub-model to obtain a projection feature.

3. The method according to claim 1, characterized in that The determining of a test result based on the test data, the trained first classification model, and the trained second classification model includes: Obtaining label data corresponding to the test data; Testing the trained first classification model based on the test data and label data corresponding to the test data to obtain a reference value; Testing the trained second classification model based on the test data and label data corresponding to the test data to obtain a test value; The reference value and the test value are compared to obtain a test result.

4. The method according to claim 1, wherein The determining of a test result based on the test data, the K cluster centers, and the N cluster centers includes: Determining a reference clustering error between the test data and the K cluster centers; Determining a test clustering error between the test data and the N cluster centers; The reference clustering error and the test clustering error are compared to obtain a test result.

5. The method according to claim 3 or 4, characterized in that Determining an evaluation result of the data source to be evaluated based on the test result includes: When the test result indicates that the test value is greater than the reference value, or the test result indicates that the reference clustering error is greater than the test clustering error, determining the evaluation result of the data source to be evaluated as allowing access; When the test result indicates that the test value is less than or equal to the reference value, or the test result indicates that the reference clustering error is less than or equal to the test clustering error, the evaluation result of the data source to be evaluated is determined to be access denied.

6. A data source quality assessment device, characterized in that: include: A first acquisition module is configured to acquire first feature data held by a first participant device and an encrypted feature sent by a second participant device, wherein the encrypted feature is obtained by encrypting second feature data in a data source to be evaluated by the second participant device; a training module, configured to train a preset first model based on the first feature data and the encrypted feature to obtain a trained first model, and further configured to project the first feature data using the preset first model to obtain a projected feature; Processing the projected features and the encrypted features using a preset loss function to determine a similarity loss sum corresponding to the first feature data; backpropagating the similarity loss sum to the preset first model to adjust parameters of the preset first model to obtain a trained first model; A second acquisition module, configured to acquire test data held by the first participant device; a first determination module configured to determine a test result based on the test data and the trained first model, the first determination module further configured to, when the first participant device holds label data, obtain training data and label data corresponding to the training data; train an initial first classification model based on the training data and the label data corresponding to the training data to obtain a trained first classification model, the initial first classification model being constructed based on the preset first model and the preset first classifier; Training an initial second classification model based on the training data and label data corresponding to the training data to obtain a trained second classification model, wherein the initial second classification model is constructed based on the trained first model and a preset second classifier; and determining a test result based on the test data, the trained first classification model, and the trained second classification model; The first determination module is further configured to obtain training data and a number K of label categories, and cluster the training data based on the number K of label categories to obtain K cluster centers; input the training data into the trained first model to obtain training features, cluster the training features to obtain N cluster centers, where K and N are positive integers; and determine a test result based on the test data, the K cluster centers, and the N cluster centers; The second determining module is configured to determine an evaluation result of the data source to be evaluated based on the test result.

7. An electronic device, characterized in that: include: a memory for storing executable instructions; The processor is configured to implement the data source quality assessment method according to any one of claims 1 to 5 when executing the executable instructions stored in the memory.

8. A computer-readable storage medium, characterized in that Executable instructions are stored, which are used to implement the data source quality assessment method described in any one of claims 1 to 5 when executed by a processor.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the data source quality assessment method according to any one of claims 1 to 5 is implemented.

Images