Abnormal user discovery method and system based on time sequence characteristics and storage medium

Through an abnormal user classification model based on time series characteristics, the XGBoost algorithm is used to judge abnormal users in the virtual community, which solves the problem of degradation of detection capabilities in the existing technology and improves the quality and reliability of the data asset database.

CN118981607APending Publication Date: 2024-11-19SICHUAN JIUZHOU ELECTRIC GROUP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411108485.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

The existing machine learning-based garbage user detection method does not consider the time characteristics, resulting in a decline in model detection capabilities, lack of improvements in the analysis of abnormal user behavior and feature extraction, which affects the quality and reliability of the data asset library.

Method used

By quantifying user behavior similarity characteristics and user basic characteristics based on time series, establishing an abnormal user classification model, using the distributed gradient enhancement model XGBoost to make abnormal user judgments, and removing invalid data from the digital asset library.

Benefits of technology

It improves the accuracy of abnormal user detection, reduces the impact of invalid data on the data asset library, and enhances the integrity and reliability of the data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981607A_ABST
    Figure CN118981607A_ABST
Patent Text Reader

Abstract

The invention discloses an abnormal user discovery method and system based on time sequence characteristics and a storage medium, and belongs to the technical field of machine learning. The method comprises the steps of quantifying user behavior similarity features and user basic features according to a time sequence, wherein the user behavior similarity features comprise user behavior attention, user behavior activeness and user behavior response speed; establishing an abnormal user classification model based on the user behavior similarity features and the user basic features; and judging whether the target user is an abnormal user by using the abnormal user classification model, and removing invalid data of the abnormal user in the digital asset library. An abnormal user classification model based on time sequence features is established through user behavior similarity features and user basic features based on the time sequence, so that the detection accuracy of abnormal users is improved, and the problem of discovery of the type of users is solved. According to the discovery of the abnormal user, the influence of invalid data on the data asset library can be reduced when the data asset library collects open source data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method, system and storage medium for discovering abnormal users based on time series features. Background Art

[0002] In the process of building and maintaining a data asset library, it is crucial to ensure the quality and reliability of the data. The existence of invalid data will not only damage the integrity of the data, but also affect its practicality. In order to improve the overall effectiveness of the data asset library, measures must be taken to identify and eliminate invalid data to ensure the accuracy and reliability of the data. In virtual communities, a large amount of invalid data is usually generated by users with abnormal behavior. The current junk user detection method based on machine learning does not take into account the time characteristics, and the detection ability of the model usually decreases over time. First, the existing technology lacks the cause analysis of such users posting invalid data; second, the existing technology needs to maintain the feature library, and does not take into account the characteristics of such users in the time flow; finally, the existing technology needs to improve the extraction of differential features between such users and normal users. Summary of the invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a method, system and storage medium for discovering abnormal users based on time series features.

[0004] The objective of the present invention is achieved through the following technical solutions: The first aspect of the present invention provides: a method for discovering abnormal users based on time series features, comprising the following steps: Quantify user behavior similarity features and user basic features according to time series, wherein the user behavior similarity features include user behavior attention, user behavior activity, and user behavior response speed; Establish an abnormal user classification model based on user behavior similarity features and user basic features; Use the abnormal user classification model to determine whether the target user is an abnormal user and remove invalid data of abnormal users in the digital asset library.

[0005] Preferably, in a certain function of a virtual community, any user u The user behavior set is C , then: in, Arranged in chronological order, the first user behavior c 1 The time point quantification value is t 1 , Second user behavior c 2 The time point quantification value is t2 , No. User Conduct The time point quantification value is ; For user behavior collection C There is a corresponding set of quantized values ​​at time points T : in, ; When user behavior collection C exist n User behavior In time series When the behavior characteristics are similar, the user behavior similarity s is calculated: .

[0006] Preferably, the first user behavior c 1 The quantified value of attention is l 1 , Second user behavior c 2 The quantified value of attention is l 2 , No. User Conduct The quantified value of attention is ; For user behavior collection C There exists a corresponding set of attention quantization values L : , then any user u User behavior attention Arr for: .

[0007] Preferably, user behavior activity Liveliness : .

[0008] Preferably, the first user behavior c 1 The time difference between the previous user behavior is , Second user behavior c 2 The time difference between the previous user behavior is , No. User Conduct The time difference between the previous user behavior is ; For user behavior collection C There is a corresponding set of response speed quantization values : , any user uUser behavior response speed for: .

[0009] Preferably, the abnormal user classification model is composed of m The construction process is composed of a number of base models, which continuously reduce the loss between the predicted value and the true value, and the construction is completed when the preset loss is reached; m The base model is i The predicted value of samples It is expressed by the following formula: in For the m -1 predicted value of the base model, For the m The predicted value of each base model; Objective function of abnormal user classification model The calculation formula is as follows: in, n is the number of samples, is the loss function of the sample true value and predicted value; is the objective function The regularization term is the complexity of all base models. sum; According to formula (5), the formula number and Taylor theorem transformation loss function are obtained: in, For the m -1 loss function of the predicted value and the true value of the base model, For the m The first-order derivative of the loss function of the base model, h i For the m The second-order derivative of the loss function of the base model; then the objective function Simplifying it gives: By solving the first-order derivative and second-order derivative of the loss function of each base model, the objective function The value of is the smallest, and the , and then gradually optimize and iterate to finally obtain the abnormal user classification model.

[0010] Preferably, the abnormal user classification model is a distributed gradient boosting model XGBoost.

[0011] The second aspect of the present invention provides: a system for discovering abnormal users based on time series features, which is used to implement any of the above-mentioned methods for discovering abnormal users based on time series features, comprising: A feature quantification module, used to quantify user behavior similarity features and user basic features according to time series, wherein the user behavior similarity features include user behavior attention, user behavior activity, and user behavior response speed; A model building module is used to build an abnormal user classification model based on user behavior similarity features and user basic features; The abnormal user judgment module is used to use the abnormal user classification model to judge whether the target user is an abnormal user and remove the data of abnormal users in the digital asset library.

[0012] The third aspect of the present invention provides: a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by a processor, any of the above-mentioned abnormal user detection methods based on time series features is implemented.

[0013] The beneficial effects of the present invention are: 1) By using the user behavior similarity characteristics based on time series and the basic characteristics of users, an abnormal user classification model based on time series characteristics is established, thereby improving the accuracy of abnormal user detection and solving the problem of discovering such users. The discovery of abnormal users can reduce the impact of invalid data on the data asset library when the data asset library collects open source data. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Workflow diagram of the model for classifying abnormal users in the digital asset library. DETAILED DESCRIPTION

[0015] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0016] See also Figure 1 The first aspect of the present invention provides: a method for discovering abnormal users based on time series features, comprising the following steps: Quantify user behavior similarity features and user basic features according to time series, wherein the user behavior similarity features include user behavior attention, user behavior activity, and user behavior response speed; Establish an abnormal user classification model based on user behavior similarity features and user basic features; Use the abnormal user classification model to determine whether the target user is an abnormal user and remove invalid data of abnormal users in the digital asset library.

[0017] In this embodiment, the user behavior similarity feature in the virtual community, as the time series changes, the behavior characteristics of abnormal users and ordinary users (including the feedback characteristics of behavior acquisition) are usually different. In order to explore the difference between the behavior of abnormal users and ordinary users, the present invention models and analyzes the user behavior similarity and quantifies it. Abnormal users hope to get more attention in the virtual community to achieve the purpose of spreading inappropriate information or seeking benefits. The present invention models and analyzes the user behavior attention characteristics, user activity characteristics, and user behavior response speed characteristics respectively. The identification problem of abnormal users is a binary classification problem. The present invention first annotates the original data set to distinguish between positive and negative samples. After that, the user behavior characteristics are extracted and quantified to construct a user behavior feature vector. To ensure the accuracy of the model, the model uses a cross-validation method to divide the feature vector group into test data and training data, and combines the XGBoost method to classify the positive and negative samples in the data set.

[0018] In some embodiments, under a certain function of a virtual community, any user u The user behavior set is C , then: in, Arranged in chronological order, the first user behavior c 1 The time point quantification value is t 1 , Second user behavior c 2 The time point quantification value is t 2 , No. User Conduct The time point quantification value is ; For user behavior collection C There is a corresponding set of quantized values ​​at time points T : in, ; When user behavior collection C exist n User behavior In time series When the behavior characteristics are similar, the user behavior similarity s is calculated: .

[0019] In this embodiment, according to the user behavior set CThe user behavior similarity s is calculated by using the continuous similar behaviors in the time series. Formula (1) is called the user behavior similarity function in the time series.

[0020] In some embodiments, the first user behavior c 1 The quantified value of attention is l 1 , Second user behavior c 2 The quantified value of attention is l 2 , No. User Conduct The quantified value of attention is ; For user behavior collection C There exists a corresponding set of attention quantization values L : , then any user u User behavior attention Arr for: .

[0021] In this embodiment, for any quantization value l i , l i The larger the value of is, the higher the behavior attention is. Formula (2) is called the user average behavior attention function.

[0022] In some embodiments, user behavior activity Liveliness : .

[0023] In this embodiment, formula (3) is called the user behavior activity function.

[0024] In some embodiments, the first user behavior c 1 The time difference between the previous user behavior is , Second user behavior c 2 The time difference between the previous user behavior is , No. User Conduct The time difference between the previous user behavior is ; For user behavior collection C There is a corresponding set of response speed quantization values : , any user u User behavior response speed for: .

[0025] In this embodiment, for any response speed quantization value ∆ t i ,∆ t i The smaller the value of is, the faster the user behavior response speed is. Formula (4) is called the user average behavior response speed function.

[0026] In some embodiments, the abnormal user classification model is composed of m The construction process is composed of a number of base models, which continuously reduce the loss between the predicted value and the true value, and the construction is completed when the preset loss is reached; m The base model is i The predicted value of samples It is expressed by the following formula: in For the m -1 predicted value of the base model, For the m The predicted value of each base model; Objective function of abnormal user classification model The calculation formula is as follows: in, n is the number of samples, is the loss function of the sample true value and predicted value; is the objective function The regularization term is the complexity of all base models. sum; According to formula (5), the formula number and Taylor theorem transformation loss function are obtained: in, For the m -1 loss function of the predicted value and the true value of the base model, For the m The first-order derivative of the loss function of the base model, h i For the m The second-order derivative of the loss function of the base model; then the objective function Simplifying it gives: By solving the first-order derivative and second-order derivative of the loss function of each base model, the objective function The value of is the smallest, and the , and then gradually optimize and iterate to finally obtain the abnormal user classification model.

[0027] In this embodiment, the abnormal user classification model is an XGBoost model, which is an optimized distributed gradient boosting model. m The additive model composed of base models is used to predict the user's label value. The model construction process involves constantly adjusting the loss between the predicted label value and the true label value to minimize the loss. The objective function of the abnormal user classification model is The role of the regularization term is to prevent the model from overfitting.

[0028] In some embodiments, the abnormal user classification model is a distributed gradient boosting model XGBoost.

[0029] An application example is given below. Figure 1 To ensure the accuracy of the model, the model uses the cross-validation method to divide the feature vector group into test data and training data, and combines the XGBoost algorithm to classify the positive and negative samples in the data set.

[0030] The present invention extracts the feature vector of any user from two aspects: the abnormal behavior characteristics of the user based on the time dimension and the basic characteristics of the user. The specific meanings represented by each feature quantization value are listed here, and the feature set used in the experiment is shown in Table 1.

[0031] Table 1 Statistical data of abnormal user characteristics Abnormal behavior characteristics: Refers to the percentage of the longest zero-value sequence obtained in the user behavior based on the time dimension for a user in the data set; is the percentage of the longest non-zero value sequence obtained; Arr Pay attention to user behavior; Liveliness is the user behavior activity; ∆T is the user behavior response speed.

[0032] Basic user characteristics: F er Refers to the number of followers any user receives in the community; Total is the total number of microblogs posted by the user, V is the user's level in the community, Lean is the average length of a user's speech.

[0033] The experimental data of this application example includes microblogs of 3,129 users with a large number of followers, and 2,301,554 user behaviors from 921,863 users under the microblogs, as shown in Table 2.

[0034] Table 2 Experimental data table The model extracts and quantifies the behavior of highly active users in time series, generates feature vectors for each user, and uses eight-fold cross validation to ensure that the model does not overfit. Finally, the effectiveness of the abnormal user classification method based on time series features is verified, and the results are shown in Table 3.

[0035] Table 3 Abnormal user detection results based on time series features As can be seen from Table 3, the detection model based on the XGBoost algorithm performs better in the comparative test, which shows that the model is feasible in detecting abnormal users. In addition, the present invention uses the T test to verify the statistical significance of the proposed user behavior features. The results show that the p value of each feature is less than 0.05, indicating that the features proposed by the present invention are also effective.

[0036] The second aspect of the present invention provides: a system for discovering abnormal users based on time series features, which is used to implement any of the above-mentioned methods for discovering abnormal users based on time series features, comprising: A feature quantification module, used to quantify user behavior similarity features and user basic features according to time series, wherein the user behavior similarity features include user behavior attention, user behavior activity, and user behavior response speed; A model building module is used to build an abnormal user classification model based on user behavior similarity features and user basic features; The abnormal user judgment module is used to use the abnormal user classification model to judge whether the target user is an abnormal user and remove the data of abnormal users in the digital asset library.

[0037] The third aspect of the present invention provides: a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by a processor, any of the above-mentioned abnormal user detection methods based on time series features is implemented.

[0038] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for discovering abnormal users based on time series features, characterized by: The following steps are involved: Quantify user behavior similarity features and user basic features according to time series, wherein the user behavior similarity features include user behavior attention, user behavior activity, and user behavior response speed; Establish an abnormal user classification model based on user behavior similarity features and user basic features; Use the abnormal user classification model to determine whether the target user is an abnormal user and remove invalid data of abnormal users in the digital asset library.

2. The abnormal user discovery method based on time series features according to claim 1 is characterized in that: In a certain function of a virtual community, any user u The user behavior set is C , then: in, Arranged in chronological order, the first user behavior c The quantitative value at the time point of 1 is t 1. Second user behavior c The quantitative value at the time point 2 is t 2. No. User Conduct The time point quantification value is ; For user behavior collection C There is a corresponding set of quantized values ​​at time points T : in, ; When user behavior collection C exist n User behavior In time series When the behavior characteristics are similar, the user behavior similarity s is calculated: 。 3. The abnormal user discovery method based on time series features according to claim 2 is characterized in that: First user behavior c The quantified value of attention is 1 l 1. Second user behavior c The quantified value of attention 2 is l 2. No. User Conduct The quantified value of attention is ; For user behavior collection C There exists a corresponding set of attention quantization values L : , then any user u User behavior attention Arr for: 。 4. The abnormal user discovery method based on time series features according to claim 2 is characterized in that: User Behavior Activity Liveness : 。 5. The abnormal user discovery method based on time series features according to claim 2 is characterized in that: First user behavior c The time difference between 1 and its previous user behavior is , Second user behavior c The time difference between 2 and its previous user behavior is , No. User Conduct The time difference between the previous user behavior is ; For user behavior collection C There is a corresponding set of response speed quantization values : , any user u User behavior response speed for: 。 6. The abnormal user discovery method based on time series features according to claim 1 is characterized in that: The abnormal user classification model is composed of m The construction process is composed of a number of base models, which continuously reduce the loss between the predicted value and the true value, and the construction is completed when the preset loss is reached; m The base model is i The predicted value of samples It is expressed by the following formula: in For the m -1 predicted value of the base model, For the m The predicted value of each base model; Objective function of abnormal user classification model The calculation formula is as follows: in, n is the number of samples, is the loss function of the sample true value and predicted value; is the objective function The regularization term is the complexity of all base models. sum; According to formula (5), the formula number and Taylor theorem transformation loss function are obtained: in, For the m -1 loss function of the predicted value and the true value of the base model, For the m The first-order derivative of the loss function of the base model, h i For the m The second-order derivative of the loss function of the base model; then the objective function Simplifying it gives: By solving the first-order derivative and second-order derivative of the loss function of each base model, the objective function The value of is the smallest, and the , and then gradually optimize and iterate to finally obtain the abnormal user classification model.

7. The abnormal user discovery method based on time series features according to any one of claims 1 to 6, characterized in that: The abnormal user classification model is a distributed gradient boosting model XGBoost.

8. An abnormal user discovery system based on time series features, characterized by: The method for discovering abnormal users based on time series features according to any one of claims 1 to 7 comprises: A feature quantification module, used to quantify user behavior similarity features and user basic features according to time series, wherein the user behavior similarity features include user behavior attention, user behavior activity, and user behavior response speed; A model building module is used to build an abnormal user classification model based on user behavior similarity features and user basic features; The abnormal user judgment module is used to use the abnormal user classification model to judge whether the target user is an abnormal user and remove the data of abnormal users in the digital asset library.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by the processor, the abnormal user discovery method based on time series features as described in any one of claims 1-7 is implemented.