Abnormal transaction classification method and device, storage medium and electronic equipment

Abstract

The invention discloses an abnormal transaction classification method and device, a storage medium and electronic equipment, and relates to the field of artificial intelligence. The method comprises the following steps: receiving target transaction data; converting the target transaction data into a target feature vector; the target feature vector is input into a target model, a classification result for the target transaction data is obtained according to priori knowledge learned by the target model in a model training stage, and the target model is a classification model which is updated according to model parameters and is obtained based on stochastic differential equation model verification, the stochastic differential equation model is used for mapping the training process of the target model. The technical problems that a traditional deep learning model is like a black box during prediction, the decision process is difficult to explain, and particularly, the reliability of a classification result is low due to the fact that the transparency of the model is insufficient during abnormal transaction recognition are solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and more specifically, to a method, device, storage medium, and electronic device for classifying abnormal transactions. Background Art

[0002] In the financial sector, especially in banks' risk control systems, deep learning models are widely used for abnormal transaction detection due to their advantages in complex data processing and pattern recognition. However, the "black box" nature of deep learning models—the opacity of their internal decision-making processes—has become a major obstacle. When identifying abnormal transactions, models often make decisions based on a large number of nonlinear combinations of features, but these decision-making processes are difficult to intuitively understand and verify, leading to doubts about the reliability of classification results, especially when faced with new or complex fraudulent behaviors. In addition, traditional interpretation methods, such as decision tree rule extraction or feature importance analysis, have limited explanatory power for high-dimensional and unstructured data (such as user operation logs), and without prior knowledge, it is difficult to effectively guide model training and optimization, which further exacerbates the uncertainty of model decisions.

[0003] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0004] The present application provides a method, device, storage medium and electronic device for classifying abnormal transactions, so as to at least solve the technical problem that traditional deep learning models are like "black boxes" when making predictions, and their decision-making processes are difficult to explain. In particular, when identifying abnormal transactions, the model lacks transparency, resulting in low reliability of classification results.

[0005] According to one aspect of the present application, a method for classifying abnormal transactions is provided, comprising: receiving target transaction data; converting the target transaction data into a target feature vector; inputting the target feature vector into a target model, and obtaining a classification result for the target transaction data based on prior knowledge learned by the target model during a model training phase, wherein the target model is a classification model obtained by updating the model parameters and verifying the result based on a stochastic differential equation model, wherein the stochastic differential equation model is used to map the training process of the target model.

[0006] Optionally, the target model is trained in the following manner: obtaining a historical transaction data set, wherein the historical transaction data set includes historical transaction data of N users and an actual category label corresponding to each historical transaction data, wherein N is an integer greater than 1; extracting at least one feature from each historical transaction data to obtain N feature vectors; iteratively training the initial model according to the N feature vectors to obtain the target model, wherein the iterative training is used to adjust the model parameters of the initial model based on the stochastic differential equation model.

[0007] Optionally, after extracting at least one feature in each historical transaction data to obtain N feature vectors, the method further comprises: setting a first value, a second value and random noise based on the N feature vectors, wherein the first value is used to represent the influence degree between samples of the same category, and the second value is used to represent the influence degree between samples of different categories; and constructing a stochastic differential equation model based on the first value, the second value and the random noise.

[0008] Optionally, after constructing the stochastic differential equation model based on the first value, the second value and the random noise, the method further comprises: step 1, solving the N feature vectors by the stochastic differential equation model to obtain a target solution set for each feature vector, wherein the target solution set is a set of solutions of the feature vector at different time points in the stochastic differential equation model; step 2, determining the distribution trend of the N target solution sets; step 3, if there are S target solution sets distributed in a first region and T target solution sets distributed in a second region in the N target solution sets, determining whether the category corresponding to the region to which each target solution set belongs is the same as the actual category label corresponding to the target solution set; wherein T and S are integers greater than or equal to 1, the sum of T and S is N, the first region is used to represent that the target solution set belongs to an abnormal transaction category, and the second region is used to represent that the target solution set belongs to a normal transaction category; step 4, when the number of target solution sets in the N target solution sets for which the category corresponding to the region to which the target solution set belongs is not the same as the actual category label corresponding to the target solution set is greater than a first preset threshold, screening the features in the N feature vectors and updating the N feature vectors based on the screening result; and step 5, repeating steps 1 to 4 until it is detected that the number of target solution sets in the N target solution sets for which the category corresponding to the region to which the target solution set belongs is not the same as the actual category label corresponding to the target solution set is less than or equal to the first preset threshold, taking the N feature vectors as N first feature vectors, and pre-training an initial model according to the N first feature vectors, wherein the pre-training is used to adjust the initial feature value of each feature in the first feature vector based on the stochastic differential equation model.

[0009] Optionally, the initial model is pre-trained according to the N first feature vectors, including: step one, analyzing an influence of an initial feature value of each feature in the N first feature vectors on a distribution trend of a solution of the random differential equation model to obtain an analysis result; step two, identifying at least one feature in the first feature vectors whose influence value on the distribution trend of the solution is greater than a second preset threshold based on the analysis result to obtain an identification result; step three, adjusting the initial feature value corresponding to each feature in each first feature vector based on the identification result, and taking the adjusted first feature vector as a second feature vector to obtain N second feature vectors; step four, pre-training the initial model through the N second feature vectors to obtain N classification results, and determining an error value between a classification result corresponding to each second feature vector and an actual class label to obtain N error values; and step five, if there are H error values greater than a third preset threshold in the N error values and / or the N classification results do not conform to a distribution trend of a solution result of the random differential equation model, repeating steps one to five until the number of error values greater than or equal to the third preset threshold is less than H and the N classification results conform to the distribution trend of the solution result of the random differential equation model, and taking the N second feature vectors as N target feature vectors, wherein H is an integer greater than or equal to 0.

[0010] Optionally, the initial model is iteratively trained according to the N feature vectors to obtain a target model, including: iteratively training the initial model according to the N target feature vectors until, for each target feature vector, a difference between an output result of each layer in the initial model and a solution of the random differential equation model at a corresponding time point of the layer is less than a fourth preset threshold to obtain the target model, wherein the number of time points for solving in the random differential equation model corresponds to the number of layers in the initial model one by one.

[0011] Optionally, each iteration training comprises the following steps: inputting the target feature vector into the initial model, processing the target feature vector according to each layer in the initial model, and taking the combination of the processing results of each layer corresponding to the target feature vector as the first output result set; inputting the target feature vector into the stochastic differential equation model for solving, obtaining the output result of the target feature vector at each time point, and taking the combination of the output result of the target feature vector at each time point as the second output result set; determining the difference between the processing result of each layer in the first output result set and the output result of the layer at the corresponding time point in the second output result set; if it is detected that the difference between the processing result of each layer in the first output result set and the output result of the layer at the corresponding time point in the second output result set is less than the fourth preset threshold, the next round of iteration training is performed; if it is detected that the difference between the processing result of the i th layer in the first output result set and the output result of the i th layer at the corresponding time point in the second output result set is greater than or equal to the fourth preset threshold, the model parameters of the initial model are adjusted, where i is an integer greater than or equal to 1.

[0012] According to another aspect of the present application, there is also provided an abnormal transaction classification apparatus, comprising: a receiving unit configured to receive target transaction data; a converting unit configured to convert the target transaction data into a target feature vector; and a determining unit configured to input the target feature vector into a target model, and obtain a classification result for the target transaction data according to prior knowledge learned by the target model in a model training stage, wherein the target model is a classification model obtained by updating model parameters and verifying based on a stochastic differential equation model, and the stochastic differential equation model is used to map a training process of the target model.

[0013] According to another aspect of the present application, there is also provided a computer readable storage medium having a computer program stored therein, wherein the computer program, when executed, causes a device in which the computer readable storage medium is located to perform the abnormal transaction classification method.

[0014] According to another aspect of the present application, there is also provided an electronic device comprising one or more processors and a memory configured to store one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to perform the abnormal transaction classification method.

[0015] According to another aspect of the present application, there is also provided a computer program product comprising computer instructions configured to implement the steps of the abnormal transaction classification method when executed by a processor.

[0016] In the present application, first, target transaction data is received, second, the target transaction data is converted into a target feature vector, and then the target feature vector is input into a target model, and according to the prior knowledge learned by the target model in the model training stage, a classification result for the target transaction data is obtained, wherein the target model is a classification model obtained by updating the model parameters and verifying based on a stochastic differential equation model, wherein the stochastic differential equation model is used to map the training process of the target model. That is, by using the stochastic differential equation to map and verify the model training process, the purpose of enhancing the model decision transparency is achieved, thereby realizing the technical effect of improving the reliability of the classification result, and further solving the technical problem that the traditional deep learning model is like a "black box" when predicting, and it is difficult to explain its decision-making process, especially when identifying abnormal transactions, the transparency of the model is insufficient, resulting in low reliability of the classification result. BRIEF DESCRIPTION OF DRAWINGS

[0017] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0018] Figure 1 is a flowchart of an optional abnormal transaction classification method according to an embodiment of the present application;

[0019] Figure 2 is a schematic diagram of an optional abnormal transaction classification method according to an embodiment of the present application;

[0020] Figure 3 is a schematic diagram of an optional abnormal transaction classification device according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0022] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and in the above drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0023] It should be noted that the collected information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, analyzed data, etc.) involved in the present application are information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards, necessary security measures are taken, do not violate public order and good customs, and provide corresponding operation portal for user to choose authorization or refusal. For example, interfaces are provided between the system and related users or institutions to provide corresponding operation portals for users to choose to agree or refuse automatic decision results; if the user chooses to refuse, the expert decision process is entered. And all data processing processes of the present application comply with the requirements of the Personal Information Protection Law and the Financial Data Security Grading Guide.

[0024] According to the embodiments of the present application, a method embodiment of an abnormal transaction classification method is provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0025] It should be noted that an intelligent classification system can be used as the execution subject of the abnormal transaction classification method of the embodiments of the present application. It can be understood that the abnormal transaction classification method provided by the embodiments of the present application can also be used as the execution subject by other systems or devices, and the embodiments of the present application do not make specific limitations.

[0026] Figure 1 is a flowchart of an optional abnormal transaction classification method according to the embodiments of the present application, as shown in Figure 1 The method comprises the following steps:

[0027] Step S101, receiving target transaction data.

[0028] Optionally, the target transaction data refers to various types of information related to user transaction behavior collected by financial institutions in their daily operations. These information may include but not limited to transaction amount, transaction time, transaction location, user account information, device information, transaction type, etc. The intelligent classification system is connected to the data source through an interface, automatically obtaining or receiving these data through manual import, providing raw information for subsequent feature conversion and model prediction.

[0029] Step S102, converting the target transaction data into a target feature vector.

[0030] Optionally, after receiving the target transaction data, the intelligent classification system converts it into a target feature vector. Feature vector is a method of converting raw data into numerical representation, which is convenient for machine learning model processing. In this step, the intelligent classification system will preprocess and extract features from the original transaction data, convert non-numeric information (such as timestamp, device model) into numerical form, and may also perform data standardization, normalization, or even advanced feature engineering (such as calculating transaction frequency within a sliding window) to form a vector containing all necessary features. The target feature vector is the standard form of model input format, which directly determines the efficiency and effectiveness of model learning.

[0031] Step S103, inputting the target feature vector into the target model, and obtaining the classification result for the target transaction data according to the prior knowledge learned by the target model in the model training stage.

[0032] In step S103, the target model is updated according to the model parameters and the classification model obtained based on the stochastic differential equation model verification.

[0033] In step S103, the stochastic differential equation model is used to map the training process of the target model.

[0034] Optionally, a fairness constraint mechanism is introduced in the training process of the target model, which ensures the consistent distribution of classification errors for different user groups through statistical testing.

[0035] Optionally, when the target feature vector is input into the target model, the model will use the prior knowledge accumulated in its training stage, i.e. the patterns and features learned from historical transaction data, to predict and classify the current transaction data. This process is actually the reasoning stage of the model, which processes the input feature vector through multiple layers of neural networks. Each layer of neural network converts the input data through a series of weight and bias parameters, and finally outputs a classification result indicating whether the transaction is abnormal. For example, for a binary classification problem, the model may output a probability value indicating the likelihood of the transaction belonging to the abnormal class.

[0036] Optionally, the target model is a trained deep learning classifier, which can be verified based on a stochastic differential equation (SDE) model. The SDE model here serves to map and verify the training process of the target model through the existence and uniqueness of the solution and the trend of the solution over time, ensuring that the decision rules learned by the model have interpretability and stability. The training process of the target model is assisted by the SDE model, which not only enhances the decision-making ability of the model, but more importantly improves the reliability and interpretability of the classification results, overcoming the limitations of the "black box" problem of deep learning models, especially in the scenario of abnormal transaction identification, providing financial institutions with more accurate and transparent risk control solutions.

[0037] As can be seen from the contents of steps S101 to S104, in the present application, first, the target transaction data is received, then the target transaction data is converted into a target feature vector, and then the target feature vector is input into the target model. According to the prior knowledge learned by the target model in the model training stage, the classification result for the target transaction data is obtained, wherein the target model is a classification model that is updated according to the model parameters and verified based on a stochastic differential equation model, wherein the stochastic differential equation model is used to map the training process of the target model. That is, by using the stochastic differential equation to map and verify the model training process, the purpose of enhancing the transparency of the model decision is achieved, thereby realizing the technical effect of improving the reliability of the classification result, and further solving the technical problem that the traditional deep learning model is like a "black box" when predicting, and it is difficult to explain its decision-making process. Especially in identifying abnormal transactions, the transparency of the model is insufficient, resulting in low reliability of the classification result.

[0038] In an optional embodiment, the intelligent classification system first obtains a historical transaction data set, wherein the historical transaction data set includes historical transaction data of N users and actual class labels corresponding to each historical transaction data, wherein N is an integer greater than 1, then extracts at least one feature in each historical transaction data to obtain N feature vectors, and then iteratively trains an initial model according to the N feature vectors to obtain a target model, wherein the iterative training is used to adjust the model parameters of the initial model according to the stochastic differential equation model.

[0039] Optionally, the intelligent classification system first obtains a historical transaction data set through a data interface or from a data warehouse, which is a data set containing N user transaction records in the past, where N is an integer greater than 1, representing the size of the data set. The historical transaction data set not only contains detailed information of the transaction, such as transaction amount, transaction time, transaction location, account information, etc., but also contains the actual class label of each transaction, i.e., indicating whether the transaction belongs to the abnormal class, which is an essential element in supervised learning, used to guide the model to learn to distinguish between normal transactions and abnormal transactions.

[0040] Optionally, after obtaining the historical transaction data set, the intelligent classification system then performs feature extraction on these raw data. Feature extraction is the process of converting raw transaction information into numerical features that can reflect the essential properties and patterns of transactions. For example, extracting numerical features from transaction amounts, transaction frequencies, extracting periodic features from transaction times, extracting geographic distribution features from geographic locations, etc. Through a series of data preprocessing and feature engineering operations, the system converts each historical transaction data into a feature vector, and finally obtains N such feature vectors. These feature vectors are the direct input of model training, which helps the model to learn and identify key patterns in transaction data.

[0041] Optionally, after obtaining the N feature vectors, the intelligent classification system takes these vectors as training samples and begins to iteratively train the initial model to obtain the final target model. The initial model is an untrained deep learning network, and its parameters are in a random or preset state. During the iterative training process, the system inputs the feature vectors into the model, uses optimization algorithms such as stochastic gradient descent, and constantly updates the model parameters according to the difference between the actual class labels and the model prediction results to minimize the prediction error. In particular, the concept of stochastic differential equation model is introduced during the iterative training process. The role of the SDE model here is not only to verify the model training effect, but also to guide the adjustment of model parameters through the existence of unique solutions and the trend of solution changes over time. This means that each update of the model parameters will be checked by the SDE model to ensure that the direction of parameter adjustment conforms to the optimization of the classification trend, thereby ensuring the stability of the model during the learning process and the accuracy of the classification.

[0042] From the above, it can be seen that through the above steps, the process of building the target model by the intelligent classification system has been significantly optimized. The introduction of the historical transaction data set provides rich learning materials for model training, ensuring the generalization ability of the model. The construction of feature vectors converts complex data into a form that is easy for the model to process, accelerating the training process. Most importantly, the application of the SDE model in the iterative training process not only improves the efficiency and accuracy of model training, but more importantly, it enhances the explainability of model decision-making, making the model training process transparent through the perspective of mathematical equations. This provides very effective technical support for financial institutions in the detection of abnormal transactions, both in pursuit of high accuracy and in ensuring that the decision-making process is compliant and transparent. Overall, these steps work together to enable the intelligent classification system to not only provide highly accurate classification results when identifying abnormal transactions, but also to clearly explain the basis for the decision, greatly improving the intelligent level of financial institutions in the field of risk control and user experience.

[0043] In an optional embodiment, the intelligent classification system sets a first value, a second value, and random noise based on the N feature vectors, wherein the first value is used to represent the influence degree between samples of the same category, and the second value is used to represent the influence degree between samples of different categories, and then constructs a stochastic differential equation model based on the first value, the second value, and the random noise.

[0044] Optionally, before constructing the stochastic differential equation (SDE) model, the intelligent classification system sets two key numerical parameters: a first value and a second value, and introduces random noise based on the N feature vectors extracted from historical transaction data. The first value is used to quantify the mutual influence degree between samples of the same category (such as normal transactions or abnormal transactions); the second value measures the distinguishing influence between samples of different categories, i.e., the quantification of the distinction between normal transactions and abnormal transactions. The setting of these two values essentially provides initial values for the within-group connection and between-group connection strength in the SDE model, ensuring that the model can correctly understand the close or distant relationship between categories.

[0045] Optionally, constructing the SDE model is one of the core steps of the present application. The SDE model is a mathematical tool used to describe the dynamic changes of a stochastic process over time. In the intelligent classification system, the SDE model is designed to map the evolution trend of feature vectors over time during the training process of the deep learning model. Specifically, the system first sets a first value and a second value, and a random noise parameter based on the N feature vectors, which together constitute the key components of the SDE model. The first value and the second value correspond to the parameters representing the within-group connection and the between-group connection in the SDE model in a mathematical sense, while the random noise simulates the randomness introduced during the deep learning training due to the uncertainty of the data or the random initialization of the model. The construction of the SDE model aims to predict and verify the stability and classification effect of the deep learning model training through mathematical means, providing guidance for subsequent model training.

[0046] Optionally, once the SDE model is constructed, it begins to play its core role. During the training process of the deep learning model, each feature vector changes with the training iteration, and these changes reflect the progress of model learning. The SDE model ensures the consistency and predictability of the deep learning model training results through the existence and uniqueness of its solution. At the same time, the trend of the solution of the SDE model over time can help the intelligent classification system monitor the process of model training and judge whether the model is progressing in the direction of optimizing the classification effect. Especially for abnormal transaction identification, the SDE model can intuitively show how the feature vectors gradually differentiate over time, allowing different categories of transactions to be clearly distinguished in the feature space, thereby verifying the classification ability of the model.

[0047] Optionally, each deep learning neural network can have a different stochastic differential equation model corresponding to it, i.e., different stochastic differential equation models can be set for different deep learning neural networks.

[0048] From the above, it can be seen that the above steps of constructing the SDE model significantly improve the abnormal transaction detection capability of the intelligent classification system. By setting the first value (in-group contact strength) and the second value (inter-group contact strength), and introducing random noise, the SDE model can mathematically predict and verify the training process and classification effect of the deep learning model. This not only enhances the transparency and interpretability of model decision-making, but also ensures that the model can fully understand and distinguish different categories of transactions during the training process, especially the subtle differences between abnormal transactions and normal transactions. With the guidance of the SDE model, the intelligent classification system can not only effectively identify abnormal transactions, but also clearly explain the basis for identification, providing financial institutions with a more reliable and transparent risk control tool. Overall, the application of the SDE model greatly improves the performance and user trust of deep learning models in abnormal transaction detection.

[0049] In an optional embodiment, after the intelligent classification system constructs the stochastic differential equation model based on the first value, the second value, and the random noise, it further performs the following steps: Step 1, solving the N feature vectors through the stochastic differential equation model to obtain a target solution set for each feature vector, wherein the target solution set is a set of solutions of the feature vector at different time points in the stochastic differential equation model; Step 2, determining the distribution trend of the N target solution sets; Step 3, if there are S target solution sets distributed in the first region and T target solution sets distributed in the second region among the N target solution sets, determining whether the category corresponding to the region to which each target solution set belongs is the same as the actual category label corresponding to the target solution set; wherein T and S are integers greater than or equal to 1, the sum of T and S is N, the first region is used to represent that the target solution set belongs to the abnormal transaction category, and the second region is used to represent that the target solution set belongs to the normal transaction category; Step 4, when the number of target solution sets in the N target solution sets for which the category corresponding to the region to which the target solution set belongs is not the same as the actual category label corresponding to the target solution set is greater than a first preset threshold, filtering the features in the N feature vectors and updating the N feature vectors based on the filtering result; Step 5, repeating steps 1 to 4 until the number of target solution sets in the N target solution sets for which the category corresponding to the region to which the target solution set belongs is not the same as the actual category label corresponding to the target solution set is less than or equal to the first preset threshold, taking the N feature vectors as N first feature vectors, and pre-training the initial model according to the N first feature vectors, wherein the pre-training is used to adjust the initial feature value of each feature of the first feature vector based on the stochastic differential equation model.

[0050] Optionally, after the SDE model is constructed, the intelligent classification system performs a mathematical solving operation on the previously extracted N feature vectors. This step aims to map the dynamic change trend of each feature vector during the training process through the SDE model. The result of the solving is a set of target solutions for each feature vector at different time points, which provides specific numerical results of the evolution of the feature vector over time, and can intuitively see how the feature vector gradually differentiates and stabilizes during the model training process. Next, the system analyzes the distribution trend of the N target solution sets. This analysis process aims to understand the overall pattern of feature vector changes, especially focusing on those solution sets that show a clear differentiation trend in the later stage of training, as they are more likely to carry key information that distinguishes normal from abnormal transactions. By analyzing the distribution of solution sets, the system can identify which features are increasing in importance and which features are decreasing in importance, providing a basis for subsequent feature selection.

[0051] Optionally, after the distribution trend of the target solution sets is basically understood, the system further checks whether the location of each solution set distribution matches the actual class label of the transaction to which the solution set belongs. Here, the first region and the second region represent the feature space of the abnormal transaction category and the normal transaction category, respectively. If the solution set is distributed in the first region, it should theoretically correspond to a label of an abnormal transaction; conversely, if the solution set is distributed in the second region, it should correspond to a label of a normal transaction. Through this check, the system can assess whether the SDE model's prediction of the training process is accurate and whether the representation of the feature vector is appropriate. If the number of solution sets whose region category does not match the actual class label exceeds a preset first threshold among the N target solution sets, the system will trigger the feature vector selection and update process. This process involves detailed analysis of the N feature vectors to identify those features that cause the distribution of the solution set to mismatch the actual category, for example, some features may have changed too dramatically during the training process, or some features may not have contributed significantly to the classification decision. The system will eliminate or weaken these features, while possibly strengthening those features that perform well in the consistency check, ultimately updating the N feature vectors to obtain more accurate classification results in the next round of solving.

[0052] Optionally, the intelligent classification system will execute the above steps in a loop until the number of solution sets in the N target solution set that do not match the actual category label of the region category to which the solution set belongs drops to equal to or below a first preset threshold. This indicates that after the feature vectors are screened and updated, the SDE model can accurately predict the trend of changes in the feature vectors during the training of the classification model, and this prediction is highly consistent with the actual classification label. At this time, the intelligent classification system uses the N screened and optimized feature vectors as a first feature vector set to pre-train the initial model. The pre-training process is based on the guidance of the SDE model, not only optimizing the initial feature values of each feature, but also adjusting the model parameters, and ensuring that the model can focus on the most relevant features in the early stage of training, thereby speeding up the model learning process and improving the overall classification effect.

[0053] Optionally, before the deep learning training starts, there are many uncertainties, such as not knowing how many deep levels of neural network need to be built, not knowing what learning rate to use, and not being able to predict whether the training result can meet the expected goal. In order to solve these unknown problems in the early stage, the embodiment introduces a prediction mechanism based on stochastic differential equations in the feature processing stage. By analyzing the feature vectors and using the intrinsic properties of stochastic differential equations, the system can gain insight into the training trajectory of the deep learning model in advance, which means that before the actual training starts, the system can have a certain degree of prediction and grasp of the training effectiveness based on the characteristics of stochastic differential equations. This innovation makes the deep learning training process more predictable and controllable, providing a strong theoretical basis for model design and parameter optimization.

[0054] From the above, through the above steps, the intelligent classification system not only builds an SDE model that guides the model training process, but also ensures the efficiency and accuracy of model training through repeated screening and optimization of feature vectors. The key to this process is that it closely integrates mathematical theory with deep learning practice, intelligently adjusts feature vectors through SDE model solving and analysis, thereby fundamentally improving the decision-making ability and interpretability of the model. Finally, the pre-training process can quickly converge to an ideal state based on the optimized first feature vector set, significantly improving the efficiency and accuracy of abnormal transaction detection, while also providing a transparent and reliable risk control decision-making tool for financial institutions.

[0055] In an alternative embodiment, the intelligent classification system pre-trains the initial model based on N first feature vectors, including the following steps: Step one, analyze the influence of the initial feature values of each feature in the N first feature vectors on the distribution trend of the solution of the random differential equation model, and obtain an analysis result; Step two, based on the analysis result, identify at least one feature in the first feature vector whose influence value on the distribution trend of the solution is greater than a second preset threshold, and obtain an identification result; Step three, based on the identification result, adjust the initial feature value corresponding to each feature in each first feature vector, and take the adjusted first feature vector as a second feature vector, to obtain N second feature vectors; Step four, pre-train the initial model based on the N second feature vectors, obtain N classification results, and determine the error value between the classification result corresponding to each second feature vector and the actual class label, to obtain N error values; Step five, if there are H error values greater than a third preset threshold in the N error values and / or the N classification results do not conform to the distribution trend of the solution of the random differential equation model, repeat steps one to five until the number of error values greater than or equal to the third preset threshold is less than H and the N classification results conform to the distribution trend of the solution of the random differential equation model, and take the N second feature vectors as N target feature vectors, wherein H is an integer greater than or equal to 0.

[0056] Optionally, the intelligent classification system first conducts a detailed analysis of the initial feature values of each feature in the N first feature vectors, examining how these values affect the distribution trend of the solution of the stochastic differential equation (SDE) model over time. This analysis step is crucial as it helps the system understand which feature values have a significant impact on the decision-making process of the model. Through mathematical methods such as differentiation, integration, or data-based statistical analysis, the system can estimate the degree of influence of each feature value on the trend of the solution of the SDE model, thereby obtaining a detailed analysis result report. Based on the analysis result obtained in the first step, the intelligent classification system then identifies those features that have a significant impact on the distribution trend of the solution of the SDE model. Here, the system sets a second preset threshold to determine whether the influence value of a feature is large enough. If the influence value of a feature exceeds the second preset threshold, it is considered a key feature that significantly contributes to the decision of the model. Through this screening mechanism, the system can focus on the most important features, avoiding wasting computational resources on irrelevant or redundant features, and improving the efficiency of model training.

[0057] Optionally, after identifying the key features, the intelligent classification system will adjust the initial feature values in the first feature vector. This adjustment process is based on the identification results obtained in step two, aiming to optimize the feature vector to better fit the actual transaction data distribution. The adjustment can be manually set or automatically completed through machine learning algorithms. The adjusted first feature vector is called the second feature vector, and this series of operations ultimately generates N optimized second feature vectors, preparing high-quality data for the subsequent pre-training process.

[0058] Optionally, with the optimized second feature vector set, the intelligent classification system then pre-trains the initial model. After completing the pre-training, the system generates a classification result for each second feature vector input to the model and compares it with the actual class label corresponding to the feature vector, calculating the error value between each classification result and the actual label, thereby obtaining a set of N error values. These error values reflect the performance of the model in the pre-training stage, providing intuitive feedback for subsequent parameter adjustment.

[0059] Optionally, the system will check whether there are more than third preset threshold error values, and also evaluate whether the N classification results follow the distribution trend of the SDE model solution. If there are more than H error values exceeding the threshold or the classification result trend does not meet expectations, the system will automatically return to the step of analyzing the N first feature vectors and re-analyze and adjust until all error values are within an acceptable range and the classification result trend is consistent with the SDE model prediction. This iterative optimization process ensures the quality of pre-training data and the preliminary generalization ability of the model, laying a solid foundation for the formal training stage.

[0060] Optionally, traditional deep learning neural networks do not always follow the simple linear relationship that "the more layers, the better performance". In practice, excessive increase in network layers may lead to high model complexity, which may affect training results and the generalization ability of the model. To solve this problem, this embodiment introduces a novel pre-training strategy that monitors the dynamic behavior of the solution as it evolves over time by introducing initial value analysis of stochastic differential equations. This method can understand the trend of the solution at a specific time during network training, and whether there is a divergence or "explosion" phenomenon of the solution, i.e., the numerical value of the solution becomes unstable or infinite. Through in-depth understanding of these dynamic characteristics, the system can more scientifically guide the architecture design of deep learning neural networks, such as the selection of the number of layers and the setting of the learning rate, rather than relying solely on trial and error, i.e., adjusting these parameters by repeatedly testing the performance of the final model. Therefore, this embodiment provides a more detailed and more predictable model parameter adjustment guide, which helps to optimize the training process and improve the overall performance of the model.

[0061] From the above, the pre-training optimization process is essentially a fine management and optimization of the model training process. By analyzing how the initial values of the feature vectors affect the solution trend of the SDE model, the system can intelligently identify and adjust key features to ensure that the distribution of data and the decision direction of the model are consistent during the model training process. This series of steps significantly enhances the efficiency of model training, reduces unnecessary computational waste, and improves the interpretability of the model, ensuring that the decision logic of the model is close to human understanding. Finally, when N second feature vectors are used as target feature vectors for formal model training, the model already has a good initial state and can quickly converge to the optimal solution, improving the accuracy and speed of abnormal transaction detection. Overall, this pre-training process greatly optimizes each link of model training, providing strong technical support for the application of intelligent classification systems in the field of abnormal transaction detection.

[0062] In an alternative embodiment, the intelligent classification system iteratively trains the initial model according to the N target feature vectors until, for each target feature vector, the difference between the output of each layer of the initial model and the solution of the stochastic differential equation model at the corresponding time point of that layer is less than a fourth predetermined threshold, resulting in a target model. The number of time points for solving in the stochastic differential equation model corresponds one-to-one to the number of layers in the initial model.

[0063] Optionally, the intelligent classification system first prepares N target feature vectors, which have been adjusted to the best state through previous screening, optimization and pre-training steps, and are used to most effectively guide the training of the deep learning model. The system then inputs these vectors into the initial model to begin the initialization of iterative training. The initial model can be a deep neural network, which may have multiple layers from the input layer to the output layer, each layer having specific weights and bias terms for extracting more abstract features from the feature vectors.

[0064] Optionally, to ensure that the output of each layer of the model training is consistent with the expected classification trend, the system also runs a stochastic differential equation (SDE) model, which is used to describe and predict the trend of the target feature vector over time (or the number of layers). The number of time points for solving in the SDE model corresponds to the number of layers of the deep model, meaning that the training of each layer will be compared with the solution of the SDE model at the corresponding time point. Specifically, the SDE model will solve its equation at each predetermined time point to generate a solution set, which mathematically represents the expected trend of the feature vector during the training process.

[0065] Optionally, the intelligent classification system begins an iterative training process, checking the output of each layer in the initial model, i.e., the state of the feature vector after the layer is trained, at each iteration, and comparing it with the solution of the SDE model at the same time point (corresponding to the number of layers). If the difference between the output of a certain layer and the solution of the SDE model exceeds the fourth preset threshold, it indicates that the current model parameter adjustment may not have achieved the ideal classification trend, and therefore parameter adjustment is needed until the difference is less than the fourth preset threshold, ensuring that the model learning is consistent with the SDE model prediction. During the entire training process, the system continuously monitors the difference between the output of the model and the solution of the SDE model, and once it finds that the difference of any layer exceeds the fourth preset threshold, it immediately takes measures to adjust the parameters of that layer, such as updating the weights and biases through the backpropagation algorithm, to ensure that the output of the model is increasingly close to the solution of the SDE model. This continuous training and parameter optimization ensures that the decision-making process of the model can be explained by the mathematical model, increasing the transparency and interpretability of the model.

[0066] Optionally, the iterative training continues until the difference between the output of each layer in the initial model and the solution of the SDE model at the corresponding time point for each of the N target feature vectors is less than the fourth preset threshold. This means that the model training has reached the predetermined accuracy standard, and the learning process strictly follows the prediction trend of the SDE model. At this time, it can be said that the model training is successful, and the target model obtained performs well in the classification task, and its decision-making process is also supported by mathematical demonstration, increasing the scientificity and reliability of the model.

[0067] From the above, it can be seen that the intelligent classification system combines deep learning model training with stochastic differential equation model, creating a new model training method. This method not only ensures the accuracy of the intelligent classification system in abnormal transaction detection, but also monitors the learning state of the model in real time during the training process, ensuring that the training of each layer strictly follows the preset mathematical model change trend, thereby greatly enhancing the interpretability of the model and the transparency of the classification decision. Compared with traditional methods, this method can better balance the complexity and interpretability of the model, providing financial institutions with a more accurate and easy-to-understand risk management tool, further improving the efficiency of the risk control system and customer trust.

[0068] In an optional embodiment, each iterative training includes the following steps: the intelligent classification system first inputs the target feature vector into the initial model, processes the target feature vector according to each layer in the initial model, and uses the combination of the processing results of each layer corresponding to the target feature vector as the first output result set; then inputs the target feature vector into the stochastic differential equation model for solution to obtain the output result of the target feature vector at each time point, and uses the combination of the output results of the target feature vector at each time point as the second output result set; then determines the difference between the processing result corresponding to each layer in the first output result set and the output result of the layer at the corresponding time point in the second output result set; if it is detected that the difference between the processing result corresponding to each layer in the first output result set and the output result of the layer at the corresponding time point in the second output result set is less than a fourth preset threshold, the next round of iterative training is performed; if it is detected that the difference between the processing result corresponding to the i-th layer in the first output result set and the output result of the i-th layer at the corresponding time point in the second output result set is greater than or equal to the fourth preset threshold, the model parameters of the initial model are adjusted, where i is an integer greater than or equal to 1.

[0069] Optionally, at the beginning of each iterative training, the intelligent classification system inputs the target feature vector into the initial deep learning model. Each layer processes the feature vector and generates a processing result. At the same time, the target feature vector is also input into the stochastic differential equation (SDE) model for solution, resulting in a series of output results, each of which corresponds to a specific time point (or layer) in the initial model. The system collects the processing results of each layer to form a first output result set, and similarly collects the output results of the SDE model at each time point to form a second output result set.

[0070] Optionally, next, the intelligent classification system will compare the processing results of each layer in the first output result set with the output results at the corresponding time point in the second output result set one by one, and calculate the difference between the two. The size of the difference reflects the deviation between the learning effect of the deep learning model at that layer and the classification trend predicted by the SDE model. If the difference of all layers is less than the fourth preset threshold, this indicates that the current state of the model is relatively ideal and can better follow the prediction of the SDE model. At this time, the system will conduct the next round of iterative training and continue to optimize the model parameters to pursue higher classification accuracy. However, if the difference of the i-th layer is found to be greater than or equal to the fourth preset threshold in the comparison, this means that there is a large deviation between the learning effect of the current model at the i-th layer and the prediction of the SDE model, and parameter adjustment is required. The system will use an optimization algorithm (such as gradient descent method) to adjust the relevant parameters in the initial model based on the difference of the layer, so that in the next round of iteration, the output of the layer is closer to the prediction of the SDE model, thereby making the decision-making process of the entire model more reasonable and accurate.

[0071] From the above, the iterative training method proposed by the above steps combines the flexible learning ability of the deep learning model and the mathematical rigor of the SDE model, forming an efficient and interpretable model training process. By comparing the output results of the initial model and the SDE model in each iteration training, the system can monitor the learning state of the model in real time, timely discover and correct the layers deviating from the expected classification trend, and ensure the stability of the model training process and the optimization of the learning effect. This method not only improves the classification accuracy of the model, but also enhances the transparency and interpretability of the model decision. For financial institutions, it can more quickly and accurately identify abnormal transactions, while reducing the risk of misjudgment, improving the overall reliability and efficiency of the risk control system. In addition, with the guidance of the SDE model, the parameter adjustment of the initial model has a scientific basis, avoiding blind adjustment and greatly shortening the time required for the model to reach a convergent state.

[0072] In an alternative embodiment, Figure 2 is a schematic diagram of an alternative method for classifying abnormal transactions according to an embodiment of the present application, as Figure 2 shown, comprising a data module, a feature processing module, a pre-training module, and a deep learning trainer. This embodiment follows the main framework of traditional deep learning, obtaining data from the data module, processing the features of the data source in the feature processing module, converting the data source into a feature vector form that can be trained by deep learning, pre-processing the original feature vector to obtain a modified feature vector that can obtain better results, and then putting it into the deep learning neural network trainer to obtain the learning result. However, the random differential equation module is used to assist in processing in the feature processing module, the pre-training module, and the deep learning trainer module. Specifically:

[0073] In the feature processing module, the system constructs a random differential equation for the corresponding problem, and uses the solution property of the differential equation, i.e., the existence and uniqueness of the solution, to pre-validate the training results of the classification problem. If the differential equation has given the final trend of the classification problem as differentiation, then the classification effect of the obtained learner must be interpretable.

[0074] Meanwhile, in the pre-training module, the system uses the relationship between the initial value and the solution of the random differential equation to assist the pre-training module in converting the feature vector into an initial value with good properties. The initial value of the differential equation often has a decisive role in the properties of its solution, and the use of the initial value requirement of the differential equation can ensure that the subsequent trained learner is interpretable.

[0075] In the training process of the deep learning trainer, the system monitors the training process with the trend of the solution of the stochastic differential equation. The time variation T of the stochastic differential equation is analogous to each layer of the neural network training, and the trend of the solution of the stochastic differential equation with time T is taken as the result variation of each layer training in the neural network training process. The trend variation of the solution of the stochastic differential equation is used to completely reproduce the black box system of the deep learning neural network training.

[0076] Optionally, the following is a specific example to illustrate how to use a stochastic differential equation to describe the deep learning training process:

[0077] Suppose there is a K-classification problem, let is the feature vector of a certain i sample belonging to the kth classification at the mth layer training, then the feature vector has such a change formula from the m-1th layer to the mth layer as shown in equation (1):

[0078]

[0079] That is, it is affected by a certain j m sample m-1th layer feature vector and the influence of other samples belonging to the k category, the parameter a describes the measure of the mutual influence between different category sample feature vectors, plus a beta representing Gaussian noise (this noise can be changed, here the most basic Gaussian noise is selected).

[0080] The system uses two quantities to represent the influence between different categories and the influence between the same category, which are called intra-group connection and inter-group connection. Next, for simplicity, assume that it is a binary classification problem, so there will be only two quantities to describe the intra-group connection and inter-group connection. Let's simply consider that the intra-group connection is the same, and let the two quantities be constants γ and θ, which represent the intra-group connection of class 1 and class 2 and the inter-group connection between class 1 and class 2. Therefore, substituting the above equation (1) can obtain the following equation group as shown in equations (2)-(3):

[0081]

[0082]

[0083] wherein, is the feature vector value of another sample i' in the m-1th layer; is the feature vector value of a certain sample j in the second category in the m-1th layer.

[0084] The left side of the equation is considered as the change with time in the differential equation At the same time, since it is a stochastic differential equation problem, the mathematical expectation needs to be considered, and finally it can be regarded as a system of ordinary differential equations, and the solution can be obtained as shown in equations (4)-(5):

[0085]

[0086] where, is the feature vector of sample i in the first class is the mathematical expectation after training the m-th layer; is the feature vector of sample j in the second class is the mathematical expectation after training the m-th layer; f(γ,δ): this function captures the influence of the intra-group connection γ and the inter-group connection δ on the expected value of the feature vector, and the values of γ and δ directly affect the form of the f function, thereby affecting the evolution trend of the feature vector, e f(γ,δ) : the use of the exponential function generally represents an exponential change with the increase of the training layer number, which indicates that the mathematical expectation value of the feature vector follows an exponential growth or decay law based on γ and δ, which corresponds to the trend of differentiation or fusion of the feature vector in the deep learning training process; g(γ,δ): this function also captures the influence of the intra-group connection γ and the inter-group connection δ on the expected value of the feature vector, but it is different from f(γ,δ) because they describe feature vectors of different categories, e g(γ,δ) : similarly, the use of the exponential function means that the mathematical expectation value of the feature vector follows another exponential variation law based on γ and δ, which reflects the evolution trend of the feature vector of the second class of samples in the deep learning process.

[0087] Therefore, when the number of training layers is sufficient, the change of the feature vector of the binary classification problem can be regarded as the change trend of a monotonic function over time. Here is a simple consideration, so if the functions f(γ,δ) and g(γ,δ) satisfy certain properties and the feature vector differentiation of the monotonic function gradually becomes obvious, then it can be sure that such a deep learning training process can train a differentiated classification result.

[0088] In the above example of a classification problem, the system uses stochastic differential equations to characterize the training process of deep learning. In binary classification, many simplifications are used, such as constantization of influence coefficients, initial value consideration, etc. However, the system can more complexly characterize the entire deep learning training process during the entire module training process, add initial value consideration to make the training converge, add randomization of influence coefficients, etc., but the system can always establish a stochastic differential equation through the change of the feature vector of each layer to overall characterize the training process of deep learning.

[0089] The embodiment of the application further provides an abnormal transaction classification device. It should be noted that the abnormal transaction classification device of the embodiment of the application can be used to execute the abnormal transaction classification method provided by the embodiment of the application. The abnormal transaction classification device provided by the embodiment of the application is introduced as follows.

[0090] According to the embodiment of the present application, a device for implementing the above-mentioned method for classifying abnormal transactions is further provided, Figure 3 is a schematic diagram of an optional device for classifying abnormal transactions according to the embodiment of the present application, as shown in the figure, the device comprises a receiving unit 301, a conversion unit 302 and a determination unit 303. Figure 3

[0091] Optionally, the receiving unit 301 is configured to receive target transaction data; the conversion unit 302 is configured to convert the target transaction data into a target feature vector; and the determination unit 303 is configured to input the target feature vector into a target model, and obtain a classification result for the target transaction data according to prior knowledge learned by the target model in a model training stage, wherein the target model is a classification model obtained by updating model parameters and verifying based on a stochastic differential equation model, and the stochastic differential equation model is used to map the training process of the target model.

[0092] Optionally, the device for classifying abnormal transactions further comprises a first acquisition unit, a first extraction unit and a first training unit. The first acquisition unit is configured to acquire a historical transaction data set, wherein the historical transaction data set comprises historical transaction data of N users and actual category labels corresponding to each historical transaction data, and N is an integer greater than 1; the first extraction unit is configured to extract at least one feature in each historical transaction data to obtain N feature vectors; and the first training unit is configured to iteratively train an initial model according to the N feature vectors to obtain a target model, wherein the iterative training is used to adjust model parameters of the initial model according to a stochastic differential equation model.

[0093] Optionally, the device for classifying abnormal transactions further comprises a first setting unit and a first construction unit. The first setting unit is configured to set a first value, a second value and random noise based on the N feature vectors, wherein the first value is used to represent the influence degree between samples of the same category, and the second value is used to represent the influence degree between samples of different categories; and the first construction unit is configured to construct a stochastic differential equation model based on the first value, the second value and the random noise.

[0094] ​Optionally, the abnormal transaction classification device further comprises a first solving unit, a first determining unit, a second determining unit, a first screening unit and a first processing unit. The first solving unit solves the N characteristic vectors by a stochastic differential equation model to obtain a target solution set for each characteristic vector, wherein the target solution set is a set of solutions of the characteristic vector at different time points in the stochastic differential equation model; the first determining unit determines the distribution trend of the N target solution sets; the second determining unit determines whether the category corresponding to the region to which each target solution set belongs is the same as the actual category label corresponding to the target solution set if S target solution sets in the N target solution sets are distributed in a first region and T target solution sets are distributed in a second region; wherein T and S are integers greater than or equal to 1, the sum of T and S is N, the first region is used to represent that the target solution set belongs to the abnormal transaction category, and the second region is used to represent that the target solution set belongs to the normal transaction category; the first screening unit screens the characteristics in the N characteristic vectors when the number of target solution sets in the N target solution sets for which the category corresponding to the region to which the target solution set belongs is not the same as the actual category label corresponding to the target solution set is greater than a first preset threshold, and updates the N characteristic vectors based on the screening result; and the first processing unit repeats steps 1 to 4 until it is detected that the number of target solution sets in the N target solution sets for which the category corresponding to the region to which the target solution set belongs is not the same as the actual category label corresponding to the target solution set is less than or equal to the first preset threshold, takes the N characteristic vectors as N first characteristic vectors, and pre-trains the initial model according to the N first characteristic vectors, wherein the pre-training is used to adjust the initial characteristic value of each characteristic of the first characteristic vector based on the stochastic differential equation model.

[0095] Optionally, the first processing unit comprises a first analysis subunit, a first identification subunit, a first adjustment subunit, a first training subunit, and a first processing subunit. The first analysis subunit analyzes the influence of the initial feature value of each feature in the N first feature vectors on the distribution trend of the solution of the random differential equation model to obtain an analysis result. The first identification subunit identifies at least one feature in the first feature vectors whose influence on the distribution trend of the solution is greater than a second preset threshold based on the analysis result to obtain an identification result. The first adjustment subunit adjusts the initial feature value corresponding to each feature in each first feature vector based on the identification result, and takes the adjusted first feature vector as a second feature vector to obtain N second feature vectors. The first training subunit pre-trains the initial model through the N second feature vectors to obtain N classification results, and determines the error value between the classification result corresponding to each second feature vector and the actual class label to obtain N error values. The first processing subunit repeats steps one to five until the number of error values greater than or equal to a third preset threshold is less than H and the N classification results conform to the distribution trend of the solution of the random differential equation model, takes the N second feature vectors as N target feature vectors, and the N target feature vectors are taken as the N target feature vectors, wherein H is an integer greater than or equal to 0.

[0096] Optionally, the first training unit comprises a second training subunit configured to iteratively train the initial model according to the N target feature vectors until, for each target feature vector, the difference between the output result of each layer in the initial model and the solution of the random differential equation model at the corresponding time point of the layer is less than a fourth preset threshold, to obtain a target model, wherein the number of time points for solving in the random differential equation model corresponds to the number of layers in the initial model one by one.

[0097] Optionally, the second training subunit comprises a first processing module, a second processing module, a first determining module, a third processing module and a first adjusting module. The first processing module is configured to input the target feature vector into the initial model, process the target feature vector according to each layer in the initial model, and combine the processing results of each layer corresponding to the target feature vector as a first output result set. The second processing module is configured to input the target feature vector into the random differential equation model for solving, obtain the output result of the target feature vector at each time point, and combine the output result of the target feature vector at each time point as a second output result set. The first determining module is configured to determine the difference between the processing result of each layer in the first output result set and the output result of the layer at the corresponding time point in the second output result set. The third processing module is configured to perform the next round of iterative training if the difference between the processing result of each layer in the first output result set and the output result of the layer at the corresponding time point in the second output result set is less than a fourth preset threshold. The first adjusting module is configured to adjust the model parameters of the initial model if the difference between the processing result of the i-th layer in the first output result set and the output result of the i-th layer at the corresponding time point in the second output result set is greater than or equal to the fourth preset threshold, where i is an integer greater than or equal to 1.

[0098] According to another aspect of the embodiments of the present application, a computer readable storage medium is also provided, and the computer readable storage medium stores a computer program. When the computer program runs, the computer readable storage medium enables an equipment to perform the above-mentioned method for classifying abnormal transactions.

[0099] According to another aspect of the embodiments of the present application, an electronic device is also provided, which comprises one or more processors and a memory. The memory is configured to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors perform the above-mentioned method for classifying abnormal transactions.

[0100] According to another aspect of the embodiments of the present application, a computer program product is also provided, which comprises computer instructions. When the computer instructions are executed by a processor, the steps of the above-mentioned method for classifying abnormal transactions are implemented.

[0101] The serial numbers of the above-mentioned embodiments of the present application are only for description, and do not represent the advantages or disadvantages of the embodiments.

[0102] In the above-mentioned embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0103] In several embodiments provided in the present application, it should be understood that the disclosed technology can be implemented by other ways. Among them, the above-described device embodiments are only schematic, for example, the division of the units can be a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, units or modules, and can be electrical or other forms.

[0104] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed to multiple units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0105] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0106] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the technical solutions or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0107] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method for classifying abnormal transactions, characterized in that: include: Receive target transaction data; Converting the target transaction data into a target feature vector; The target feature vector is input into the target model, and a classification result for the target transaction data is obtained based on the prior knowledge learned by the target model during the model training phase, wherein the target model is updated based on the model parameters and a classification model is obtained by verification based on a stochastic differential equation model, wherein the stochastic differential equation model is used to map the training process of the target model.

2. The abnormal transaction classification method according to claim 1, characterized in that: The target model is trained in the following way: Obtaining a historical transaction data set, wherein the historical transaction data set includes historical transaction data of N users and an actual category label corresponding to each historical transaction data, wherein N is an integer greater than 1; Extract at least one feature from each historical transaction data to obtain N feature vectors; The initial model is iteratively trained according to the N eigenvectors to obtain the target model, wherein the iterative training is used to adjust the model parameters of the initial model according to the stochastic differential equation model.

3. The abnormal transaction classification method according to claim 2, characterized in that: After extracting at least one feature from each historical transaction data to obtain N feature vectors, the method further includes: Based on the N eigenvectors, a first value, a second value, and random noise are set, wherein the first value is used to represent the degree of influence between samples of the same category, and the second value is used to represent the degree of influence between samples of different categories; The stochastic differential equation model is constructed based on the first value, the second value, and the random noise.

4. The abnormal transaction classification method according to claim 3, characterized in that: After constructing the stochastic differential equation model based on the first value, the second value, and the random noise, the method further includes: Step 1: solving the N eigenvectors using the stochastic differential equation model to obtain a target solution set for each eigenvector, wherein the target solution set is a set of solutions of the eigenvector at different time points in the stochastic differential equation model; Step 2: determine the distribution trend of the N target solution sets; Step 3: If there are S target solution sets distributed in the first area and T target solution sets distributed in the second area among the N target solution sets, determine whether the category corresponding to the area to which each target solution set belongs is the same as the actual category label corresponding to the target solution set; Wherein, T and S are both integers greater than or equal to 1, the sum of T and S is N, the first region is used to indicate that the target solution set belongs to the abnormal transaction category, and the second region is used to indicate that the target solution set belongs to the normal transaction category; Step 4: When the number of differences between the category corresponding to the region to which the target solution set belongs and the actual category label corresponding to the target solution set in the N target solution sets is greater than a first preset threshold, filtering the features in the N feature vectors, and updating the N feature vectors based on the filtering results; Step 5: Repeat steps 1 to 4 until it is detected that the number of categories corresponding to the region to which the target solution set belongs in the N target solution sets that are different from the actual category labels corresponding to the target solution set is less than or equal to the first preset threshold, and use the N feature vectors as N first feature vectors, and pre-train the initial model based on the N first feature vectors, wherein the pre-training is used to adjust the initial eigenvalue of each feature in the first feature vector based on the stochastic differential equation model.

5. The abnormal transaction classification method according to claim 4, characterized in that: Pre-training the initial model according to the N first eigenvectors includes: Step 1: Analyze the influence of the initial eigenvalue of each feature in the N first eigenvectors on the distribution trend of the solution of the stochastic differential equation model to obtain an analysis result; Step 2: Based on the analysis result, identify at least one feature in the first eigenvector whose influence on the distribution trend of the solution is greater than a second preset threshold, and obtain an identification result; Step 3: adjusting the initial eigenvalue corresponding to each feature in each first eigenvector based on the recognition result, and using the adjusted first eigenvector as the second eigenvector to obtain N second eigenvectors; Step 4: pre-training the initial model using the N second eigenvectors to obtain N classification results, and determining an error value between the classification result corresponding to each second eigenvector and the actual category label to obtain N error values; Step 5: If there are H error values ​​among the N error values ​​that are greater than the third preset threshold and / or the N classification results do not conform to the distribution trend of the solution results of the stochastic differential equation model, repeat steps 1 to 5 until the number of the N error values ​​that are greater than or equal to the third preset threshold is less than H and the N classification results conform to the distribution trend of the solution results of the stochastic differential equation model, and use the N second eigenvectors as N target eigenvectors, where H is an integer greater than or equal to 0.

6. The abnormal transaction classification method according to claim 5, characterized in that: Iteratively training the initial model according to the N feature vectors to obtain the target model includes: The initial model is iteratively trained according to the N target feature vectors until, for each target feature vector, the difference between the output result of each layer in the initial model and the solution of the stochastic differential equation model at the time point corresponding to the layer is less than a fourth preset threshold, thereby obtaining the target model, wherein the number of time points for solving the stochastic differential equation model corresponds one-to-one to the number of layers in the initial model.

7. The abnormal transaction classification method according to claim 6, characterized in that: Each training iteration consists of the following steps: Inputting a target feature vector into the initial model, processing the target feature vector according to each layer in the initial model, and taking a combination of processing results of each layer corresponding to the target feature vector as a first output result set; Inputting the target feature vector into the stochastic differential equation model for solving, obtaining an output result of the target feature vector at each time point, and taking a combination of the output results of the target feature vector at each time point as a second output result set; Determine a difference between a processing result corresponding to each layer in the first output result set and an output result of the layer at a corresponding time point in the second output result set; If it is detected that the difference between the processing result corresponding to each layer in the first output result set and the output result of the layer at the corresponding time point in the second output result set is less than the fourth preset threshold, the next round of iterative training is performed; If it is detected that the difference between the processing result corresponding to the i-th layer in the first output result set and the output result of the i-th layer at the corresponding time point in the second output result set is greater than or equal to the fourth preset threshold, the model parameters of the initial model are adjusted, where i is an integer greater than or equal to 1.

8. A device for classifying abnormal transactions, characterized in that: include: A receiving unit, configured to receive target transaction data; a conversion unit, configured to convert the target transaction data into a target feature vector; A determination unit is used to input the target feature vector into a target model, and obtain a classification result for the target transaction data based on the prior knowledge learned by the target model during the model training phase, wherein the target model is a classification model updated based on model parameters and verified based on a stochastic differential equation model, wherein the stochastic differential equation model is used to map the training process of the target model.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located executes the abnormal transaction classification method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The device comprises one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the abnormal transaction classification method according to any one of claims 1 to 7.

11. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by a processor, the steps of the abnormal transaction classification method according to any one of claims 1 to 7 are implemented.

Images