Fraud detection methods, devices, electronic devices and readable storage media

By acquiring multiple target data points of fraud incidents from different data sources, performing feature extraction and time series splicing, and utilizing a hidden layer feature model, the problem of insufficient comprehensiveness and accuracy in existing fraud incident identification methods is solved, achieving more accurate fraud incident prediction.

CN117591955BActive Publication Date: 2026-11-10CHINA TELECOM CORP LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311422166.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-10-30
Publication Date
2026-11-10
Estimated Expiration
2043-10-30

AI Technical Summary

Technical Problem

In existing technologies, fraud identification methods ignore the correlation between events during their evolution, resulting in low comprehensiveness and accuracy of identification.

Method used

By acquiring multiple target data related to the target event from different data sources, performing feature extraction and feature vector concatenation, and utilizing time series sets and hidden feature models, the predicted probability of fraud events is obtained.

Benefits of technology

It improves the comprehensiveness and accuracy of fraud identification, and can more fully show the trajectory and nature of the event, thus improving the reliability of probability prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117591955B_ABST
    Figure CN117591955B_ABST
Patent Text Reader

Abstract

This invention provides a method, apparatus, electronic device, and readable storage medium for identifying fraud incidents. The method, for any target event, extracts features from the target data set corresponding to the target event to obtain a target feature vector set; concatenates multiple time vectors corresponding to the time series set with multiple target feature vectors from the target feature vector set to obtain strong time data features corresponding to each target data; the time series set contains target sequence values ​​used to characterize the time order of occurrence of at least two target data; based on the strong time data features, obtains target hidden layer features corresponding to each target data to obtain a target hidden layer feature set; based on the target hidden layer feature set, obtains a prediction probability; the prediction probability is used to characterize the probability that the target event belongs to a fraud event, which can improve the comprehensiveness and accuracy of the fraud incident identification method to a certain extent.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and in particular relates to a method, apparatus, electronic device and readable storage medium for identifying fraud incidents. Background Technology

[0002] Current telecommunications fraud cases are exhibiting some new characteristics, such as the chain-like nature of fraud patterns. Fraud is rarely an isolated event, but rather a complex process involving multiple steps and participants. For example, a fraudulent activity might begin with a phone call, then guide the victim through text messages and websites, ultimately leading to financial loss.

[0003] In related technologies, traditional fraud detection often judges based on data from a single event, neglecting the interrelationships between events as they evolve. For example, judging fraud solely based on phone records or text message content fails to provide a comprehensive understanding of the diversity and evolution of fraudulent activities, resulting in low comprehensiveness and accuracy for such fraud detection methods. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and readable storage medium for identifying fraud incidents, in order to solve the problems of low comprehensiveness and accuracy of fraud incident identification methods.

[0005] To solve the above-mentioned technical problems, the present invention is implemented as follows:

[0006] In a first aspect, the present invention provides a method for identifying fraud incidents, the method comprising:

[0007] For any target event, feature extraction is performed on the target data set corresponding to the target event to obtain a target feature vector set; the target data set contains at least two target data related to the target event obtained from different data sources;

[0008] Multiple time vectors corresponding to the time series set are concatenated with multiple target feature vectors in the target feature vector set to obtain strong time data features corresponding to each target data; the time series set contains target sequence values ​​used to characterize the occurrence time order of the at least two target data.

[0009] Based on the strong temporal data features, the target hidden layer features corresponding to each target data are obtained to obtain the target hidden layer feature set.

[0010] Based on the target hidden layer feature set, a prediction probability is obtained; the prediction probability is used to characterize the probability that the target event belongs to a fraud event.

[0011] Optionally, the step of obtaining target hidden layer features corresponding to each target data based on the strong temporal data features to obtain a target hidden layer feature set includes:

[0012] For any target strong temporal data feature among the plurality of strong temporal data features, a first hidden layer feature corresponding to the target strong temporal data feature is obtained based on the target strong temporal data feature; the first hidden layer feature corresponds to the target sequence value corresponding to the target strong temporal data feature;

[0013] After determining the first hidden layer feature corresponding to each of the target strong temporal data features, for any target sequence value, the first hidden layer feature corresponding to the previous sequence value of the target sequence value is fused with the first hidden layer feature corresponding to the target sequence value to obtain the target hidden layer feature corresponding to the target sequence value.

[0014] Optionally, the at least two target data include: telephone data, SMS data, and network data; the method further includes:

[0015] Obtain telephone data, SMS data, and network data related to the target event from different data sources;

[0016] The telephone data, SMS data, and network data are sorted according to the first occurrence time of the telephone data, the second occurrence time of the SMS data, and the third occurrence time of the network data to obtain the target data set;

[0017] Arrange the first occurrence time, the second occurrence time, and the third occurrence time in chronological order to obtain the time series set.

[0018] Optionally, the step of extracting features from the target data set corresponding to the target event to obtain a target feature vector set includes:

[0019] For any target data in the target dataset, the target data is convolved using a convolutional layer in the feature extraction model to obtain the target features corresponding to the target data.

[0020] The target feature vector set is obtained by performing pooling and fully connected processing on the target features corresponding to each target data in the target data set through the pooling layer and fully connected layer in the feature extraction model.

[0021] Optionally, the time series set includes multiple target series values; the method further includes:

[0022] For any target sequence value in the time series set, the target sequence value is encoded according to the first formula to obtain the time vector corresponding to the target sequence value;

[0023] The first formula includes:

[0024]

[0025] Among them, the The time vector represents the time vector, t represents the target sequence value, constant represents the quantity constant, and d represents the quantity constant. Φ(X) The d represents the number of target data in the target data set. T This represents the dimension of the target data set.

[0026] Optionally, obtaining the prediction probability based on the target hidden layer feature set includes:

[0027] Multiple target hidden layer features in the target hidden layer feature set are concatenated to obtain concatenated features;

[0028] The spliced ​​features are input into the prediction model, and the prediction probability output by the prediction model is obtained based on the spliced ​​features and the second formula.

[0029] The second formula includes:

[0030]

[0031] Where σ is the activation function, W n H represents the weight, H′ represents the concatenation feature, and b n Y represents the bias value, and Y represents the predicted probability.

[0032] Optionally, fusing the first hidden layer feature corresponding to the previous sequence value of the target sequence value with the first hidden layer feature corresponding to the target sequence value to obtain the target hidden layer feature corresponding to the target sequence value includes:

[0033] Based on the first hidden layer feature corresponding to the target sequence value, the first hidden layer feature corresponding to the previous sequence value of the target sequence value, and the third formula, the target hidden layer feature corresponding to the target sequence value is determined.

[0034] The third formula includes:

[0035]

[0036] Wherein, h t The h represents the first hidden layer feature corresponding to the target sequence value. t-1The first hidden layer feature represents the previous sequence value of the target sequence value, and ξ represents a selection factor with a fixed value.

[0037] Secondly, the present invention provides a fraud incident identification device, the device comprising:

[0038] The first extraction module is used to extract features from the target data set corresponding to any target event to obtain a target feature vector set; the target data set includes at least two target data related to the target event obtained from different data sources;

[0039] The first concatenation module is used to concatenate multiple time vectors corresponding to the time series set with multiple target feature vectors in the target feature vector set to obtain strong time data features corresponding to each target data; the time series set includes target sequence values ​​used to characterize the occurrence time order of the at least two target data.

[0040] The first acquisition module is used to acquire the target hidden layer features corresponding to each target data based on the strong temporal data features, and obtain the target hidden layer feature set;

[0041] The second acquisition module is used to acquire a prediction probability based on the target hidden layer feature set; the prediction probability is used to characterize the probability that the target event belongs to a fraud event.

[0042] Thirdly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the above-mentioned fraud event identification method when executing the program.

[0043] Fourthly, the present invention provides a readable storage medium that, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to perform the aforementioned fraud detection method.

[0044] In this embodiment of the invention, by acquiring at least two target data related to a target time from different data sources and performing feature extraction, relevant data from different data sources at different stages of the same target event can be collected. Then, feature extraction is performed on the target data to obtain feature vectors corresponding to target data at different dimensions and stages, resulting in a target feature vector set. Further, by concatenating multiple time vectors corresponding to the time series set with the target feature vectors in the target feature vector set, strong time data features are obtained. Since strong time data features are obtained by concatenating time series sets and target feature vectors, they essentially contain the occurrence time sequence of at least two target data points and the information of the target feature vectors corresponding to at least two target data points. This allows strong time data features to reveal the event evolution process of the target event to a certain extent based on the time series set and target feature vectors. This enables a more comprehensive presentation of the "development trajectory" of the target event and a more accurate capture of the essence of the target event during the subsequent acquisition of target hidden features and prediction probabilities based on strong time data features. The resulting prediction probability regarding whether the target event is a fraud event is more reliable, thus improving the comprehensiveness and accuracy of the fraud event identification method to a certain extent. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the steps of a fraud event identification method provided in an embodiment of the present invention;

[0047] Figure 2 This is a flowchart illustrating the specific steps of a fraud event identification method provided in an embodiment of the present invention;

[0048] Figure 3 This is a structural diagram of a fraud event identification device provided in an embodiment of the present invention;

[0049] Figure 4 This is a structural diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Figure 1 This is a flowchart illustrating the steps of a fraud event identification method provided in an embodiment of the present invention, as follows: Figure 1 As shown, the method may include:

[0052] Step 101: For any target event, extract features from the target data set corresponding to the target event to obtain a target feature vector set; the target data set contains at least two target data related to the target event obtained from different data sources.

[0053] In this embodiment of the invention, the target event can be an event to be identified, such as an event occurring in a scenario where fraud is prevalent. For the target event, at least two target data related to the target event are obtained from different data sources to obtain a target data set. The target data in the target data set is multi-source heterogeneous data. The target data set may include at least two target data arranged in chronological order of occurrence. The different data sources may include data sources corresponding to at least two sub-events involved in the evolution of the target event, such as telephone, SMS, MMS, and network. The event types of the at least two sub-events may include: telephone type, SMS type, network type, etc. Sub-events can be divided based on the main processing actions in the event. Multiple sub-events may belong to the same event type or multiple event types. For example, sub-events may include a user downloading an application software; a user making a payment in the application software, etc., where the two sub-events belong to the same event type, namely, network type. The at least two target data related to the target event may include: telephone data, SMS data, network data, etc. A feature extraction model is used to extract features from the target data set corresponding to the target time to obtain a target feature vector set. The feature extraction model can be a model that has been trained in advance based on a sample dataset. The feature extraction model can be a convolutional neural network, and the architecture of the feature extraction model can include convolutional layers, pooling layers, and fully connected layers.

[0054] For example, suppose the evolution of a target event includes: 1) a phone call inquiring about funding needs; 2) sending an SMS with an application link recommending the application; 3) guiding the user to register for the application and recommending services; 4) the user purchasing the service. In this event evolution, the different data sources involved in the target event can include telephone, SMS, and the internet. Target data is obtained from these data sources, including: obtaining telephone data, which may include call duration, caller information, called party information, and call content; obtaining SMS data, which may include SMS reception time, sender information, and SMS content; and obtaining internet data, which may include user behavior data related to the application and application-related information.

[0055] Step 102: Concatenate multiple time vectors corresponding to the time series set with multiple target feature vectors in the target feature vector set to obtain strong time data features corresponding to each target data; the time series set contains target sequence values ​​used to characterize the occurrence time order of the at least two target data.

[0056] In this embodiment of the invention, while acquiring at least two target data sets, the occurrence times of at least two target data sets are also acquired. The at least two target data sets are then sorted based on their occurrence times to obtain a target data set. Furthermore, a target sequence value is determined for each target data set according to the chronological order of its occurrence times. The target sequence value characterizes the occurrence order of the target data corresponding to that target sequence value within the at least two target data sets. Based on the at least two sequence values ​​corresponding to the at least two target data sets, a time series set is determined. The time series set corresponds to the target data set, and correspondingly, it also corresponds to the target feature vector set.

[0057] Encoding multiple target sequence values ​​in a time series set yields multiple time vectors. Following the correspondence between target sequence values ​​and target feature vectors in a target feature vector set, these time vectors are concatenated with the target feature vectors in the target feature vector set to obtain the strong temporal data features corresponding to each target data point. Strong temporal features integrate the time vector corresponding to the target data and the target feature vectors characterizing the target data, giving them a temporal sequence quality, facilitating subsequent deduction of event evolution. In other words, one target data point can correspond to one target sequence value, representing the order in which the target data occurs among at least two target data points; one target data point can correspond to one target feature vector, obtained through feature extraction; one target data point can correspond to one time vector, obtained by encoding the target sequence value; and one target data point can correspond to one strong temporal data feature, obtained by concatenating the time vector and the target feature vector. It can be understood that any two of the target sequence value, target feature vector, time vector, and strong temporal data feature corresponding to a single target data point can be mutually correspondent.

[0058] For example, suppose the target dataset is X. The data source methods are defined as {tel, sms, net, mms, ...}, where tel represents telephone, sms represents SMS, net represents the internet, and mms represents MMS. The time series set includes T = {1, 2, 3, ..., t}, and the target sequence value t represents the target data. In the target data set, the occurrence time is t, i.e., the target data... The target sequence value corresponding to the order of occurrence is t. Accordingly, it can be understood that the target data... The corresponding order of occurrence is 1, meaning that in the entire evolution of the target event, the event whose data source is a telephone call (which could be the content of an event communicated via telephone) occurs first. Each target sequence value in the time series set T is encoded to obtain multiple time vectors. For any given target sequence value, the time vector corresponding to that target sequence value is concatenated with the target feature vector corresponding to that target sequence value to obtain the strong time data feature corresponding to that target sequence value.

[0059] Step 103: Based on the strong temporal data features, obtain the target hidden layer features corresponding to each target data to obtain the target hidden layer feature set.

[0060] In this embodiment of the invention, a pre-trained transformer sequence model can be used to obtain the target hidden features corresponding to any target data based on the strong temporal data features. Based on the target hidden features corresponding to the strong temporal data features of each target data, a target hidden feature set is determined. The target hidden features are used to characterize the deeper features obtained after fusing temporal features into the target data.

[0061] Step 104: Based on the target hidden layer feature set, obtain the prediction probability; the prediction probability is used to characterize the probability that the target event belongs to a fraud event.

[0062] In this embodiment of the invention, a prediction probability can be obtained based on the target hidden layer feature set using a pre-trained prediction model. This prediction probability can be used to characterize the probability that the target time belongs to a fraudulent time. It is understood that the prediction model can be a Multilayer Perceptron (MLP) model, and the prediction model can be pre-trained based on a sample dataset to be trained prediction model. Specific training processes can be found in existing technologies, and will not be elaborated upon here.

[0063] In summary, in this embodiment of the invention, by acquiring at least two target data points related to a target time from different data sources and performing feature extraction, relevant data from different data sources at different stages of the same target event can be collected. Then, feature extraction is performed on the target data to obtain feature vectors corresponding to target data at different dimensions and stages, resulting in a target feature vector set. Furthermore, by concatenating multiple time vectors corresponding to the time series set with the target feature vectors in the target feature vector set, strong time data features are obtained. Since strong time data features are obtained by concatenating the time series set with the target feature vectors, they essentially contain the occurrence time sequence of at least two target data points and the information of the target feature vectors corresponding to at least two target data points. This allows strong time data features to reveal the event evolution process of the target event to a certain extent based on the time series set and the target feature vectors. This enables a more comprehensive presentation of the "development trajectory" of the target event and a more accurate capture of the essence of the target event during the subsequent acquisition of target hidden features and prediction probabilities based on strong time data features. The resulting prediction probability regarding whether the target event is a fraud event is more reliable, thus improving the comprehensiveness and accuracy of the fraud event identification method to a certain extent.

[0064] Optionally, step 103 may include the following steps:

[0065] Step 201: For any target strong time data feature among the plurality of strong time data features, obtain the first hidden layer feature corresponding to the target strong time data feature based on the target strong time data feature; the first hidden layer feature corresponds to the target sequence value corresponding to the target strong time data feature.

[0066] In this embodiment of the invention, for any target strong temporal data feature among multiple strong temporal data features, a first hidden layer feature corresponding to the target strong temporal data feature is obtained using a transformer sequence model. The first hidden layer feature represents the feature corresponding to the target feature vector of the target data after fusing temporal characteristics. Since the target strong temporal data feature corresponds to a target sequence value in the time series set, the first hidden layer feature obtained based on the target strong temporal data feature also corresponds to that target sequence value. That is, each target sequence value in the time series set corresponds to a first hidden layer feature obtained based on the strong temporal data feature.

[0067] Specifically, based on the strong temporal data characteristics of any target, a first autocorrelation matrix is ​​obtained through linear transformation using a transformer sequence model. The specific formula can be:

[0068]

[0069] in, W is used to represent the strong temporal data characteristics of the target. Q Used to represent a weight matrix whose parameters are the optimal matrix parameters obtained through parameter tuning during the training of the transformer sequence model, and Q is used for the first autocorrelation matrix.

[0070] Based on the attention mechanism network (Multi-Head Attention) in the transformer sequence model, and combining the transpose of the first autocorrelation matrix, the second autocorrelation matrix, and the third autocorrelation matrix, the attention score corresponding to the target strong temporal data features is obtained. The specific formula can be:

[0071]

[0072] in, Q is used to represent the attention score corresponding to the target's strong temporal data features. T The matrix Q is used to represent the transpose of the first autocorrelation matrix Q, K is used to represent the second autocorrelation matrix, V is used to represent the third autocorrelation matrix, and d is used to represent the dimension of the second autocorrelation matrix.

[0073] The second autocorrelation matrix K and the third autocorrelation matrix V can be determined by the following formulas:

[0074]

[0075] Among them, h t W is used to represent the strong temporal data characteristics of the target. K W V Used to represent a weight matrix whose parameters are the optimal matrix parameters obtained through parameter tuning during the training of the transformer sequence model; II represents the connection operation.

[0076] After determining the attention score Then, based on the activation function, attention score, and weight matrix, the first hidden layer feature corresponding to the target strong temporal data feature is determined. The specific formula can be:

[0077]

[0078] Among them, h t The first hidden layer feature is represented by ReLU, the activation function is represented by ReLU, and the weight matrices are represented by W1 and W2. The parameters of the weight matrices are the optimal matrix parameters obtained by parameter adjustment during the training of the transformer sequence model.

[0079] Understandably, since the model's parameters are randomly initialized, W Q W K W V W1 and W2 are obtained by iteratively updating the parameters of the transformer sequence model through backpropagation algorithm and optimizer, so that the transformer sequence model can adapt to the training data and minimize the loss function.

[0080] Step 202: After determining the first hidden layer feature corresponding to each of the target strong temporal data features, for any target sequence value, the first hidden layer feature corresponding to the previous sequence value of the target sequence value is fused with the first hidden layer feature corresponding to the target sequence value to obtain the target hidden layer feature corresponding to the target sequence value.

[0081] In this embodiment of the invention, after determining the first hidden layer feature corresponding to each target strong temporal data feature, for any target sequence value in the time series set, the target hidden layer feature corresponding to the target sequence value is obtained and determined based on the first hidden layer feature corresponding to the target sequence value and the first hidden layer feature corresponding to the previous sequence value of the target sequence value. At this time, the target hidden layer feature is equivalent to combining some temporal features of the entire target event in the event evolution process, so that the target hidden layer feature can learn the development process of the target event more comprehensively.

[0082] Optionally, step 202 may include the following steps:

[0083] Step 2021: Based on the first hidden layer feature corresponding to the target sequence value, the first hidden layer feature corresponding to the previous sequence value of the target sequence value, and the third formula, determine the target hidden layer feature corresponding to the target sequence value.

[0084] In this embodiment of the invention, the target hidden layer features can be determined by the following formula:

[0085]

[0086] Among them, h t h represents the first hidden layer feature corresponding to the target sequence value. t-1 ξ represents the first hidden layer feature corresponding to the previous sequence value of the target sequence value, and ξ represents the selection factor with a fixed value.

[0087] In other words, the target hidden features can be obtained by combining the first hidden features with a portion of the first hidden features corresponding to the preceding sub-event in the development process of the target event. The target hidden features are equivalent to combining the features of the preceding sub-event in the evolution of the target event with the temporal characteristics of the target data, taking into account the event development characteristics, increasing the path of information transmission and flow, and improving the feature representation capability of the target hidden features.

[0088] Optionally, the at least two target data may include: telephone data, SMS data, and network data. Embodiments of the present invention may include the following steps:

[0089] Step 301: Obtain telephone data, SMS data, and network data related to the target event from different data sources.

[0090] In this embodiment of the invention, telephone data, SMS data, and network data related to the target event are obtained from telephone, SMS, and network data sources. Telephone data may include: call duration, caller information, called party information, call content, etc.; SMS data may include: SMS reception time, SMS sender information, SMS content, etc.; network data may include: user behavior data corresponding to application software operations, application software-related information, etc.

[0091] Step 302: Sort the telephone data, the SMS data, and the network data according to the first occurrence time of the telephone data, the second occurrence time of the SMS data, and the third occurrence time of the network data to obtain the target data set.

[0092] In this embodiment of the invention, during the acquisition of telephone data, SMS data, and network data, the occurrence time of each data acquisition corresponds to a specific time. The occurrence time can refer to the execution time of the main processing action corresponding to the target data. Specifically, the first occurrence time of telephone data can be the time of the main processing action (answering a call), the second occurrence time of SMS data can be the time of the main processing action (receiving an SMS message), and the third occurrence time of network data can be the execution time corresponding to the main processing action (such as payment or download). The telephone data, SMS data, and network data are sorted according to the first, second, and third occurrence times to obtain the target data set.

[0093] Step 303: Arrange the first occurrence time, the second occurrence time, and the third occurrence time in chronological order to obtain the time series set.

[0094] In this embodiment of the invention, a target sequence value is determined for each target data according to the chronological order of occurrence of at least two target data, namely, the first occurrence time, the second occurrence time, and the third occurrence time. The target sequence value is used to characterize the occurrence order of the target data corresponding to the target sequence value among the at least two target data. For example, if telephone data A and B, SMS data C, and network data D and E are obtained, and the first occurrence time corresponding to A is 7:00, the first occurrence time corresponding to B is 9:00, the second occurrence time corresponding to C is 7:30, the third occurrence time corresponding to D is 8:00, and the third occurrence time corresponding to E is 8:30, then according to the chronological order, the target sequence value corresponding to A is determined to be 1, the target sequence value corresponding to C is determined to be 2, the target sequence value corresponding to D is determined to be 3, the target sequence value corresponding to E is determined to be 4, and the target sequence value corresponding to B is determined to be 5. Then, the target data set X = (A, C, D, E, B) and the time series set T = (1, 2, 3, 4, 5) can be obtained.

[0095] In this embodiment of the invention, by obtaining the occurrence time of the target data, sorting the target data according to the chronological order of their occurrence times, and determining the sequence value corresponding to the target data, a target data set and a time series set corresponding to the target data set can be obtained. The target data set takes into account the occurrence time order of each target data corresponding to the target event, which further improves the accuracy of fraud identification in the subsequent fraud identification process.

[0096] Optionally, step 101 may include the following steps:

[0097] Step 401: For any target data in the target data set, perform convolution processing on the target data using the convolutional layer in the feature extraction model to obtain the target features corresponding to the target data.

[0098] In this embodiment of the invention, the feature extraction model may include convolutional layers, pooling layers, and fully connected layers. The target dataset is input into the feature extraction model. For any target data point in the dataset, the convolutional layers in the feature extraction model perform convolution processing on that target data, obtaining the target features corresponding to the target data. The specific formula can be as follows:

[0099]

[0100] in, Used to represent target data, The w is used to represent the target features corresponding to the target data, conv(·) is used to represent the convolution operation, and w l This is used to represent the weight parameters, which are determined during the training of the feature extraction model. l This is the bias value, which is determined during the continuous optimization of the feature extraction model during training.

[0101] After performing convolution on each target data point in the target dataset, a target feature set containing multiple target features is obtained. The specific formula can be:

[0102]

[0103] Where C(X) represents the target feature set, methods represents the data source, and T represents the time series set.

[0104] Step 402: The target features corresponding to each target data in the target data set are processed by pooling and fully connected layers in the feature extraction model to obtain the target feature vector set.

[0105] In this embodiment of the invention, after obtaining the target features corresponding to each target data in the target data set, the target features are pooled and fully connected through the pooling layer and fully connected layer in the feature extraction model to obtain a target feature vector set.

[0106] In this embodiment of the invention, feature extraction is performed on multiple target data using a feature extraction model to obtain the data features of the target data, providing a data foundation for subsequent fraud event identification.

[0107] Optionally, embodiments of the present invention may further include the following steps:

[0108] Step 501: For any target sequence value in the time series set, encode the target sequence value according to the first formula to obtain the time vector corresponding to the target sequence value.

[0109] In this embodiment of the invention, for any target sequence value in the time series set, the target sequence value is encoded according to a first formula to obtain the time vector corresponding to the target sequence value. The first formula can be as follows:

[0110]

[0111] Among them, the The time vector represents the time vector, t represents the target sequence value, constant represents the quantity constant, and d represents the quantity constant. Φ(X) The d represents the number of target data in the target data set. T This represents the dimension of the target data set.

[0112] Optionally, step 104 may include the following steps:

[0113] Step 1041: Concatenate multiple target hidden layer features in the target hidden layer feature set to obtain concatenated features.

[0114] In this embodiment of the invention, after obtaining the target hidden layer features corresponding to each target data, the multiple target hidden layer features are concatenated to obtain the concatenated features. The specific formula may be:

[0115]

[0116] Where H represents the splicing feature. It is used to represent multiple target hidden layer features, and concat is used to represent the concatenation function.

[0117] Step 1042: Input the spliced ​​features into the prediction model, and obtain the prediction probability output by the prediction model based on the spliced ​​features and the second formula.

[0118] In this embodiment of the invention, the concatenated features are input into the prediction model, and the prediction probability is determined using the concatenated features and the second formula, which serves as the output of the prediction model. The second formula may specifically include:

[0119]

[0120] Where σ is the activation function, W n The weights are represented by b, and their values ​​can be determined through continuous optimization of the model during training. H′ represents the concatenated features. n Y represents the bias value, which can be determined by continuously optimizing the model during the training process of the prediction model.

[0121] For example, for a target data point of a target event, the target sequence value corresponding to the target data is used to characterize the chronological order of the occurrence time of the target data among at least two target data points. Encoding the target sequence value yields the time vector corresponding to the target data. Feature extraction of the target data yields the target feature vector. Concatenating the target feature vector with the time vector yields the strong temporal data feature corresponding to the target data. Based on the strong temporal data feature, the first hidden layer feature corresponding to the target data can be obtained. Then, the first hidden layer feature corresponding to the target data is fused with the first hidden layer feature corresponding to the previous target sequence value (i.e., the first hidden layer feature corresponding to the previous target data point in the target data set) to obtain the target hidden layer feature corresponding to the target data.

[0122] For example, Figure 2 A flowchart illustrating the specific steps of a fraud incident identification method is shown, such as... Figure 2 As shown, multiple target data points and their corresponding target sequence values ​​are obtained from different data sources. These target data points and their sequence values ​​are then input into a feature extraction model to obtain a set of target feature vectors. Based on the time vectors obtained by encoding the multiple target sequence values ​​in the time series set and the target feature vector set, a strong time data feature set is obtained. This set is then input into a transformer sequence model to obtain the target hidden layer features. Based on these hidden layer features, the prediction probability is obtained through a prediction model.

[0123] Figure 3 This is a structural diagram of a fraud event identification device provided in an embodiment of the present invention. The device 60 may include:

[0124] The first extraction module 601 is used to extract features from the target data set corresponding to any target event to obtain a target feature vector set; the target data set includes at least two target data related to the target event obtained from different data sources;

[0125] The first splicing module 602 is used to splice multiple time vectors corresponding to the time series set with multiple target feature vectors in the target feature vector set to obtain strong time data features corresponding to each target data; the time series set includes target sequence values ​​used to characterize the occurrence time order of the at least two target data.

[0126] The first acquisition module 603 is used to acquire the target hidden layer features corresponding to each target data based on the strong time data features, and obtain the target hidden layer feature set;

[0127] The second acquisition module 604 is used to acquire a prediction probability based on the target hidden layer feature set; the prediction probability is used to characterize the probability that the target event belongs to a fraud event.

[0128] Optionally, the first acquisition module may include:

[0129] The first acquisition submodule is used to acquire, based on any target strong temporal data feature among the plurality of strong temporal data features, a first hidden layer feature corresponding to the target strong temporal data feature; the first hidden layer feature corresponds to the target sequence value corresponding to the target strong temporal data feature.

[0130] The first fusion module is used to, after determining the first hidden layer feature corresponding to each of the target strong temporal data features, fuse the first hidden layer feature corresponding to the previous sequence value of the target sequence value with the first hidden layer feature corresponding to the target sequence value for any target sequence value, so as to obtain the target hidden layer feature corresponding to the target sequence value.

[0131] Optionally, the at least two target data include: telephone data, SMS data, and network data; the device 60 may include:

[0132] The third acquisition module is used to acquire telephone data, SMS data, and network data related to the target event from different data sources;

[0133] The first sorting module is used to sort the telephone data, SMS data, and network data according to the first occurrence time of the telephone data, the second occurrence time of the SMS data, and the third occurrence time of the network data to obtain a target data set.

[0134] The second sorting module is used to arrange the first occurrence time, the second occurrence time, and the third occurrence time in chronological order to obtain the time series set.

[0135] Optionally, the first extraction module 601 may include:

[0136] The first processing module is used to perform convolution processing on any target data in the target data set using the convolutional layer in the feature extraction model to obtain the target features corresponding to the target data.

[0137] The second processing module is used to perform pooling and full-connection processing on the target features corresponding to each target data in the target data set through the pooling layer and the fully connected layer in the feature extraction model, respectively, to obtain the target feature vector set.

[0138] Optionally, the time series set includes multiple target sequence values; the device 60 may include:

[0139] The first encoding module is used to encode any target sequence value in the time series set according to a first formula to obtain a time vector corresponding to the target sequence value;

[0140] The first formula includes:

[0141]

[0142] Among them, the The time vector represents the time vector, t represents the target sequence value, constant represents the quantity constant, and d represents the quantity constant. Φ(X) The d represents the number of target data in the target data set. T This represents the dimension of the target data set.

[0143] Optionally, the second acquisition module 604 may include:

[0144] The second splicing module is used to splice multiple target hidden layer features in the target hidden layer feature set to obtain spliced ​​features;

[0145] The second acquisition submodule is used to input the spliced ​​features into the prediction model, and based on the spliced ​​features and the second formula, obtain the prediction probability output by the prediction model.

[0146] The second formula includes:

[0147]

[0148] Where σ is the activation function, W n H represents the weight, H′ represents the concatenation feature, and b n Y represents the bias value, and Y represents the predicted probability.

[0149] Optionally, the first fusion module may include:

[0150] The first determining module is used to determine the target hidden layer feature corresponding to the target sequence value based on the first hidden layer feature corresponding to the target sequence value, the first hidden layer feature corresponding to the previous sequence value of the target sequence value, and the third formula.

[0151] The third formula includes:

[0152]

[0153] Wherein, h t The h represents the first hidden layer feature corresponding to the target sequence value.t-1 The first hidden layer feature represents the previous sequence value of the target sequence value, and ξ represents a selection factor with a fixed value.

[0154] In summary, the fraud event identification device provided in this embodiment of the invention can collect relevant data from different data sources related to the target time by acquiring at least two target data related to the target time from different data sources and performing feature extraction. This allows for the collection of relevant data from different data sources at different stages of the same target event, followed by feature extraction of the target data to obtain feature vectors corresponding to target data at different dimensions and stages, thus obtaining a target feature vector set. Furthermore, by concatenating multiple time vectors corresponding to the time series set with the target feature vectors in the target feature vector set, strong time data features are obtained. Since strong time data features are obtained by concatenating the time series set with the target feature vectors, they essentially contain the information of the occurrence time sequence of at least two target data points and the target feature vectors corresponding to at least two target data points. This allows strong time data features to reveal the event evolution process of the target event to a certain extent based on the time series set and the target feature vectors. This enables a more comprehensive presentation of the "development trajectory" of the target event and a more accurate capture of the essence of the target event during the subsequent acquisition of target hidden features and prediction probability based on the strong time data features. The resulting prediction probability of whether the target event is a fraud event is more reliable, thus improving the comprehensiveness and accuracy of the fraud event identification method to a certain extent.

[0155] The present invention also provides an electronic device, see [link to relevant documentation]. Figure 4 The system includes: a processor 701, a memory 702, and a computer program 7021 stored in the memory and executable on the processor. When the processor executes the program, it implements the fraud event identification method of the foregoing embodiments.

[0156] The present invention also provides a readable storage medium, wherein when the instructions in the storage medium are executed by the processor of an electronic device, the electronic device is able to perform the fraud event identification method of the foregoing embodiments.

[0157] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0158] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0159] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0160] Similarly, it should be understood that, in order to simplify the invention and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the invention, various features of the invention are sometimes grouped together in a single embodiment, figure, or description thereof. However, this disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim. Rather, as reflected in the following claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into this detailed description, wherein each claim itself is a separate embodiment of the invention.

[0161] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0162] The various component embodiments of the present invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components in the sorting device according to the present invention. The present invention can also be implemented as a device or apparatus program for performing part or all of the methods described herein. Such a program implementing the present invention can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0163] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

[0164] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0165] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0166] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying fraud incidents, characterized in that, The method includes: For any target event, feature extraction is performed on the target data set corresponding to the target event to obtain a target feature vector set; the target data set contains at least two target data related to the target event obtained from different data sources; Multiple time vectors corresponding to the time series set are concatenated with multiple target feature vectors in the target feature vector set to obtain strong time data features corresponding to each target data; the time series set contains target sequence values ​​used to characterize the occurrence time order of the at least two target data. Based on the strong temporal data features, the target hidden layer features corresponding to each target data are obtained to obtain the target hidden layer feature set. Based on the target hidden layer feature set, a prediction probability is obtained; the prediction probability is used to characterize the probability that the target event belongs to a fraud event; The step of obtaining target hidden layer features corresponding to each target data based on the strong temporal data features, to obtain a target hidden layer feature set, includes: For any target strong temporal data feature among the plurality of strong temporal data features, a first hidden layer feature corresponding to the target strong temporal data feature is obtained based on the target strong temporal data feature; the first hidden layer feature corresponds to the target sequence value corresponding to the target strong temporal data feature; After determining the first hidden layer feature corresponding to each of the target strong temporal data features, for any target sequence value, the first hidden layer feature corresponding to the previous sequence value of the target sequence value is fused with the first hidden layer feature corresponding to the target sequence value to obtain the target hidden layer feature corresponding to the target sequence value. The different data sources include: telephone, SMS, MMS, and the Internet.

2. The method according to claim 1, characterized in that, The at least two target data include: telephone data, SMS data, and network data; the method further includes: Obtain telephone data, SMS data, and network data related to the target event from different data sources; The telephone data, SMS data, and network data are sorted according to the first occurrence time of the telephone data, the second occurrence time of the SMS data, and the third occurrence time of the network data to obtain the target data set; Arrange the first occurrence time, the second occurrence time, and the third occurrence time in chronological order to obtain the time series set.

3. The method according to claim 1, characterized in that, The step of extracting features from the target data set corresponding to the target event to obtain a target feature vector set includes: For any target data in the target dataset, the target data is convolved using a convolutional layer in the feature extraction model to obtain the target features corresponding to the target data. The target feature vector set is obtained by performing pooling and fully connected processing on the target features corresponding to each target data in the target data set through the pooling layer and fully connected layer in the feature extraction model.

4. The method according to claim 1, characterized in that, The time series set includes multiple target sequence values; the method further includes: For any target sequence value in the time series set, the target sequence value is encoded according to the first formula to obtain the time vector corresponding to the target sequence value; The first formula includes: Among them, the The time vector represents the target sequence value, and constant represents the quantity constant. This indicates the number of target data items in the target data set. This represents the dimension of the target data set.

5. The method according to claim 1, characterized in that, The step of obtaining the prediction probability based on the target hidden layer feature set includes: Multiple target hidden layer features in the target hidden layer feature set are concatenated to obtain concatenated features; The spliced ​​features are input into the prediction model, and the prediction probability output by the prediction model is obtained based on the spliced ​​features and the second formula. The second formula includes: in, For activation function, H represents the weight. This indicates the splicing feature. Y represents the bias value, and Y represents the predicted probability.

6. The method according to claim 1, characterized in that, The step of fusing the first hidden layer feature corresponding to the previous sequence value of the target sequence value with the first hidden layer feature corresponding to the target sequence value to obtain the target hidden layer feature corresponding to the target sequence value includes: Based on the first hidden layer feature corresponding to the target sequence value, the first hidden layer feature corresponding to the previous sequence value of the target sequence value, and the third formula, the target hidden layer feature corresponding to the target sequence value is determined. The third formula includes: Among them, the The first hidden layer feature corresponding to the target sequence value is represented by the following: The first hidden layer feature represents the value of the previous sequence corresponding to the target sequence value. This represents a selection factor with a fixed value.

7. A fraud detection device, characterized in that, The device includes: The first extraction module is used to extract features from the target data set corresponding to any target event to obtain a target feature vector set; the target data set contains at least two target data related to the target event obtained from different data sources; The first concatenation module is used to concatenate multiple time vectors corresponding to the time series set with multiple target feature vectors in the target feature vector set to obtain strong time data features corresponding to each target data; the time series set includes target sequence values ​​used to characterize the occurrence time order of the at least two target data. The first acquisition module is used to acquire the target hidden layer features corresponding to each target data based on the strong temporal data features, and obtain the target hidden layer feature set; The second acquisition module is used to acquire a prediction probability based on the target hidden layer feature set; the prediction probability is used to characterize the probability that the target event belongs to a fraud event; The first acquisition module includes: The first acquisition submodule is used to acquire, based on any target strong temporal data feature among the plurality of strong temporal data features, a first hidden layer feature corresponding to the target strong temporal data feature; the first hidden layer feature corresponds to the target sequence value corresponding to the target strong temporal data feature. The first fusion module is used to, after determining the first hidden layer feature corresponding to each of the target strong temporal data features, fuse the first hidden layer feature corresponding to the previous sequence value of the target sequence value with the first hidden layer feature corresponding to the target sequence value for any target sequence value, so as to obtain the target hidden layer feature corresponding to the target sequence value. The different data sources include: telephone, SMS, MMS, and the Internet.

8. An electronic device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the fraud event identification method as described in any one of claims 1-6.

9. A readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device is able to perform the fraud event identification method according to any one of claims 1-6.

Citation Information

Patent Citations

  • Method and apparatus for trainning a fraud detection model and a method and apparatus for fraud detection

    CN109345260A

  • Telecommunication fraud event identification method and device, and computer readable storage medium

    CN114331473A