Risk number identification method based on adaptive learning
By optimizing the positive sample labeling and feature coefficients of unlabeled samples through adaptive learning and a soft threshold objective function, the error diffusion problem of risk identification models in existing technologies is solved, improving the accuracy and recall of risk identification and generating a number identification model for suspected risks.
Patent Information
- Application Number
- CN202511077174.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-01
- Publication Date
- 2025-12-30
AI Technical Summary
Existing risk number identification models, in the case of PU-Learning, suffer from classification errors because there are positive samples among the unlabeled samples. Directly using these positive samples as negative samples to train the classifier leads to classification errors. Furthermore, the identification effect is heavily dependent on prior knowledge of the samples, making the model's screening results susceptible to influence. Classification errors result in diffuse errors.
An adaptive learning approach is adopted, which constructs a soft threshold objective function and a cross-entropy loss function, and updates the model parameters through multiple iterations to optimize the positive sample labeling and feature coefficients of unlabeled samples. Interconnected behavioral data is used as features to improve the data feature representation capability and construct a risk number behavior recognition model.
It improves the accuracy and recall of risk number identification, effectively reduces errors, enhances the model's identification performance, prevents overfitting, generates a list of high-risk suspected risk numbers, and provides early warnings.
Smart Images
Figure CN121239780A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of big data and risk identification, and particularly relates to a risk number identification method based on adaptive learning. BACKGROUND
[0002] The risk number identification model scene usually faces the situation of PU-Learning: because all positive samples cannot be exhausted, part of the positive sample data in the unlabeled sample exists, and if it is directly used as a negative sample, the prediction accuracy of the model will be seriously reduced.
[0003] PU-Learning is a research direction of semi-supervised learning, and the basic concept of PU-Learning is that only a small number of positive samples are labeled, and no negative samples are labeled, so the unlabeled samples contain both positive samples and negative samples. In the training process, training the classifier by taking the unlabeled sample as a negative sample may lead to classification errors, because the negative sample may contain positive samples. There are two solutions: 1. Heuristically find reliable negative samples from unlabeled samples to train a binary classifier. 2. Train the classifier by taking the unlabeled sample as a negative sample, and because the negative sample contains positive samples, the incorrect label assignment leads to classification errors. The classifier will give a score (probability value) for each item, and usually the score of the positive sample is higher than that of the negative sample, so for those unlabeled items, the higher the score, the more likely it is a positive sample.
[0004] Currently, such risk number identification has the following problems: 1. If the unlabeled sample is directly used as a negative sample, the problem of this method is that the classification effect seriously depends on the prior knowledge of the sample. 2. By training the classifier to try to find positive sample data from the unlabeled sample, and then using the updated positive and negative sample data to build a model, the process of screening reliable positive samples from the unlabeled sample seriously depends on the reliable negative sample identification model, and the quality of the model determines the screening result, which directly affects the effect of the subsequent risk number identification model; 3. The method of using the classifier to identify reliable positive sample data from the unlabeled data will set the identification result to be a positive sample 1 or a negative sample 0, and if the classification is wrong, it will cause errors to spread. SUMMARY
[0005] The purpose of the present application is to provide a risk number identification method based on adaptive learning, which effectively improves the identification effect of the risk number by adaptive learning parameters and multiple iterations.
[0006] The technical scheme adopted by the present application is:
[0007] A risk number identification method based on adaptive learning, comprising the following steps:
[0008] S1, collect encrypted call behavior data of suspected risk behavior numbers;
[0009] S2, extract call behavior features of suspected risk behavior numbers from call behavior data, extend data statistical features of suspected risk behavior numbers through statistical methods, and perform data cleaning and filling preprocessing on the extended data to obtain a data feature set;
[0010] S3, construct a soft threshold objective function as a risk number behavior recognition model, and use the preprocessed data feature set data as the features X of the objective function to calculate the target value y, and combine the cross-entropy loss function to iteratively update and optimize the parameters of the risk number behavior recognition model, to obtain the final risk number behavior recognition model;
[0011] S4, use the final risk number behavior recognition model to recognize the data feature set, generate a suspected risk number list and issue a warning for the suspected risk number.
[0012] Further, in S1, the encrypted call behavior data of the number is obtained based on the time interval; the call behavior data includes voice, SMS, and interconnection behavior features.
[0013] Further, in S1, the number call behavior data collection refers to taking a picture of the call number through a data collection system and big data technology, extracting the call behavior feature data of the number, and encrypting and desensitizing the key fields through encryption and desensitization technology, while realizing the automatic collection of target numbers and behavior features for subsequent model persistence.
[0014] Further, S2 specifically includes the following steps:
[0015] S2-1, obtain the voice call sheet of the number communication, including the number of calls, the call duration, the number of calls, the call time, the call dispersion, the call interval, the call frequency, and the two-way call behavior; wherein the call dispersion refers to the ratio of the number of calls to the number of opposite numbers;
[0016] S2-2, obtain the SMS call sheet of the number communication, including the number of SMS, the number of SMS, the call time, the dispersion, the SMS interval, the SMS frequency, and the two-way SMS; wherein the dispersion refers to the ratio of the number of SMS to the number of SMS opposite numbers, and the two-way SMS refers to the number of numbers with two-way SMS behavior and the proportion;
[0017] S2-3, obtain the interconnection call sheet of the number communication, including the number of calls, the number of calls, the proportion of calls, the number of abnormal call sheets, the connection rate, the call dispersion, and the call frequency of abnormal call sheets; wherein the abnormal call sheet refers to the number of times the opposite number is not connected, does not respond, and is not in the service range, and the number ratio between the abnormal call sheet and the normal call sheet;
[0018] S2-4, expanding the time interval-based data statistical features by statistical functions, including maximum value max, minimum value min, unique value unique, average value mean, quartile quartile;
[0019] S2-5, filling the different missing data according to the business features;
[0020] Further, the interconnection call record in S2-3 represents all call behaviors between numbers, and the interconnection call record includes the number behavior data of missed calls, non-responses, and numbers out of service range. The present application introduces the interconnection call record to more truly express the distribution and characteristics of data features.
[0021] Further, S3 specifically includes the following steps:
[0022] S3-1, setting a marking matrix Y, the marking matrix Y is a Boolean matrix composed of 0-1 elements, which is used to update only the target samples in Y that are not marked each time the model coefficient is iteratively calculated; wherein the value 1 indicates that the element in Y at the corresponding position is a marked sample, and the value 0 indicates that the element in Y at the corresponding position is an unmarked sample;
[0023] S3-2, designing a soft threshold target function based on the marking matrix Y as the expression of the prediction model, and the specific expression is:
[0024] y*Y=σ(z)+λ|W| 2 =σ(WX T )+λ|W| 2
[0025] Wherein, y represents the target value function, y takes the value 0 indicating a negative sample, and y takes the value 1 indicating a positive sample; y*Y represents the point multiplication of y and Y, which ensures that only the values of the unmarked samples are updated each time; W is the feature coefficient, W∈R k×m ; X is the feature value σ(z) represents a logistic regression model, i.e. λ is the penalty coefficient.
[0026] S3-3, solving the optimal feature coefficient W and penalty coefficient λ of the target function y*Y expression by iterative updating, which ensures that only the values of the unmarked samples are updated each time; specifically including the following steps:
[0027] S3-3-1, deriving σ(z) to get σ'(z)=σ(z)(1-σ(z)), using the cross-entropy loss function, By solving, we get
[0028] S3-3-2, set the initial value W = [0] * number of features, the maximum number of iterations, the learning rate α to 0.01, and the convergence threshold θ to 0.01.
[0029] S3-3-3, learning coefficient W, target sample y, the specific update expression is as follows:
[0030]
[0031] y t+1 =α(σ(w) t+1 X T )+λ|w t+1 |^2)*Y
[0032] S3-3-4, Calculate the change ||w t+1 -w t ||2 and||y t+1 *Yy t *Y||2, when any change ||w t+1 -w t If ||2 is less than the convergence threshold θ, stop the iteration; otherwise, repeat S3-3-3 and S3-3-4 until the number of iterations exceeds the maximum number of iterations.
[0033] In S3-3-5, the penalty coefficient λ is obtained through np.logspace(-2,3,num=6,base=10), resulting in an array([1.e-02,1.e-01,1.e+00,1.e+01,1.e+02,1.e+03]). The results are iterated through, and the y_pred value of the validation data is calculated using a soft-threshold objective function. This value is then compared with the true y value of the validation data to calculate f1, and the λ with the highest f1 is selected. The penalty coefficient is added to prevent overfitting. The f1 evaluation function balances the precision and recall of the classification model and can be considered a harmonic average of these two metrics.
[0034] Furthermore, as a preferred implementation, in S4, the call behavior list is predicted and identified periodically based on the risk number identification model, high-risk suspected risk numbers are output and pushed to the risk shutdown platform, and warnings are issued for reliable suspected numbers in conjunction with the business review mechanism.
[0035] This invention employs the above technical solution, collecting interconnection behavior data as features for a suspected risk number identification model, calculating the proportional relationship between interconnection list call features and normal call features, thereby improving the expressive power of data features. This invention constructs a risk number behavior identification model code based on a soft threshold function and an adaptive learning method. The soft threshold function characterizes the reliability of positive samples among unlabeled samples. Through adaptive learning parameters and multiple iterations, the labeling of positive risk number samples and feature coefficients are simultaneously optimized within the same model. Attached Figure Description
[0036] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;
[0037] Figure 1 This is a flowchart illustrating a risk number identification method based on adaptive learning according to the present invention.
[0038] Figure 2 This is a schematic diagram illustrating the process of solving the objective function of this invention. Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.
[0040] This invention acquires target numbers and their behavioral characteristics, and expands multi-dimensional statistical features based on the behavioral characteristics; by adaptively learning parameters and iterating multiple times, it simultaneously optimizes risk number marking and feature coefficients in the same model, and establishes a suspected risk number identification model to solve a traditional problem of learning illegal risk behaviors.
[0041] like Figure 1 As shown in Figure 2, this invention discloses a risk number identification method based on adaptive learning, which includes the following steps:
[0042] S1, Collect encrypted call behavior data of numbers suspected of risky behavior;
[0043] S2, extract call behavior features of suspected risky numbers from call behavior data, expand the data statistical features of suspected risky numbers through statistical methods, and perform data cleaning and filling preprocessing on the expanded data to obtain a set of data features;
[0044] S3. Construct a soft threshold objective function as a risk number behavior recognition model, and use the preprocessed data feature set as the feature X of the objective function to calculate the target value y. Combine the cross-entropy loss function to iteratively update and optimize the parameters of the risk number behavior recognition model to obtain the final risk number behavior recognition model.
[0045] S4 uses the final risk number behavior recognition model to identify the data feature set, generate a list of suspected risk numbers, and issue warnings for suspected risk numbers.
[0046] Furthermore, S1 obtains encrypted call behavior data of the number based on time intervals; the call behavior data includes voice, SMS, and interconnection behavior characteristics.
[0047] Furthermore, in S1, the collection of call behavior data refers to taking pictures of the call numbers through a data collection system and big data technology, extracting the call behavior feature data of the numbers, and encrypting and desensitizing key fields through encryption and desensitization technologies. At the same time, it realizes the automatic collection of target numbers and behavioral features so that the model can be persisted in the future.
[0048] Furthermore, S2 specifically includes the following steps:
[0049] S2-1, Obtain the voice call detail records (CDRs) of the number, including the number of outgoing calls, call duration, number of call days, call time, call dispersion, call interval, call frequency, and two-way call behavior; where call dispersion refers to the ratio of the number of calls to the number of numbers on the other end.
[0050] S2-2, Obtain the SMS call detail records of the number, including the number of SMS messages, SMS days, call duration, dispersion, SMS interval, SMS frequency, and two-way SMS; where dispersion refers to the ratio of the number of SMS messages to the number of SMS recipient numbers, and two-way SMS refers to the number and proportion of numbers that have two-way SMS behavior.
[0051] S2-3, obtain the call details of the number communication interconnection, including the number of calls, the number of outgoing calls, the percentage of outgoing calls, the number of abnormal call details, the connection rate, the call dispersion, and the frequency of abnormal call details; among them, abnormal call details refer to the number of times the other party's number is not connected, does not respond, or is out of service range, as well as the ratio between the number of abnormal call details and normal call details.
[0052] S2-4 extends the statistical features of time-interval data through statistical functions, including maximum value (max), minimum value (min), unique value (unique), mean value (mean), and quartiles (quartiles).
[0053] S2-5, Perform corresponding filling processing on different missing data according to business characteristics;
[0054] Furthermore, in S2-3, the interconnected call records represent all call behavior between numbers, including call behavior data for numbers that were not connected, did not respond, or were dialed outside the service area. This invention introduces interconnected call records to more accurately represent the distribution and characteristics of data features.
[0055] Furthermore, S3 specifically includes the following steps:
[0056] S3-1, Set a label matrix Y, which is a Boolean matrix consisting of 0-1 elements. It is used to update only the unlabeled target samples in Y each time the model coefficients are calculated iteratively; where a value of 1 indicates that the element at the corresponding position in Y is a labeled sample, and a value of 0 indicates that the element at the corresponding position in Y is an unlabeled sample.
[0057] S3-2, Based on the label matrix Y, a soft threshold objective function is designed as the expression for the prediction model. The specific expression is as follows:
[0058] y*Y=σ(z)+λ|W| 2 =σ(WX T )+λ|W| 2
[0059] Where y represents the target value function, y takes the value 0 to represent a negative sample, and y takes the value 1 to represent a positive sample; y*Y represents the dot product of y and Y, ensuring that only the values of unlabeled samples are updated each time; W is the feature coefficient, W∈R k×m X is the eigenvalue σ(z), representing the logistic regression model, i.e. λ is the penalty coefficient.
[0060] S3-3, by iteratively updating the optimal feature coefficients W and penalty coefficient λ of the objective function y*Y, ensures that only the values of the unlabeled sample portion are updated each time; specifically, it includes the following steps:
[0061] S3-3-1, Differentiating σ(z) gives σ ′ (z) = σ(z)(1-σ(z)), using the cross-entropy loss function, The solution is obtained by solving the problem.
[0062] S3-3-2, set the initial value W = [0] * number of features, the maximum number of iterations, the learning rate α to 0.01, and the convergence threshold θ to 0.01.
[0063] S3-3-3, learning coefficient W, target sample y.
[0064]
[0065] y t+1 =α(σ(w) t+1 X T )+λ|w t+1 |^2)*Y
[0066] S3-3-4, Calculate the change ||w t+1-w t ||2 and||y t+1 *Yy t *Y||2, when any change ||w t+1 -w t If ||2 is less than the convergence threshold θ, stop the iteration; otherwise, repeat S3-3-3 and S3-3-4 until the number of iterations exceeds the maximum number of iterations.
[0067] In S3-3-5, the penalty coefficient λ is obtained through np.logspace(-2,3,num=6,base=10), resulting in an array([1.e-02,1.e-01,1.e+00,1.e+01,1.e+02,1.e+03]). The results are iterated through, and the y_pred value of the validation data is calculated using a soft-threshold objective function. This value is then compared with the true y value of the validation data to calculate f1, and the λ with the highest f1 is selected. The penalty coefficient is added to prevent overfitting. The f1 evaluation function balances the precision and recall of the classification model and can be considered a harmonic average of these two metrics.
[0068] Furthermore, as a preferred implementation, in S4, the call behavior list is predicted and identified periodically based on the risk number identification model, high-risk suspected risk numbers are output and pushed to the risk shutdown platform, and warnings are issued for reliable suspected numbers in conjunction with the business review mechanism.
[0069] This invention acquires target numbers and their behavioral characteristics, and expands multi-dimensional statistical features based on the behavioral characteristics; it designs a soft threshold recognition objective function, and optimizes risk number marking and feature coefficients simultaneously in the same model through adaptive learning parameters and multiple iterations, establishing a suspected risk number recognition model to solve a traditional problem of learning illegal risk behaviors, effectively improving the model's recognition performance.
[0070] The performance of this invention was analyzed through ablation experiments as follows:
[0071] Table 1: Comparison of the effects of the present invention and existing technical solutions
[0072]
[0073] As shown in Table 1, the model parameters of all three methods were optimized, and the training and test sets used were consistent. The F1-score results show that the method of this invention (12.02%) is superior to the results of PU-learning (9.45%) and the single-model decision tree model (6.21%). PU-learning specifically uses a binary classifier for reliable positive sample identification, and then utilizes the identified sample data to perform a risk number identification model through a decision tree model, achieving better results than using the decision tree model alone.
[0074] This invention employs the above technical solution, collecting interconnection behavior data as features for a suspected risk number identification model, calculating the proportional relationship between interconnection list call features and normal call features, thereby improving the expressive power of data features. This invention constructs a risk number behavior identification model code based on a soft threshold function and adaptive learning methods. The soft threshold function characterizes the reliability of positive samples among unlabeled samples. Through adaptive learning parameters and multiple iterations, the labeling of positive risk number samples and feature coefficients are simultaneously optimized within the same model. This invention effectively improves the identification performance of risk numbers through adaptive learning parameters and multiple iterations.
[0075] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
Claims
1. A risk number recognition method based on adaptive learning, characterized in that: It comprises the following steps: S1, collect encrypted call behavior data of suspected risk behavior number; S2, extract call behavior features of suspected risk behavior number from call behavior data, extend data statistical features of suspected risk behavior number by statistical method, and perform data cleaning and filling preprocessing on the extended data to obtain a data feature set; S3, construct a soft threshold objective function as a risk number behavior recognition model, and take the preprocessed data feature set data as the features X of the objective function to calculate the target value y, and combine the cross entropy loss function to iteratively update and optimize the learning coefficient W and the penalty coefficient λ of the risk number behavior recognition model to obtain the final risk number behavior recognition model; S4, use the final risk number behavior recognition model to recognize the data feature set, generate a suspected risk number list, and issue a warning for the suspected risk number. 2.The risk number identification method based on adaptive learning according to claim 1, characterized in that: In S1, the encrypted call behavior data of the number is obtained based on the time interval; the call behavior data includes voice, SMS, and interconnection behavior features. 3.The risk number identification method based on adaptive learning according to claim 1, characterized in that: In S1, the number call behavior data collection refers to taking a photo of the call number through a data collection system and big data technology, extracting the call behavior feature data of the number, and encrypting and desensitizing the key fields through encryption and desensitization technology, while realizing the automatic collection of target numbers and behavior features for subsequent model persistence.
4. The risk number identification method based on adaptive learning according to claim 1, characterized in that: S2 specifically comprises the following steps: S2-1, obtain the voice call sheet of the number communication, including the number of calls, the call duration, the number of calls, the call time, the call dispersion, the call interval, the call frequency, and the two-way call behavior; wherein the call dispersion is the ratio of the number of calls to the number of opposite numbers; S2-2, obtain the SMS call sheet of the number communication, including the number of SMS, the number of SMS, the call time, the dispersion, the SMS interval, the SMS frequency, and the two-way SMS; wherein the dispersion is the ratio of the number of SMS to the number of SMS opposite numbers, and the two-way SMS is the number and proportion of numbers with two-way SMS behavior; S2-3, obtain the interconnection call sheet of the number communication, including the number of calls, the number of calls, the proportion of calls, the number of abnormal call sheets, the connection rate, the call dispersion, and the call frequency of abnormal call sheets; wherein the abnormal call sheet refers to the number of times the opposite number is not connected, not responded, and not in the service range, as well as the number ratio between abnormal call sheets and normal call sheets; S2-4, extend the data statistical features based on the time interval by statistical functions, including the maximum value max, the minimum value min, the unique value unique, the average value mean, and the quartile quartile; S2-5, fill in the corresponding missing data according to the business features.
5. The risk number identification method based on adaptive learning according to claim 4, characterized in that: In S2-3, the interconnection call sheet represents all call behaviors between numbers, and the interconnection call sheet includes the number of behaviors of numbers that are not connected, not responded, and not in the service range.
6. The risk number identification method based on adaptive learning according to claim 1, characterized in that: S3 specifically comprises the following steps: S3-1, a label matrix Y is set, the label matrix Y is a Boolean matrix composed of 0-1 elements, and is used to update only the target samples not labeled in Y each time the model coefficient is iteratively calculated; wherein the 1 value indicates that the element at the corresponding position in Y is a labeled sample, and the 0 value indicates that the element at the corresponding position in Y is an unlabeled sample; S3-2, a soft threshold objective function is designed based on the label matrix Y as an expression of the prediction model, and the specific expression is: y * Y = σ(z) + λ|W| 2 = σ(WX T )+ λ|W| 2 Wherein, y represents the target value function, y value 0 represents a negative sample, and y value 1 represents a positive sample; y * Y represents the point multiplication of y and Y, ensuring that the value of only the unlabeled sample part is updated each time; W is a feature coefficient, W is R k×m ; X is a feature value σ(z) represents a logistic regression model, that is λ is a penalty coefficient; S3-3, solving the objective function y by iterative update * The characteristic coefficient W and the penalty coefficient λ of the Y expression optimization ensure that only the values of the unlabeled sample part are updated each time; specifically, the following steps are included: S3-3-1, σ(z) is derived σ ′ (z) = σ(z)(l - σ(z)), using a cross-entropy loss function, By solving, we get S3-3-2, an initial value W = [0]*feature number is set, the maximum number of iterations is set, the learning rate alpha is set to 0.01, and the convergence threshold theta is set to 0.01; S3-3-3, the learning coefficient W and the target sample y are updated, and the specific update expression is as follows: y t+1 = a (s(w t+1 X T )+ l |w t+1 |^2) * Y S3-3-4, calculate the change amount ||w t+1 -w t ||2 and ||y t+1* Y-y t* Y||2, when any one of the change amount ||w t+1 -w t When the change amount ||2 is less than the convergence threshold θ, stop iteration, otherwise repeat S3-3-3 and S3-3-4 until the number of iterations exceeds the maximum number of iterations; S3-3-5, the penalty coefficient λ is valued by np.logspace(-2, 3, num=6, base=10), the value result is array([1.e-02, 1.e-01, 1.e+00, 1.e+01, 1.e+02, 1.e+03]); traverse the value result, calculate the y_pred value of the verification data through the soft threshold objective function, compare it with the true y value of the verification data to calculate f1, and select the λ with the highest f1.
7. The risk number identification method based on adaptive learning according to claim 1, characterized in that: In S4, the call behavior list is predicted and identified according to the risk number identification model by period, and high-risk suspected risk numbers are output and pushed to the risk shutdown platform, and a reliable suspected number is warned by combining a business audit mechanism.