Intelligent identification system of cheating behavior in advertising alliances based on deep learning

Abstract

The present invention relates to an intelligent identification system for cheating behavior in advertising alliances based on deep learning, which collects user information monitored in historical periods and the corresponding cheating behavior; uses a redundant analysis model to identify the correlation between user information and cheating behavior; and constructs a random forest model between the correlated user information and cheating behavior. The user information entering the advertisement is monitored in real time, and input into the random forest model, and the predicted cheating behavior is output; when the proportion of the number of cheating users exceeds the set proportion threshold, it is determined that the advertisement is cheating. The present invention realizes intelligent identification of cheating behavior in advertising through deep learning of the complex relationship between user information and its cheating behavior, solving the problems of low accuracy and low efficiency of the prior art.

Description

Technical Field

[0001] The present invention belongs to the field of computer Internet technology and relates to an intelligent recognition system for cheating behavior of advertising alliances based on deep learning. Background Art

[0002] Ad alliance is an online advertising model that optimizes advertising delivery and management by pooling the resources of advertisers and publishers. In this model, advertisers pay for advertising, and publishers earn revenue by displaying or clicking on ads. However, with the rapid growth of the online advertising market, ad fraud is becoming increasingly rampant. Common ad fraud includes click fraud, display fraud, and installation fraud, which not only wastes advertisers' funds, but also damages the healthy development of the entire advertising ecosystem.

[0003] Traditional ad fraud detection methods usually rely on rule engines and statistical analysis. Although these methods can identify common fraud behaviors to a certain extent, they often seem powerless when faced with ever-changing fraud methods and large amounts of data. Rule engines require continuous manual updates and maintenance, while statistical analysis may have difficulty capturing complex fraud patterns. As an important branch of artificial intelligence, deep learning has powerful data processing and pattern recognition capabilities. By training deep learning models, features can be automatically extracted from large amounts of complex data and potential ad fraud behaviors can be identified. Summary of the invention

[0004] The present invention provides an intelligent identification system for cheating behavior in advertising alliances based on deep learning, which aims to utilize the rich user information (behavior information, device information and geographic information) collected in historical periods and its corresponding cheating behavior data, and realize intelligent identification of cheating behavior in advertising through deep learning of the complex relationship between user information and its cheating behavior, thereby solving the problems of low accuracy and efficiency of the existing technology.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] The present application provides an intelligent identification system for cheating behavior in advertising alliances based on deep learning, including a data monitoring module, a cheating analysis module and a statistical application module, wherein the data monitoring module, the cheating analysis module and the statistical application module are communicatively connected, wherein:

[0007] The data monitoring module is used to monitor the user information of the user who enters the advertisement in real time, and the user information includes behavior information, device information and geographic information;

[0008] The cheating analysis module is used to input the real-time monitored user information into a preset cheating analysis unit and output the predicted cheating behavior; the cheating behavior includes click cheating, display cheating, installation cheating and no cheating behavior; the cheating analysis unit includes the following construction steps:

[0009] S1. Collect user information monitored in historical periods and their corresponding cheating behaviors;

[0010] S2, using the redundancy analysis model to identify the correlation between user information and cheating behavior;

[0011] S3, build a random forest model between relevant user information and cheating behavior;

[0012] The statistical application module is used to count the number of cheating users whose cheating behavior is not non-cheating behavior, and calculate the proportion of the number of cheating users.

[0013] Furthermore, in the data monitoring module, the behavioral information includes click patterns, browsing behaviors and interactive behaviors, and the click pattern is the frequency, time and source of user clicks on advertisements; the device information includes the device operating system type and its version, device model and device network messages, and the device network information includes the IP address and network type; the geographic information includes the IP geographic location and the geographic area when the user logs in or operates.

[0014] Furthermore, the redundancy analysis model includes the following construction steps:

[0015] S21. Use user information as the explanatory variable and cheating behavior as the response variable to build a redundancy analysis model.

[0016] S22. Use the Monte Carlo permutation test to check the significance of the full model and the first axis of the model. When the full model and the first axis of the model are significant, the model passes the test;

[0017] S23, drawing a ranking diagram according to the score of each variable in the ranking axis, and determining the angle between the user information and the cheating behavior arrow in the ranking diagram;

[0018] S24. When the angle between the arrows of the user information and the cheating behavior variable is less than a set angle threshold, it is determined that there is a correlation between the two.

[0019] Furthermore, in step S23, the Pearson correlation coefficient of the scores of the user information and the cheating behavior in the sorting axis is used to assist in analyzing the correlation between the user information and the cheating behavior variables. When the Pearson correlation coefficient is greater than 0.7, it is determined that there is a correlation between the two.

[0020] Furthermore, the Pearson correlation coefficient is calculated as follows:

[0021]

[0022] Where: r ij Represents user information variable x i and the cheating behavior variable x j Pearson correlation coefficient between ki Represents x i The kth sample, x kj Represents x j The k-th sample of , k = 1, 2, ..., n, where n is the number of samples.

[0023] Furthermore, the random forest model includes the following construction steps:

[0024] S31, data preparation: prepare a data set including explanatory variables and response variables, the explanatory variables are user information, and the response variables are cheating behavior;

[0025] S32, data set division: the collected data is divided into a training set and a test set, wherein 70% of the data samples in the data set are used as the training set and 30% of the data samples are used as the test set;

[0026] S33, Random Sampling: Randomly extract a certain number of explanatory variables and data samples from the training set, and construct multiple different training sets with replacement;

[0027] S34, constructing a decision tree: for each randomly sampled training set, construct a decision tree model;

[0028] S35, Ensemble Decision Tree: Integrate multiple decision tree models into a random forest model and use the voting mechanism to make the final prediction;

[0029] S36. Model evaluation: Use evaluation indicators to evaluate the prediction performance of the random forest model on the test set and determine the final model with good prediction performance.

[0030] Furthermore, in step S34, the decision tree model is configured as a CART model.

[0031] Furthermore, in step S36, the evaluation indicators include accuracy, precision, recall and F1 value.

[0032] Furthermore, the calculation formula for calculating the proportion of cheating users is as follows:

[0033]

[0034] In the formula, Z r Indicates the proportion of cheating users; Z aIndicates the number of cheating users; Z sum Indicates the total number of users who monitor the ad.

[0035] Furthermore, when the proportion of the number of cheating users exceeds a set proportion threshold, it is determined that cheating exists in the advertisement and the delivery is cancelled.

[0036] Beneficial effects of the present invention:

[0037] By collecting user information monitored in historical periods and its corresponding cheating behavior; using a redundant analysis model to identify the correlation between user information and cheating behavior; and building a random forest model between the correlated user information and cheating behavior. Real-time monitoring of user information entering the advertisement, and input into the random forest model, outputting the predicted cheating behavior; when the proportion of cheating users exceeds the set proportion threshold, it is determined that the advertisement is cheating. The present invention achieves intelligent identification of cheating advertising behavior through deep learning of the complex relationship between user information and its cheating behavior, solving the problems of low accuracy and efficiency of the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0039] Figure 1 This is a structural diagram of an intelligent identification system for cheating behavior in advertising alliances based on deep learning in the present invention. DETAILED DESCRIPTION

[0040] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0041] See also Figure 1 The present application provides an intelligent identification system for cheating behavior in advertising alliances based on deep learning, including a data monitoring module, a cheating analysis module and a statistical application module, wherein the data monitoring module, the cheating analysis module and the statistical application module are communicatively connected, wherein:

[0042] The data monitoring module is used to monitor the user information of the user who enters the advertisement in real time, and the user information includes behavior information, device information and geographic information;

[0043] In this embodiment, behavior information, device information, and geographic information are important factors for identifying and analyzing user cheating behavior. They help the system better distinguish between normal user behavior and potential fraudulent behavior by extracting and analyzing user data:

[0044] Behavioral information includes:

[0045] 1) Click patterns: Identify how often, when, and from where users click on ads.

[0046] 2) Browsing behavior: Analyze users’ browsing behavior on web pages or in applications, including dwell time and page jump patterns.

[0047] 3) Interactive behavior: Capture the user’s interaction with the advertising content, such as video playback duration, clicks on links within the advertisement, etc.

[0048] Device information includes:

[0049] 1) Operating System and Version: Determine the type of operating system and its version on the user’s device to identify operating system vulnerabilities or malware potential.

[0050] 2) Device model: Detects device hardware features, such as resolution, processor type, etc., which helps identify simulator or virtual machine environments.

[0051] 3) Network information: including IP address, network type (such as Wi-Fi or mobile data), etc., used to detect abnormal network traffic or IP proxy usage.

[0052] Geographic information includes:

[0053] 1) IP geolocation: Obtain the user's geolocation information through reverse lookup of the IP address to verify whether the user's claimed location is consistent with the actual location.

[0054] 2) Geographical area: Based on the geographic area information when the user logs in or operates, determine whether the user's activities match their usual activity locations.

[0055] Therefore, further, in the data monitoring module, the behavioral information includes click patterns, browsing behaviors and interactive behaviors, and the click pattern is the frequency, time and source of user clicks on advertisements; the device information includes the device operating system type and its version, device model and device network messages, and the device network information includes the IP address and network type; the geographic information includes the IP geographic location, and the geographic area when the user logs in or operates.

[0056] The cheating analysis module is used to input the real-time monitored user information into a preset cheating analysis unit and output the predicted cheating behavior; the cheating behavior includes click cheating, display cheating, installation cheating and no cheating behavior;

[0057] In this embodiment, cheating behaviors include: click fraud, repeatedly clicking on ads through automated scripts or malicious users to inflate the number of clicks; display fraud: increasing the number of displays by falsifying ad display data to obtain more revenue; installation fraud: in app installation ads, falsifying user installation behavior to obtain advertising fees; no cheating: that is, there is no advertising cheating.

[0058] The cheating analysis unit includes the following construction steps:

[0059] S1. Collect user information monitored in historical periods and their corresponding cheating behaviors;

[0060] S2, using the redundancy analysis model to identify the correlation between user information and cheating behavior;

[0061] S3. Construct a random forest model between relevant user information and cheating behavior.

[0062] In this embodiment, the cheating analysis unit is constructed by using the user information collected in the historical period and the corresponding cheating behavior big data. The big data can be used to identify cheating behaviors related to different user information. However, the correlation between different user information (behavior information, device information and geographic information) and different cheating behaviors (click cheating, display cheating, installation cheating and no cheating behavior) is different. When establishing the prediction relationship between the two, it is necessary to sort out the user information variables that are significantly correlated with each cheating behavior to establish the prediction relationship between them. Therefore, in order to make the big data analysis results more accurate, the redundancy analysis model is first used to identify the correlation between user information and cheating behavior, and then the prediction model of cheating behavior is constructed through the variables of correlation.

[0063] Among them, redundancy analysis (RDA) is an ordination method that combines regression analysis and principal component analysis (PCA) for the expansion of multiresponse regression analysis. RDA extracts and summarizes the changes in a set of response variables and explains these changes through a set of explanatory variables by performing PCA analysis on the fitted value matrix of the multivariate linear regression between the response variable matrix and the explanatory variables. This method allows multiple response variables to be regressed on multiple explanatory variables, and then the fitted value matrix generated by the multivariate linear regression is processed by principal component analysis (PCA). RDA can be regarded as a constrained version of PCA, in which the canonical axes (constructed by linear combinations of response variables) must also be linear combinations of explanatory variables. The interpretation of redundancy analysis plots (RDA plots) involves understanding the ordination axes and linear combinations of explanatory variables shown in the plots. These plots are often used to show how explanatory variables affect changes in response variables. In RDA plots, the ordination axes are actually linear combinations of explanatory variables, and these axes allow you to intuitively see which explanatory variables have the greatest impact on the response variable. In addition, RDA plots can also be used to compare differences between different samples or treatment groups, as well as to identify patterns and trends. The difference between redundancy analysis and correlation analysis is that correlation analysis focuses on the strength of the direct relationship between variables, while redundancy analysis goes a step further and extracts and summarizes the changes in a set of response variables through a combination of regression analysis and principal component analysis, and explains these changes through a set of explanatory variables. The advantage of this method is that it can handle the relationship between multiple response variables and explanatory variables, not just the correlation between them. Compared with principal component analysis (PCA), redundancy analysis (RDA) is a constrained version of principal component analysis, in which the canonical axes (constructed by linear combinations of response variables) must also be linear combinations of explanatory variables. This means that in RDA, the extracted principal components take into account not only the direction of maximum variation in the data, but also the constraints introduced by the explanatory variables. Therefore, RDA provides a more direct way to understand how explanatory variables affect the changes in the response variable.

[0064] Furthermore, the redundancy analysis model includes the following construction steps:

[0065] S21. Use user information as the explanatory variable and cheating behavior as the response variable to build a redundancy analysis model.

[0066] S22. Use the Monte Carlo permutation test to check the significance of the full model and the first axis of the model. When the full model and the first axis of the model are significant, the model passes the test;

[0067] S23, drawing a ranking diagram according to the score of each variable in the ranking axis, and determining the angle between the user information and the cheating behavior arrow in the ranking diagram;

[0068] S24. When the angle between the arrows of the user information and the cheating behavior variable is less than a set angle threshold, it is determined that there is a correlation between the two.

[0069] In this embodiment, the interpretation of the RDA ordination diagram needs to start from the following aspects: the direction and length of the arrows, the angle between the arrows, the distribution of sample points, etc.:

[0070] 1) Direction and length of arrow:

[0071] Arrow direction: represents the direction of the variable on the sorting axis. The closer the projection position of the variable is on the corresponding axis, the greater the contribution of the variable to the axis.

[0072] Arrow length: Indicates the importance of the variable. The longer the arrow, the greater the variation explained by the variable in the ordination diagram and the more significant the impact.

[0073] 2) Angle between arrows

[0074] Angle less than 90 degrees: indicates that there is a positive correlation between the variables. The smaller the angle, the stronger the correlation.

[0075] Angle equal to 90 degrees: indicates that there is no correlation between the variables.

[0076] Angle greater than 90 degrees but less than 180 degrees: indicates that there is a negative correlation between the variables. The closer the angle is to 180 degrees, the stronger the negative correlation.

[0077] 3) Distribution of sample points

[0078] The position of the sample point in the ordination diagram: The position of the sample point indicates its performance on different variables. If the sample point is close to an arrow, it means that the sample has a higher value on the variable.

[0079] Clustering of sample points: The clustering of sample points can indicate the similarity between samples. If the sample points form obvious clusters in the graph, it means that these samples have similarities in the performance of the variable.

[0080] It should be noted that, since there is basically a positive correlation between the user information and the cheating behavior variables in this embodiment, for two variables whose angle between arrows is less than a set angle threshold, it is determined that there is a correlation between the two variables.

[0081] Furthermore, in step S23, the Pearson correlation coefficient of the scores of the user information and the cheating behavior in the sorting axis is used to assist in analyzing the correlation between the user information and the cheating behavior variables. When the Pearson correlation coefficient is greater than 0.7, it is determined that there is a correlation between the two.

[0082] Furthermore, the Pearson correlation coefficient is calculated as follows:

[0083]

[0084] Where: r ij Represents user information variable x i and the cheating behavior variable x j Pearson correlation coefficient between ki Represents x i The kth sample, x kj Represents x j The k-th sample of , k = 1, 2, ..., n, where n is the number of samples.

[0085] In this embodiment, after determining the correlation between user information and cheating behavior, the corresponding random forest model is constructed as a prediction model for intelligent identification of cheating behavior. Random Forest model is a machine learning algorithm widely used in classification and regression problems, which improves the accuracy and stability of the model by integrating multiple decision trees. The basic idea of ​​the algorithm is to construct multiple independently trained decision trees, and then determine the final prediction value by majority voting or averaging their prediction results. The construction process includes randomly sampling data to generate multiple training sets, and each tree only selects part of the features for splitting when splitting nodes, thereby increasing the differences between trees. Random forest has many advantages, such as high prediction accuracy, strong anti-overfitting ability and the ability to evaluate the importance of features. Its important parameters include the number of decision trees, the number of features that can be selected at each split, the maximum depth of the decision tree, etc. Due to the combination of the results of multiple models, random forests usually perform better than single decision trees, especially when processing high-dimensional data and large data sets. In addition, random forests are very common in practical applications and are often used in tasks such as spam detection, image classification, medical diagnosis, and house price prediction. In general, the random forest model is an effective tool for improving prediction performance through ensemble learning technology and has high practical application value.

[0086] Furthermore, the random forest model includes the following construction steps:

[0087] S31, data preparation: prepare a data set including explanatory variables and response variables, the explanatory variables are user information, and the response variables are cheating behavior;

[0088] S32, data set division: the collected data is divided into a training set and a test set, wherein 70% of the data samples in the data set are used as the training set and 30% of the data samples are used as the test set;

[0089] S33, Random Sampling: Randomly extract a certain number of explanatory variables and data samples from the training set, and construct multiple different training sets with replacement;

[0090] S34, constructing a decision tree: for each randomly sampled training set, construct a decision tree model;

[0091] Furthermore, in step S34, the decision tree model is configured as a CART model.

[0092] Classification and Regression Trees (CART) is a decision tree algorithm used for classification and regression analysis. CART recursively divides the data set into smaller subsets, selects the best features and split points at each node to minimize the Gini impurity (for classification) or the mean square error (for regression), and finally generates a tree for prediction. The CART model is easy to understand and interpretable, and can handle nonlinear relationships and high-dimensional data. By repeatedly dividing the data set, CART can effectively capture complex patterns in the data and is also robust to missing values ​​and outliers. In a random forest, each decision tree is built based on the Classification and Regression Tree (CART) algorithm. Specifically, each tree in a random forest is an independent CART decision tree that randomly selects features for modeling. Random forests make the final decision by integrating the prediction results of multiple CART decision trees, thereby improving the accuracy and generalization ability of the overall model. Each tree is built for a different random subset of the training data, and only one random subset of features is considered when selecting the best feature at each node, which makes the random forest have a certain degree of randomness and reduces the risk of overfitting. Therefore, the random forest takes advantage of the CART decision tree, combines the prediction results of multiple trees, improves the stability and prediction ability of the model, and has achieved widespread success in practical applications.

[0093] S35, Ensemble Decision Tree: Integrate multiple decision tree models into a random forest model and use the voting mechanism to make the final prediction;

[0094] It should be noted that in the random forest, each decision tree is constructed based on different training data sets and feature subsets, and has a certain degree of randomness. The final prediction result can be determined by a voting mechanism or by taking the average. In classification problems, a majority vote can be used to select the category with the most votes as the final prediction result; in regression problems, the prediction results of multiple trees can be averaged as the final output. By integrating multiple decision trees into a random forest model, the risk of overfitting can be effectively reduced and the robustness and accuracy of the model can be improved. The prediction of cheating behavior in this embodiment belongs to a classification problem, so a voting mechanism is used to make the final prediction.

[0095] S36. Model evaluation: Use evaluation indicators to evaluate the prediction performance of the random forest model on the test set and determine the final model with good prediction performance.

[0096] Furthermore, in step S36, the evaluation indicators include accuracy, precision, recall and F1 value.

[0097] In this embodiment, the accuracy rate is the ratio of the number of samples correctly predicted by the model to the total number of samples, which is one of the most commonly used evaluation indicators. The calculation formula is: Accuracy Rate = (TP+TN) / (TP+TN+FP+FN), where TP represents true positive examples, TN represents true negative examples, FP represents false positive examples, and FN represents false negative examples.

[0098] Precision measures how many of the positive examples predicted by the model are truly positive examples. It is an important indicator for evaluating the accuracy of model predictions. The calculation formula is: Precision = TP / (TP+FP).

[0099] The recall rate measures the ratio of the number of positive examples that the model can correctly predict to the actual number of positive examples, also known as the recall rate. The calculation formula is: Recall rate = TP / (TP+FN).

[0100] The F1 value is the harmonic mean of precision and recall, which comprehensively considers the accuracy and comprehensiveness of the model. The higher the F1 value, the better the balance between precision and recall. The calculation formula is: F1 = 2×(precision×recall) / (precision+recall).

[0101] The statistical application module is used to count the number of cheating users whose cheating behavior is not non-cheating behavior, and calculate the proportion of the number of cheating users.

[0102] Furthermore, the calculation formula for calculating the proportion of cheating users is as follows:

[0103]

[0104] In the formula, Z r Indicates the proportion of cheating users; Z a Indicates the number of cheating users; Z sum Indicates the total number of users who monitor the ad.

[0105] Furthermore, when the proportion of the number of cheating users exceeds a set proportion threshold, it is determined that cheating exists in the advertisement and the delivery is cancelled.

[0106] In this embodiment, in order to prevent the entire advertisement from being misjudged as cheating because a certain user is determined to be a cheating user, the number of cheating users must exceed a certain percentage before the advertisement is determined to be cheating and its advertising eligibility is revoked.

[0107] Beneficial effects of the present invention:

[0108] By collecting user information monitored in historical periods and its corresponding cheating behavior; using a redundant analysis model to identify the correlation between user information and cheating behavior; and building a random forest model between the correlated user information and cheating behavior. Real-time monitoring of user information entering the advertisement, and input into the random forest model, outputting the predicted cheating behavior; when the proportion of cheating users exceeds the set proportion threshold, it is determined that the advertisement is cheating. The present invention achieves intelligent identification of cheating advertising behavior through deep learning of the complex relationship between user information and its cheating behavior, solving the problems of low accuracy and efficiency of the prior art.

[0109] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. An intelligent identification system for cheating behavior in advertising alliances based on deep learning, characterized by: It includes a data monitoring module, a cheating analysis module and a statistical application module, wherein the data monitoring module, the cheating analysis module and the statistical application module are communicatively connected, wherein: The data monitoring module is used to monitor the user information of the user who enters the advertisement in real time, and the user information includes behavior information, device information and geographic information; The cheating analysis module is used to input the real-time monitored user information into a preset cheating analysis unit and output the predicted cheating behavior; the cheating behavior includes click cheating, display cheating, installation cheating and no cheating behavior; the cheating analysis unit includes: Collect user information monitored in historical periods and their corresponding cheating behaviors; Use redundancy analysis models to identify the correlation between user information and cheating behavior; Build a random forest model between relevant user information and cheating behavior; The statistical application module is used to count the number of cheating users whose cheating behavior is not non-cheating behavior, and calculate the proportion of the number of cheating users; The redundancy analysis model comprises: Using user information as the explanatory variable and cheating behavior as the response variable, a redundancy analysis model is constructed; Monte Carlo permutation test is used to check the significance of the whole model and the first axis of the model. When the whole model and the first axis of the model are significant, the model passes the test. Draw a ranking diagram based on the scores of each variable in the ranking axis, and determine the angle between the arrows of user information and cheating behavior in the ranking diagram; When the angle between the arrows of the user information and cheating behavior variables is less than the set angle threshold, it is determined that there is a correlation between the two; The random forest model includes: Data preparation: prepare a data set containing explanatory variables and response variables, the explanatory variables are user information, and the response variables are cheating behavior; Dataset division: The collected data is divided into training set and test set, where 70% of the data samples in the data set are used as the training set and 30% of the data samples are used as the test set; Random sampling: randomly extract a certain number of explanatory variables and data samples from the training set, and construct multiple different training sets with replacement; Construct a decision tree: For each randomly sampled training set, construct a decision tree model; Ensemble decision tree: Integrate multiple decision tree models into a random forest model and use the voting mechanism to make the final prediction; Model evaluation: Evaluation metrics are used to evaluate the prediction performance of the random forest model on the test set and determine the final model with good prediction performance.

2. According to claim 1, a deep learning-based intelligent identification system for cheating in advertising alliances is characterized by: The behavioral information includes click patterns, browsing behaviors and interactive behaviors. The click patterns are the frequency, time and source of user clicks on advertisements. The device information includes the device operating system type and version, device model and device network information. The device network information includes IP address and network type. The geographic information includes IP geographic location and the geographic area when the user logs in or operates.

3. According to claim 1, a deep learning-based intelligent identification system for cheating in advertising alliances is characterized by: The redundancy analysis model also assists in analyzing the correlation between the user information and the cheating behavior variables through the Pearson correlation coefficient of the user information and the cheating behavior scores in the sorting axis. Specifically, when the Pearson correlation coefficient is greater than 0.7, it is determined that there is a correlation between the two.

4. According to claim 3, a deep learning-based intelligent identification system for cheating in advertising alliances is characterized by: The Pearson correlation coefficient is calculated as follows: ， Where: r ij Represents user information variables x i Variables related to cheating behavior x j Pearson correlation coefficient between ; x ki express x i No. k samples, x kj express x j No. k samples, k =1, 2, …, n ,in n is the number of samples.

5. According to claim 1, a deep learning-based intelligent identification system for cheating in advertising alliances is characterized by: The decision tree model is configured as a CART model.

6. According to claim 1, a deep learning-based intelligent identification system for cheating in advertising alliances is characterized by: The evaluation indicators include accuracy, precision, recall and F1 value.

7. According to claim 1, a deep learning-based intelligent identification system for cheating in advertising alliances is characterized by: The calculation formula for calculating the proportion of cheating users is as follows: ， In the formula, Z r Indicates the percentage of cheating users; Z a Indicates the number of cheating users; Z sum Indicates the total number of users who monitor the ad.

8. According to claim 1, a deep learning-based intelligent identification system for cheating in advertising alliances is characterized by: When the proportion of the number of cheating users exceeds a set proportion threshold, it is determined that the advertisement is cheating and the advertisement is cancelled.

Images