A data processing method and device, computer equipment and storage medium

Abstract

Embodiments of the present application provide a data processing method and device, computer equipment and a storage medium, which are applied to the field of artificial intelligence technology. The method specifically comprises: obtaining a training sample set, the training sample set comprising a plurality of training samples; obtaining feature data of each training sample in the plurality of training samples, the feature data comprising initial label data and a labeling time of the initial label data; and performing adjustment processing on the initial label data of each training sample according to the labeling time of the initial label data of the plurality of training samples to obtain adjusted target label data of each training sample, the adjusted target label data being used for training of a business model. Through the present solution, the label can be smoothed by using the time factor to obtain the adjusted target label data, and the business model is trained by using the adjusted target label data, thereby improving the accuracy of the recognition effect of the business model.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a data processing method, apparatus, computer equipment, and storage medium. Background Technology

[0002] With the rapid development of mobile internet and artificial intelligence technologies, the training of business models is increasingly being applied in the field of artificial intelligence and various other sectors. Business models can be specifically categorized as security models, service processing models, exchange models, credit models, and other related models. These models typically require training with sample data and labels before they can be deployed.

[0003] In existing technologies, the label changes over time are not fully considered, resulting in insufficient label smoothness, which leads to poor training performance of the business model and affects its recognition performance. Summary of the Invention

[0004] This application proposes a data processing method, apparatus, computer equipment, and storage medium that can smooth labels using time factors to obtain adjusted target label data, and use the adjusted target label data to train a business model, thereby improving the accuracy of the business model's recognition effect.

[0005] One embodiment of this application provides a data processing method, including:

[0006] Obtain a training sample set, which includes multiple training samples;

[0007] Obtain feature data for each training sample from the plurality of training samples, wherein the feature data includes initial label data and the labeling time of the initial label data;

[0008] Based on the annotation time of the initial label data of the multiple training samples, the initial label data of each training sample is adjusted to obtain the adjusted target label data of each training sample. The adjusted target label data is used for training the business model.

[0009] One embodiment of this application provides a data processing apparatus, including:

[0010] An acquisition unit is used to acquire a training sample set, wherein the training sample set includes multiple training samples;

[0011] The acquisition unit is further configured to acquire feature data of each training sample among the plurality of training samples, the feature data including initial label data and the labeling time of the initial label data;

[0012] The processing unit is configured to adjust the initial label data of each training sample according to the labeling time of the initial label data of the plurality of training samples, so as to obtain the adjusted target label data of each training sample, and the adjusted target label data is used for training the business model.

[0013] One aspect of this application provides a computer device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the methods described in the above embodiments.

[0014] One aspect of this application provides a computer storage medium storing a computer program, which includes program instructions. When the program instructions are executed by a processor, they perform the methods described in the above embodiments.

[0015] One aspect of this application provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. When the computer instructions are executed by the processor of a terminal device, the methods described in the above embodiments are performed.

[0016] The data processing method provided in this application allows a computer device to acquire a training sample set including multiple training samples and feature data for each training sample. The feature data includes initial label data and the annotation time of the initial label data. Then, the computing device adjusts the initial label data of each training sample based on the annotation time of the initial label data of the multiple training samples to obtain adjusted target label data for each training sample. The adjusted target label data is used for training a business model. Since the target label data for each training sample is determined after adjustment based on the annotation time of each training sample, the business model trained using the adjusted target label data considers the label changes caused by time factors, thereby improving the accuracy of the business model's recognition performance. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the architecture of a data processing system provided in an embodiment of this application;

[0019] Figure 2This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0020] Figure 3 This is a schematic diagram of a data cleaning process provided in an embodiment of this application;

[0021] Figure 4 This is a schematic flowchart of a model training method provided in an embodiment of this application;

[0022] Figure 5 This is a flowchart illustrating another data processing method provided in an embodiment of this application;

[0023] Figure 6 This is a schematic diagram of the model structure of a neural network model provided in an embodiment of this application;

[0024] Figure 7 This is a flowchart illustrating another model training method provided in an embodiment of this application;

[0025] Figure 8 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0026] Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0028] It should be noted that the terms "first," "second," etc., used in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature specified with "first" or "second" may explicitly or implicitly include at least one of those features.

[0029] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0030] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0031] The solutions provided in this application belong to the deep learning technology under the field of artificial intelligence.

[0032] Deep Learning (DL) is an interdisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0033] In this application, a computer device can acquire a training sample set including multiple training samples and feature data for each training sample. The feature data includes initial label data and the annotation time of the initial label data. Then, the computer device adjusts the initial label data of each training sample based on the annotation time of the initial label data of the multiple training samples to obtain adjusted target label data for each training sample. Subsequently, the adjusted target label data can be used to train a business model using deep learning technology.

[0034] This application can be applied to the following scenarios: This solution can be used to obtain adjusted target label data corresponding to training samples, and then train a business model using the adjusted target label data. In a business risk prediction scenario, when a business prediction request carrying business data is received, the business model is invoked to identify and process the business data, obtaining the risk prediction probability corresponding to the business data. This risk prediction probability can be used for risk prediction. Specifically, if the risk prediction probability is greater than or equal to a reference probability threshold, the business data is considered to pose a business risk in that business scenario. Therefore, when the business data is actually used online, after detecting the business data, online risk control can be performed on the business data in real time, such as intercepting or filtering the business data, thereby achieving the effect of risk control.

[0035] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the architecture of a data processing system provided in an embodiment of this application. The system architecture includes a server 140 and a computer device cluster, wherein the computer device cluster may include computer device 110, computer device 120, computer device 130, etc. The computer device cluster and the server 140 can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0036] Figure 1 The server 140 shown can be a standalone 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 communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0037] Figure 1 The computer devices 110, 120, and 130 shown can be mobile phones, tablets, laptops, PDAs, mobile internet devices (MIDs), vehicles, in-vehicle devices, roadside devices, aircraft, wearable devices such as smartwatches, smart bracelets, pedometers, and other smart devices with data processing functions.

[0038] In one possible implementation, taking computer device 110 as an example, computer device 110 acquires a training sample set, which includes multiple training samples. Then, computer device 110 acquires feature data for each training sample, including initial label data and the annotation time of the initial label data. Next, computer device 110 sends the training sample set and the feature data of each training sample to server 140. Server 140 adjusts the initial label data of each training sample based on the annotation time of the initial label data of the multiple training samples, obtaining adjusted target label data for each training sample. Then, server 140 sends the adjusted target label data of each training sample back to computer device 110. Subsequently, computer device 110 can train a business model based on the adjusted target label data.

[0039] Of course, based on the annotation time of the initial label data of multiple training samples, the initial label data of each training sample is adjusted to obtain the adjusted target label data for each training sample. This process does not necessarily have to be performed by server 140; it can also be performed by computer device 110 or any other computer device in the computer device cluster. Furthermore, training the business model based on the adjusted target label data does not necessarily have to be performed by computer device 110; it can also be performed by server 140.

[0040] It is understood that the system architecture diagrams described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0041] Please see Figure 2 , Figure 2 This is a flowchart illustrating a data processing method provided in an embodiment of this application. The method is applied to a computer device, which may be, for example, a smartphone, tablet, smart wearable device, in-vehicle device, or other intelligent device. The computer device can also be a server working collaboratively with a server. Figure 2 As shown, the data processing method may include steps S210 to S230. Wherein:

[0042] Step S210: Obtain a training sample set, which includes multiple training samples.

[0043] In practice, the training sample set can include multiple training samples, each of which can include both raw data and text data. The raw data can be numerical data, and the text data can be text-based data.

[0044] In one possible implementation, the detailed process of a computer device acquiring a training sample set is as follows: First, raw data is extracted from the underlying data platform. Raw data may include, for example, transaction data and security data. Specifically, transaction data can be data extracted from a payment platform, such as an information exchange center payment platform, a shopping transaction payment platform, etc. Security data can be data extracted from a security model; for example, security data could be data generated or uploaded by user A in the security model's usage scenario, or data exchanged between user A and user B in the security model's usage scenario. Furthermore, each piece of data, whether transaction data or security data, can correspond to an account. Of course, in addition to transaction data and security data, raw data can also include business processing data, including but not limited to exchange-related data generated in exchange models and credit-related data generated in credit models. In this application, the raw data extracted from the underlying data platform only needs to include data affected by time factors; there is no limitation on the specific business scenarios from which the raw data originates. Then, multi-table matching is performed based on the ID (each ID corresponds to one account), the original data is merged, and then the original data is aggregated according to a specific dimension (feature engineering). Feature engineering is to create a new X' from X. The basic operations include derivation (dimensionality increase) and filtering (dimensionality reduction).

[0045] In this context, multi-table matching by ID means: for example, if 10,000 samples are needed, and the total number of users is 10e, with features distributed across different tables 'a' and 'b', then first, the features of the 10,000 accounts need to be matched in table 'a', and then the features of the 10,000 accounts need to be matched in table 'b'. Then, the different features matched from tables 'a' and 'b' are merged into a single table by ID. Features corresponding to the same account are merged into the same table. Finally, principles such as missing rate and uniqueness can be used to clean the raw data before storing it at the computing platform layer.

[0046] Specifically, some steps of data cleaning, such as Figure 3 As shown, Figure 3This is a schematic diagram of a data cleaning process provided in an embodiment of this application. The sample layer stores training samples extracted from the underlying platform, including raw data and text data. Specifically, the data cleaning method for the training samples in the sample layer can include filtering or deleting training samples with a missing rate greater than 90%, etc. Furthermore, feature data obtained after feature engineering based on the training samples is stored in the feature layer. Specifically, the data cleaning method for the feature data in the feature layer can include: 1. deleting feature data with a missing rate greater than 90%; 2. deleting feature data with a uniqueness rate greater than 90%; 3. deleting feature data with an IV (Information Value) less than 0.02, etc.

[0047] It's important to note that the IV value is primarily used to encode and assess the predictive power of input variables. The magnitude of the IV value of a feature variable indicates the strength of its predictive power. The IV value ranges from [0, positive infinity]. Next, we will explain the calculation of the IV value in detail. Since the IV value is calculated based on the Weight of Evidence (WOE), we will first introduce WOE. WOE is an encoding form of the original variable. To encode a variable using WOE, the variable first needs to be grouped, i.e., binned or discretized. Common discretization methods include equal-width grouping, equal-height grouping, or grouping using decision trees. WOE represents the difference between "the proportion of responding customers in the current group" and "the proportion of non-responding customers in the current group." The IV measures the information content of a variable, essentially a weighted sum of the WOE values ​​of the independent variables. Its magnitude determines the degree of influence of the independent variable on the target variable.

[0048] This solution cleans the raw data extracted from the underlying platform, which involves re-examining and verifying the data, such as filling in missing data, deleting duplicate data, and correcting erroneous data. This ensures the accuracy, consistency, and integrity of the data, which requires cleaning the underlying data.

[0049] Then, text data is extracted from the underlying data platform, merged by ID, and redundant text such as spaces and punctuation marks are removed. Of course, the text data obtained from the underlying platform can also be cleaned. For details on the specific steps of data cleaning, please refer to the aforementioned steps for cleaning raw data. Then, a lookup table is used to convert the text into an n-dimensional 0-1 vector, multiple vectors are merged into a feature matrix, and then the feature matrix is ​​stored in the computing platform layer. The specific conversion method is shown in Table 1 below:

[0050] Taking a 10-dimensional dictionary as an example, we can convert "Who am I" into a 10-dimensional 01 vector.

[0051] Table 1. Data Transformation of Text Data

[0052]

[0053] As shown in Table 1 above, the 10-dimensional 01 vector "0101000010" is obtained after transforming the text data "Who am I" using a 10-dimensional dictionary. It should be noted that different dimensional dictionaries can be manually selected or set for different business scenarios. For example, in business scenarios with large data demands, a larger-dimensional dictionary can be selected, and the text data can be vectorized based on this dictionary. This ensures that the dictionary covers as much of the text data as possible, improving the accuracy of the data transformation. Conversely, in business scenarios with smaller data demands, a smaller-dimensional dictionary can be selected, and the text data can be vectorized based on the larger-dimensional dictionary. Because the data demand is small, the resulting 01 vector has a smaller dimension, thus saving storage space on computer equipment and improving processing efficiency for subsequent data calculations. In addition to selecting an appropriate dictionary based on specific business scenarios, a general dictionary can also be used for matching and data transformation. This means using a default general dictionary with sufficiently broad dimensions to cover as much of the possible text data as possible. For any business scenario, this general dictionary can be directly selected for data transformation, eliminating the need to set up separate dictionaries for different scenarios, thus saving time and labor costs. Of course, in actual business scenarios, specific situations can be analyzed to select a dictionary with appropriate dimensions, as long as it can accurately and effectively transform the text data. This application does not impose specific limitations in this regard.

[0054] Step S220: Obtain feature data for each training sample among the plurality of training samples, wherein the feature data includes initial label data and the annotation time of the initial label data.

[0055] In practice, each training sample includes feature data, which includes the initial label data and the annotation time of the initial label data. The initial label data specifically refers to numerical labels, and in binary or multi-class classification scenarios, the initial label data is either 0 or 1. The annotation time refers to the time corresponding to the annotation of the initial label data. The expression for the initial label data is shown in formula (1):

[0056] Formula (1)

[0057] In formula (1), i represents the category to which a training sample belongs, and y is one of the categories. That is, if a training sample belongs to category y, its initial label data is 1; if it does not belong to category y, its initial label data is 0. In addition to being applicable to binary classification scenarios, it is also applicable to multi-class classification scenarios.

[0058] The annotation time can be in days, such as January 1, 2021 for the initial label data of training sample 1; it can also be in hours, such as 12:00 noon on January 1, 2021; it can be in minutes, such as 12:30 noon on January 1, 2021; or it can be in months, such as January 2020. For different business scenarios, appropriate time units (including months, days, hours, minutes, etc.) can be selected to obtain the corresponding annotation time. Specifically, if the label data is significantly affected by time in a certain business scenario, a more precise time unit, such as days, can be chosen; if the label data is not significantly affected by time in another business scenario, a coarser time unit, such as months, can be chosen.

[0059] Step S230: Based on the annotation time of the initial label data of the multiple training samples, adjust the initial label data of each training sample to obtain the adjusted target label data of each training sample.

[0060] In practice, adjusting the initial label data for each training sample can include label smoothing. Specifically, this label smoothing in this solution primarily targets initial label data affected by time factors, such as gambling labels or credit labels in a risk control scenario. Finally, the adjusted target label data for multiple training samples obtained after label smoothing the original label data for each training sample can be used to train the business model.

[0061] In one possible implementation, the computer device acquires the time difference between the labeling time and the reference time of the initial label data for each training sample. Then, the computer device determines the maximum time difference from the time differences corresponding to each training sample. Finally, the computer device uses the labeling time of the initial label data for each training sample and the maximum time difference to determine the adjusted target label data for each training sample.

[0062] The reference time can refer to the current time or the time when the training sample set was acquired, as long as all training samples in the training sample set use the same reference time as the standard. This application does not impose specific limitations on this. Furthermore, the time unit corresponding to the reference time should be consistent with the time unit corresponding to the labeling time. Specifically, if the labeling time of the initial label data of the training samples is in "days", then the reference time should also be in "days".

[0063] For example, the initial labeling time for training sample 1 is January 1, 2021, and the reference time can be the current time, specifically March 1, 2021. Then, the computer calculates the time difference between the initial labeling time and the reference time for each training sample. If the training sample set includes 10,000 training samples, after obtaining the initial labeling time for each training sample, the time difference between the initial labeling time and the reference time for each training sample is calculated. Assume the time differences and initial labeling data corresponding to the training samples extracted from the sample pool are as shown in Table 2 below:

[0064] Table 2. Time difference and initial label of training samples

[0065]

[0066] Table 2 provides examples of the time differences and initial labels for 10 training samples. The same method can be used to determine the time differences for all training samples in the sample pool, where each ID corresponds to an account for a training sample. Specifically, the time difference can be the difference in the number of days between the labeled time and the current time. Then, the computer selects samples with a maximum time difference of 180 days from these 10 training samples.

[0067] In one possible implementation, the computer device obtains the ratio between the labeling time and the maximum time difference of the initial label data for each training sample; then, the computer device uses the ratio corresponding to each training sample and the adjustment expression to determine the adjusted target label data for each training sample.

[0068] The ratio between the initial labeling time and the maximum time difference for each training sample can be denoted as f(t), where f(t) = t / N. Here, t is the time difference for each training sample, and N is the maximum time difference among multiple training samples. The adjustment expression can then be specifically shown in formula (2):

[0069] Formula (2)

[0070] Where K is the number of categories corresponding to the training sample set. In a binary classification scenario, K=2, in a three-class classification scenario, K=3, and so on.

[0071] According to formula (2), the original label data of each training sample can be converted into labels smoothed by the time function (i.e., the adjusted target label data) through the label smoothing function.

[0072] In one possible implementation, if the target label data of the target training sample is a first preset value, then the adjusted target label data of the target training sample is determined using the ratio corresponding to the target training sample and a first adjustment expression, wherein the target training sample is any one of multiple training samples. Then, if the target label data of the target training sample is a second preset value, the computer device determines the adjusted target label data of the target training sample using the ratio corresponding to the target training sample, the number of categories corresponding to the training sample set, and the second adjustment expression.

[0073] The first preset value can be 1, and the second preset value can be 0. If the target label data of the target training sample is 1, the computer device uses the ratio corresponding to the target training sample and the first adjustment expression to determine the adjusted target label data of the target training sample. The first adjustment expression can be as shown in formula (3):

[0074] Formula (3)

[0075] Similarly, if the target label data of the target training sample is 0, the computer device uses the ratio corresponding to the target training sample, the number of categories corresponding to the training sample set, and the second adjustment expression to determine the adjusted target label data of the target training sample. The second adjustment expression can be specifically shown in formula (4):

[0076] Formula (4)

[0077] Where K is the number of categories corresponding to the training sample set. In a binary classification scenario, K=2, in a three-class classification scenario, K=3, and so on.

[0078] For example, the number of categories corresponding to a training sample set refers to the number of categories that classify multiple training samples in the training sample set. For instance, a training sample set includes 10 training samples: id1, id2, id3, id4, id5, id6, id7, id8, id9, and id10. Here, id1 corresponds to category 1, id2 to category 2, id3 to category 1, id4 to category 3, id5 to category 1, id6 to category 3, id7 to category 3, id8 to category 1, id9 to category 3, and id10 to category 1. Therefore, the total number of categories for all training samples in the training sample set is 3, and the number of categories corresponding to the training sample set is equal to 3.

[0079] For example, as shown in Table 2, assuming the target training sample is id1, since the initial label data is equal to 1, the label data of id1 is adjusted according to the first adjustment expression, resulting in the adjusted target label data of id1: 1 - 180 / 180 = 0. Similarly, for id2, the label data is adjusted according to the first adjustment expression, resulting in the adjusted target label data of id2: 1 - 170 / 180 = 0.05. Furthermore, the label data of id7 is adjusted according to the second adjustment expression, assuming K = 3, resulting in the adjusted target label data of id7: 150 / 180 * (3 - 1) = 0.42. And the label data of id8 is adjusted according to the second adjustment expression, assuming K = 3, resulting in the adjusted target label data of id8: 130 / 180 * (3 - 1) = 0.42. This process continues, adjusting the original label data of each training sample using either the first or second adjustment expression to obtain the corresponding adjusted target label data. The adjusted target label data for each training sample is shown in Table 3:

[0080] Table 3. Adjusted target labels for training samples

[0081]

[0082] As shown in Tables 2 and 3, smoothing the initial labels by referencing their annotation time results in smoother target labels compared to the original labels. For example, for training sample id2, its initial label is 1. Since this is too far removed from the current time, using the label of this training sample id2 for subsequent model training has little reference value. Therefore, the label is changed from 1 to 0.05 using the time factor. Similarly, for training sample id9, its initial label is 0. Since this is too far removed from the current time, the label may have changed. Therefore, the label is changed from 0 to 0.28 using the time factor, allowing the label of this training sample id9 to account for a certain proportion during model training. In summary, this method utilizes the time factor to smooth labels, resulting in smoothed target labels.

[0083] The data processing method provided in this application allows a computer device to acquire a training sample set including multiple training samples and feature data for each training sample. The feature data includes initial label data and the annotation time of the initial label data. Then, the computing device adjusts the initial label data of each training sample based on the annotation time of the initial label data of the multiple training samples to obtain adjusted target label data for each training sample. This adjusted target label data is used for training a business model. Since the target label data for each training sample is determined after adjustment based on the annotation time of each training sample, the business model trained using the adjusted target label data considers the label changes caused by time factors, thereby improving the accuracy of training the business model.

[0084] Please see Figure 4 , Figure 4 This is a flowchart illustrating a model training method provided in an embodiment of this application. The method is applied to a computer device, which may be, for example, a smartphone, tablet, smart wearable device, in-vehicle device, or other intelligent device. The computer device can also be a server working collaboratively with a server. Figure 4 As shown, the model training method may include steps S410 to S430. Wherein:

[0085] Step S410: Extract features from each of the multiple training samples to obtain the feature matrix of each training sample.

[0086] In one possible implementation, the computer device extracts features from each of the multiple training samples to obtain the feature matrix of each training sample. For details on this process, please refer to [link to relevant documentation]. Figure 5 , Figure 5This is a flowchart illustrating another data processing method provided in an embodiment of this application. For example... Figure 5 As shown in the diagram, the process flow includes three modules: a payment platform module, a computing platform module, and a model management platform module. The payment platform module is primarily used to acquire training samples, that is, to extract raw data and text data from various payment platforms. The raw data can specifically include numerical data, and the text data can specifically include text data; both raw data and text data can be collectively referred to as training samples. The computing platform module is mainly used for data processing and storing the processed data. This primarily involves feature engineering on the raw data and / or text data. Feature engineering can include data cleaning, data anonymization, and other operations to obtain corresponding data features, which are then stored in the computing platform module. Each training sample corresponds to one account, and each training sample's data feature also corresponds to one account. Each training sample includes raw label data. The computing platform module can adjust the raw label data of each training sample to obtain adjusted target label data for each training sample, which is also stored in the computing platform module.

[0087] Finally, the model management platform module is primarily used for model training. The data features of each training sample and the corresponding adjusted target label data can be used to train the business model. Since the target label data of each training sample used for model training has been adjusted using a time function to account for its time-related influence, the adjusted target label data can better train the business model, thereby enabling the business model to have more accurate model recognition capabilities and improving model accuracy.

[0088] Specifically, some steps of data cleaning, such as Figure 3 As shown, Figure 3This is a schematic diagram of a data cleaning process provided in an embodiment of this application. The sample layer stores training samples extracted from the underlying platform, including raw data and text data. Specifically, the data cleaning method for the training samples in the sample layer can include filtering or deleting training samples with a missing rate greater than 90%, etc. Furthermore, feature data obtained after feature engineering based on the training samples is stored in the feature layer. Specifically, the data cleaning method for the feature data in the feature layer can include: 1. deleting feature data with a missing rate greater than 90%; 2. deleting feature data with a uniqueness rate greater than 90%; 3. deleting feature data with an IV (Information Value) less than 0.02, etc.

[0089] In one possible implementation, based on the data characteristics of the text data, the computer device can further convert the 0-1 matrix of the text data into output features that can be used by the business model. Specifically, a neural network model can be used to identify and process the features of the text data to obtain the text features. It should be noted that the neural network model can specifically be a Transformer model, an RNN (Recurrent Neural Network) model, an LSTM (Long Short-Term Memory) network, a GRU (Gated Recurrent Neural Network) model, or other types of network models, such as recurrent neural network models, etc., which are not limited in this invention.

[0090] For example, let's take the Transformer model as an example of a neural network model for a detailed explanation. Please refer to [link / reference needed]. Figure 6 , Figure 6 This is a schematic diagram of the model structure of a Transformer model provided in an embodiment of this application. For example... Figure 6As shown, the Transformer model includes an encoder and a decoder. The parameters input to the encoder can be the input parameters processed by the input word embedding module and position encoding. Then, the input parameters processed by the input word embedding module are input into the encoder. Specifically, the encoder may include a multi-head attention module and a feedforward neural network module. The decoder may include a multi-head attention module, a feedforward neural network module, a masked multi-head attention module, a normalization layer, and an activation function layer, etc.

[0091] Specifically, let's assume the text data is "Who am I?". Figure 2 In this embodiment, the text data undergoes a data transformation process to obtain a 10-dimensional 01 vector. Specifically, the text data "Who am I?" is transformed using a 10-dimensional dictionary to obtain the 10-dimensional 01 vector "0101000010". Then, "0101000010" can be used as the input parameter for the Transformer model. Through the Transformer model's processing, the feature matrix corresponding to the text data "Who am I?" is obtained.

[0092] The specific steps for obtaining transaction features from raw data through feature engineering are as follows: First, raw data is extracted from the underlying data platform. Multi-table matching is performed based on ID (each ID corresponds to one account), and the raw data is merged. Then, the raw data is aggregated according to specific dimensions (feature engineering). Feature engineering involves creating a new X' from X. Basic operations include derivation (dimensionality increase) and filtering (dimensionality reduction). The raw data can specifically be data extracted from a payment platform, such as an information exchange center payment platform, a shopping transaction payment platform, etc.

[0093] In this context, multi-table matching by ID means: for example, if 10,000 samples are needed, and the total number of users is 10e, with features distributed across different tables 'a' and 'b', then first, the features of the 10,000 accounts need to be matched in table 'a', and then the features of the 10,000 accounts need to be matched in table 'b'. Then, the different features matched from tables 'a' and 'b' are merged into a single table by ID. Text features corresponding to the same account are also merged into the same table. Finally, principles such as missing rate and uniqueness can be used to clean the raw data before storing it at the computing platform layer.

[0094] Then, the transaction features obtained after feature engineering of the original data are combined or spliced ​​with the text features to obtain the feature matrix of the training samples.

[0095] Step S420: Train the business model based on the feature matrix of each training sample and the adjusted target label data of each training sample.

[0096] In one possible implementation, the computer device calls the business model to perform recognition processing on the feature matrix of each training sample to obtain the predicted label data of each training sample; then, the computer device determines the loss function of the business model based on the difference between the predicted label data of each training sample and the adjusted target label data of each training sample; finally, the computer device adjusts the model parameters of the business model according to the loss function.

[0097] In one possible implementation, the business model can specifically be a classification model, which can specifically include the XGBOOST classification model. The model parameters of the XGBOOST classification model are as follows:

[0098]

[0099] Here, eta represents the learning rate, which is used to control the weights of the trees. After each iteration, the XGBOOST classification model multiplies the scores of the leaf nodes by this coefficient to weaken the influence of each tree and avoid overfitting.

[0100] `max_depth` represents the maximum depth of the tree. It is also used to avoid overfitting. When its value is larger, the model will learn more specific and local samples, which may lead to overfitting.

[0101] Lambda refers to the L2 regularization term of the weights. This parameter controls the regularization part of the XGBoost classification model, and its uses can be further explored in reducing overfitting.

[0102] Subsample controls the proportion of random sampling for each tree. Decreasing this parameter makes the algorithm more conservative and avoids overfitting. However, if this value is set too small, it may lead to underfitting.

[0103] Of course, the business model provided in this application embodiment can be other classification models besides the XGBOOST classification model, such as the GBDT model, random forest model, neural network model, decision tree algorithm model, etc. The decision tree algorithm model can specifically include: ID3 algorithm model, C4.5 algorithm model, C5.0 algorithm model, and Classification and Regression Tree (CART) algorithm model, etc.

[0104] The XGBOOST classification model can be considered an engineering implementation of the GBDT model algorithm. Specifically, the GBDT model is an ensemble of multiple CART decision trees, where the training objective of each CART decision tree is the negative gradient direction of the current loss function. During node splitting, feature selection is still based on either least squares error or minimum Gini coefficient. The XGBOOST classification model is also an additive model, but its node splitting criteria have changed. Whether it's a regression or classification problem, we first define a loss function, which is any differentiable loss function.

[0105] For example, let's take the XGBOOST classification model as an example for a detailed explanation. Please refer to [link / reference]. Figure 7 , Figure 7 This is a flowchart illustrating another model training method provided in this application embodiment. First, during the training process, initial parameter values ​​are set, the XGBOOST classification model is trained, and then the loss function value is output. This data is saved and input into the Gaussian stochastic process model for the next round of parameter prediction. The parameters predicted by the Gaussian process are then input into the XGBOOST classification model for the second round of training. These steps are then repeated until the loss function drops to the expected value.

[0106] This solution utilizes the difference between the adjusted target label data of each training sample and the predicted label data of each training sample obtained by the model to adjust the model parameters of the business model. Since the target label data of each training sample used for model training has been adjusted by a time function to take into account its influence by time factors, the adjusted target label data can better train the business model, thereby enabling the business model to have a more accurate model recognition ability and thus improving the accuracy of the model.

[0107] Step S430: When the trained business model meets the model convergence condition, the training of the business model is stopped.

[0108] In one possible implementation, the model convergence condition can refer to the following: when the number of training iterations of the business model reaches a preset training threshold, such as 100, the business model satisfies the model convergence condition; that is, training of the business model stops after 100 iterations; when the difference between the predicted label data of each training sample and the adjusted target label data of each training sample is less than an error threshold, the business model satisfies the model convergence condition; when the change between the predicted label data obtained from two consecutive training iterations of the business model is less than a change threshold, the business model satisfies the model convergence condition. Specifically, the business model can include exchange rate models, credit models, etc.

[0109] In one possible implementation, a computer device receives a business prediction request, which carries business data. The computer device then extracts features from the business data to obtain a feature matrix. Next, the computer device uses a business model to identify the business features of the business data, obtaining a risk prediction probability. Finally, if the risk prediction probability is greater than or equal to a reference probability threshold, the business data is marked as risky data.

[0110] In one possible implementation, after stopping the training of the business model, the trained business model can be obtained. This business model is used to predict the risks of business data. Specifically, in the exchange rate model within a risk control scenario, this solution can be used. When a business prediction request carrying business data is received, the trained business model is invoked to perform risk prediction on the business data, obtaining the risk prediction result, which can be a risk prediction probability. If the risk prediction probability of the business data is greater than or equal to a reference probability threshold, the business data is considered to have business risk. Therefore, the business data can be marked as risky data and then sent to the sender of the business acquisition request. Subsequently, when the business data is obtained in an actual business scenario, it can be directly intercepted or filtered, thereby achieving the effect of risk control.

[0111] This solution utilizes the feature matrix of each training sample and its adjusted target label data to train a business model that can be used for risk control. In exchange rate scenarios under risk control, the performance of the model is improved, with a Kolmogorov-Smirnov (KS) improvement of 10%. KS is used to evaluate the model's risk discrimination capability, measuring the difference between the cumulative distributions of good and bad samples. The greater the cumulative difference between good and bad samples, the higher the KS index, and the stronger the model's risk discrimination capability. Since the target label data for each training sample is adjusted based on the labeling time of each training sample, the business model trained using this adjusted target label data considers the label changes caused by time factors, thereby improving the accuracy of the trained business model. Furthermore, the trained business model performs better in risk prediction in business scenarios, improving the accuracy of risk prediction.

[0112] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of a data processing apparatus provided in an embodiment of this application. The data processing apparatus 800 can be applied to... Figures 2-7The computer device in the corresponding method embodiment. The data processing device 800 may be a computer program (including program code) running on a lightweight node, for example, the data processing device 800 is an application software; the device can be used to perform the corresponding steps in the method provided in the embodiments of this application. The data processing device 800 may include:

[0113] The acquisition unit 810 is used to acquire a training sample set, which includes multiple training samples.

[0114] The acquisition unit 810 is further configured to acquire feature data of each training sample among the plurality of training samples, the feature data including initial label data and the labeling time of the initial label data;

[0115] The processing unit 820 is used to adjust the initial label data of each training sample according to the labeling time of the initial label data of the multiple training samples to obtain the adjusted target label data of each training sample, and the adjusted target label data is used for training the business model.

[0116] In one possible implementation, the processing unit 820 adjusts the initial label data of each training sample according to the annotation time of the initial label data of the plurality of training samples to obtain the adjusted target label data of each training sample, including:

[0117] Obtain the time difference between the initial labeling time and the reference time for each training sample;

[0118] The maximum time difference is determined from the time difference corresponding to each training sample;

[0119] The adjusted target label data for each training sample is determined by using the annotation time of the initial label data for each training sample and the maximum time difference.

[0120] In one possible implementation, the processing unit 820 uses the annotation time of the initial label data of each training sample and the maximum time difference to determine the adjusted target label data of each training sample, including:

[0121] Obtain the ratio between the labeling time of the initial label data for each training sample and the maximum time difference;

[0122] The adjusted target label data for each training sample is determined using the ratio and adjustment expression corresponding to each training sample.

[0123] In one possible implementation, the adjustment expression includes a first adjustment expression and a second adjustment expression. The processing unit 820 uses the ratio corresponding to each training sample and the adjustment expression to determine the adjusted target label data for each training sample, including:

[0124] If the initial label data of the target training sample is a first preset value, then the adjusted target label data of the target training sample is determined by using the ratio corresponding to the target training sample and the first adjustment expression, wherein the target training sample is any one of the plurality of training samples;

[0125] If the initial label data of the target training sample is a second preset value, then the adjusted target label data of the target training sample is determined by using the ratio corresponding to the target training sample, the number of categories corresponding to the training sample set, and the second adjustment expression.

[0126] In one possible implementation, the processing unit 820 is also configured to perform the following operations:

[0127] Feature extraction is performed on each of the plurality of training samples to obtain the feature matrix of each training sample;

[0128] The business model is trained based on the feature matrix of each training sample and the adjusted target label data of each training sample;

[0129] When the trained business model meets the model convergence condition, the training of the business model is stopped, and the trained business model is obtained. The trained business model is used to predict risks from business data.

[0130] In one possible implementation, the processing unit 820 trains a business model based on the feature matrix of each training sample and the adjusted target label data of each training sample, including:

[0131] The business model is invoked to identify the feature matrix of each training sample, thereby obtaining the predicted label data for each training sample.

[0132] The loss function of the business model is determined based on the difference between the predicted label data of each training sample and the adjusted target label data of each training sample.

[0133] The model parameters of the business model are adjusted according to the loss function.

[0134] In one possible implementation, the processing unit 820 is also configured to perform the following operations:

[0135] Obtain a business forecast request, wherein the business forecast request carries business data;

[0136] Feature extraction is performed on the business data to obtain the feature matrix of the business data;

[0137] The business model is invoked to identify and process the feature matrix of the business data to obtain the risk prediction probability of the business data.

[0138] If the predicted risk probability is greater than or equal to the reference probability threshold, the business data will be marked as risk data.

[0139] The data processing apparatus provided in this application embodiment can acquire a training sample set including multiple training samples and feature data of each training sample. The feature data includes initial label data and the annotation time of the initial label data. Then, the annotation time of the initial label data of the multiple training samples is used to adjust the initial label data of each training sample, resulting in adjusted target label data for each training sample. This adjusted target label data is used to train the business model. Since the target label data of each training sample is determined after adjustment based on the annotation time of each training sample, the business model trained using the adjusted target label data considers the label changes caused by time factors, thereby improving the accuracy of the business model's recognition performance.

[0140] Please see Figure 9 , Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. The computer device 900 is used to perform... Figures 2-7 The steps performed by the computer device in the corresponding method embodiment, the computer device 900 includes: one or more processors 910; one or more input devices 920; one or more output devices 930; and a memory 940. The processors 910, input devices 920, output devices 930, and memory 940 are connected via a bus 950. The memory 940 is used to store a computer program, the computer program including program instructions, and the processor 910 is used to execute the program instructions stored in the memory 940 to perform the following operations:

[0141] A training sample set, which includes multiple training samples;

[0142] Obtain feature data for each training sample from the plurality of training samples, wherein the feature data includes initial label data and the labeling time of the initial label data;

[0143] Based on the annotation time of the initial label data of the multiple training samples, the initial label data of each training sample is adjusted to obtain the adjusted target label data of each training sample. The adjusted target label data is used for training the business model.

[0144] In one possible implementation, the processor 910 adjusts the initial label data of each training sample based on the annotation time of the initial label data of the plurality of training samples to obtain the adjusted target label data of each training sample, including:

[0145] Obtain the time difference between the initial labeling time and the reference time for each training sample;

[0146] The maximum time difference is determined from the time difference corresponding to each training sample;

[0147] The adjusted target label data for each training sample is determined by using the annotation time of the initial label data for each training sample and the maximum time difference.

[0148] In one possible implementation, the processor 910 uses the annotation time of the initial label data of each training sample and the maximum time difference to determine the adjusted target label data of each training sample, including:

[0149] Obtain the ratio between the labeling time of the initial label data for each training sample and the maximum time difference;

[0150] The adjusted target label data for each training sample is determined using the ratio and adjustment expression corresponding to each training sample.

[0151] In one possible implementation, the adjustment expression includes a first adjustment expression and a second adjustment expression. The processor 910 uses the ratio corresponding to each training sample and the adjustment expression to determine the adjusted target label data for each training sample, including:

[0152] If the initial label data of the target training sample is a first preset value, then the adjusted target label data of the target training sample is determined by using the ratio corresponding to the target training sample and the first adjustment expression, wherein the target training sample is any one of the plurality of training samples;

[0153] If the initial label data of the target training sample is a second preset value, then the adjusted target label data of the target training sample is determined by using the ratio corresponding to the target training sample, the number of categories corresponding to the training sample set, and the second adjustment expression.

[0154] In one possible implementation, processor 910 is also used to perform the following operations:

[0155] Feature extraction is performed on each of the plurality of training samples to obtain the feature matrix of each training sample;

[0156] The business model is trained based on the feature matrix of each training sample and the adjusted target label data of each training sample;

[0157] When the trained business model meets the model convergence condition, the training of the business model is stopped, and the trained business model is obtained. The trained business model is used to predict risks from business data.

[0158] In one possible implementation, the processor 910 trains a business model based on the feature matrix of each training sample and the adjusted target label data of each training sample, including:

[0159] The business model is invoked to identify the feature matrix of each training sample, thereby obtaining the predicted label data for each training sample.

[0160] The loss function of the business model is determined based on the difference between the predicted label data of each training sample and the adjusted target label data of each training sample.

[0161] The model parameters of the business model are adjusted according to the loss function.

[0162] In one possible implementation, processor 910 is also used to perform the following operations:

[0163] Obtain a business forecast request, wherein the business forecast request carries business data;

[0164] Feature extraction is performed on the business data to obtain the feature matrix of the business data;

[0165] The business model is invoked to identify and process the feature matrix of the business data to obtain the risk prediction probability of the business data.

[0166] If the predicted risk probability is greater than or equal to the reference probability threshold, the business data will be marked as risk data.

[0167] It should be understood that the computer device described in the embodiments of this application can perform the foregoing... Figures 2-7 The description of the data processing method in the corresponding embodiments can also be performed as described above. Figure 8The description of the data processing apparatus in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0168] Furthermore, it should be noted that this application embodiment also provides a computer storage medium, which stores a computer program executed by the aforementioned data processing device, and the computer program includes program instructions. When the processor executes the aforementioned program instructions, it can execute the aforementioned... Figures 2-7 The methods described in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same methods will also not be repeated. For technical details not disclosed in the computer storage medium embodiments related to this application, please refer to the description of the method embodiments of this application. As an example, program instructions can be deployed on a computer device, or executed on multiple computer devices located in one location, or executed on multiple computer devices distributed across multiple locations and interconnected via a communication network. These multiple computer devices distributed across multiple locations and interconnected via a communication network can constitute a blockchain system.

[0169] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned... Figures 2-5 The methods described in the corresponding embodiments will not be repeated here.

[0170] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0171] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A data processing method, characterized in that, The method includes: Obtain a training sample set, which includes multiple training samples; Obtain feature data for each training sample from the plurality of training samples, wherein the feature data includes initial label data and the labeling time of the initial label data; Based on the annotation time of the initial label data of the multiple training samples, the initial label data of each training sample is adjusted to obtain the adjusted target label data of each training sample. This includes: obtaining the time difference between the annotation time of the initial label data of each training sample and a reference time; determining the maximum time difference from the time differences corresponding to each training sample; the reference time is the ratio between the annotation time of the initial label data of each training sample and the maximum time difference, using the same reference time as the standard for all training samples in the training sample set; if the initial label data of the target training sample is a first preset value, then the adjusted target label data of the target training sample is determined using the ratio corresponding to the target training sample and a first adjustment expression. The target training sample is any one of the plurality of training samples; if the initial label data of the target training sample is a second preset value, then the adjusted target label data of the target training sample is determined by using the ratio corresponding to the target training sample, the number of categories corresponding to the training sample set, and the second adjustment expression; the original label data of each training sample is converted into a label smoothed by a time function through a label smoothing function; the adjusted target label data is used for training a business model; the business model includes a model for risk prediction of business data in a risk control scenario; the label smoothing function includes the first adjustment expression and the second adjustment expression, the first adjustment expression and the second adjustment expression being different expressions; The loss function for training the business model is determined based on the difference between the predicted label data of each training sample and the adjusted target label data of each training sample.

2. The method according to claim 1, characterized in that, The business model is a classification model.

3. The method according to claim 1, characterized in that, The acquisition of the training sample set includes: extracting raw data from the underlying data platform, the raw data including transaction data, security data and business processing data, and the raw data including relevant data affected by time factors.

4. The method according to claim 1, characterized in that, The initial label data includes either gambling labels or credit labels in risk control scenarios.

5. The method according to any one of claims 1 to 4, characterized in that, The method further includes: Feature extraction is performed on each of the plurality of training samples to obtain the feature matrix of each training sample; The business model is trained based on the feature matrix of each training sample and the adjusted target label data of each training sample; When the trained business model meets the model convergence condition, the training of the business model is stopped, and the trained business model is obtained. The trained business model is used to predict risks from business data.

6. The method according to claim 5, characterized in that, The step of training a business model based on the feature matrix of each training sample and the adjusted target label data of each training sample includes: The business model is invoked to identify the feature matrix of each training sample, thereby obtaining the predicted label data for each training sample. The loss function of the business model is determined based on the difference between the predicted label data of each training sample and the adjusted target label data of each training sample. The model parameters of the business model are adjusted according to the loss function.

7. The method according to claim 5, characterized in that, The method further includes: Obtain a business forecast request, wherein the business forecast request carries business data; Feature extraction is performed on the business data to obtain the feature matrix of the business data; The business model is invoked to identify and process the feature matrix of the business data to obtain the risk prediction probability of the business data. If the predicted risk probability is greater than or equal to the reference probability threshold, the business data will be marked as risk data.

8. A data processing apparatus, characterized in that, The device includes: An acquisition unit is used to acquire a training sample set, wherein the training sample set includes multiple training samples; The acquisition unit is further configured to acquire feature data of each training sample among the plurality of training samples, the feature data including initial label data and the labeling time of the initial label data; The processing unit is configured to adjust the initial label data of each training sample based on the annotation time of the initial label data of the plurality of training samples to obtain the adjusted target label data of each training sample. The adjustment includes: obtaining the time difference between the annotation time of the initial label data of each training sample and a reference time; determining the maximum time difference from the time differences corresponding to each training sample; the reference time is the ratio between the annotation time of the initial label data of each training sample and the maximum time difference, using the same reference time as a standard for all training samples in the training sample set; if the initial label data of the target training sample is a first preset value, then using the ratio corresponding to the target training sample and a first adjustment expression, determining the adjusted target label data of the target training sample; the target training sample is any one of the plurality of training samples. If the initial label data of the target training sample is a second preset value, then the adjusted target label data of the target training sample is determined using the ratio corresponding to the target training sample, the number of categories corresponding to the training sample set, and the second adjustment expression. The original label data of each training sample is converted into a label smoothed by a time function through a label smoothing function. The adjusted target label data is used for training a business model. The business model includes a model for risk prediction of business data in a risk control scenario. The label smoothing function includes the first adjustment expression and the second adjustment expression, which are different expressions. The loss function for training the business model is determined based on the difference between the predicted label data of each training sample and the adjusted target label data of each training sample.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a set of program code, and the processor calls the program code stored in the memory to execute the method described in any one of 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1 to 7.

Images