A mobile phone APP advertisement false volume identification method and system

Abstract

The application relates to a mobile phone APP advertisement false volume identification method and system, obtains mobile phone APP advertisement click records of a user, obtains index information, inputs at least one index information into a trained false volume identification model, outputs a probability A that the user is a false volume, compares the probability A with a preset probability range (A1-A2), when A exceeds A2, it is judged that the user is a false volume, when A1 The application identifies the advertisement click behavior of the user, obtains a false volume probability value, and further identifies the false volume of the user in the gray area, improves the precision of the mobile phone APP advertisement false volume identification, and effectively reduces the misjudgment rate and the omission rate.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and specifically relates to a method and system for identifying fake ad traffic in mobile apps. Background Technology

[0002] With the booming development of the mobile internet, mobile devices have gradually replaced PCs as the main entry point for internet traffic. As a result, mobile app advertising, which is the advertising displayed when users use apps on mobile devices, has become increasingly popular among advertisers.

[0003] In mobile advertising, billing is typically based on clicks (CPC). To maximize revenue, advertising platforms and media outlets often engage in fraudulent practices, resulting in a large amount of fake traffic for advertisers. This severely impacts advertisers' profits and disrupts market balance. Therefore, efficient anti-fraud methods and systems are crucial. Existing technologies offer various methods for identifying fake ad traffic, including analyzing IP addresses, user agents, geographic regions, click-through rates, click frequencies, cookie records, and access times. However, current methods for identifying fake ad traffic are relatively simplistic, resulting in high false positive and false negative rates. Furthermore, there are few reported methods specifically for identifying fake ad traffic in mobile app advertising. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method for identifying fake ad traffic in mobile apps, which can accurately identify fake ad traffic in mobile apps. The specific solution is as follows:

[0005] A method for identifying fake ad traffic in mobile apps, comprising the following steps:

[0006] Obtain the user's mobile app ad click records and get at least one of the following metrics within a preset time period: user's mobile device machine code, number of clicks on the ad, click-through rate of the ad, short-term fluctuation rate of the ID of the same mobile device machine code, and total ad clicks of the ID of the user's mobile device.

[0007] Input at least one of the above-mentioned indicator information into the trained fake data identification model, and output the probability A that the user is a fake data;

[0008] The probability A is compared with the preset probability range (A1-A2). When A exceeds A2, the user is judged to be a fake user.

[0009] Furthermore, before obtaining users' mobile app ad click records, the following steps are also included:

[0010] Select M user information entries, of which at least N entries have been identified as fake entries. Extract at least one of the following metrics from different dimensions within a preset time period: user's mobile device machine code, user's clicks on ads, user's click-through rate on ads, short-term volatility of the same mobile device machine code, and total ad clicks for the user's mobile device ID.

[0011] Set up a training set, with M training samples in each training set. Each training sample includes at least one indicator information. Input the M training samples into the model to be trained to obtain the trained spurious quantity identification model.

[0012] Preferably, the model to be trained is selected from one or more of gradient regression trees, logistic regression models, and random forests.

[0013] On the other hand, the present invention also provides a mobile APP advertising fake traffic identification system, the system comprising:

[0014] The sample acquisition module is used to extract at least one of the following indicators from M user information, of which at least N user information has been identified as fake data, within a preset time period: user's mobile device machine code, user's clicks on advertisements, user's click-through rate on advertisements, short-term volatility of the same mobile device machine code, and total number of advertisement clicks for the user's mobile device ID.

[0015] The model training module is used to input M training samples into the model to be trained for training, and obtain the trained spurious sample recognition model.

[0016] The record acquisition module acquires the user's mobile APP ad click records and obtains at least one of the following indicators within a preset time period: the user's mobile device machine code, the number of clicks on the ad, the click-through rate of the ad, the short-term fluctuation rate of the ID of the same mobile device machine code, and the total number of ad clicks for the ID of the user's mobile device.

[0017] The first fake user identification module inputs at least one of the above-mentioned indicator information into the trained fake user identification model and outputs the probability A that the user is a fake user; it compares the probability A with the preset probability range (A1-A2), and when A exceeds A2, it determines that the user is a fake user.

[0018] The mobile app ad fake traffic identification method and system provided by this invention, based on the characteristics of mobile app fake traffic, sets reasonable indicator information, adopts a trained fake traffic identification model, identifies users' ad click behavior, obtains fake traffic probability values, and further identifies fake traffic from users located in the gray area, thereby improving the accuracy of mobile app ad fake traffic identification and effectively reducing the false positive rate and false negative rate. Attached Figure Description

[0019] Figure 1 Example 1: Flowchart of a method for identifying fake ad impressions on mobile apps;

[0020] Figure 2 Example 2: Flowchart of the method for identifying fake ad traffic in mobile apps;

[0021] Figure 3 Example 4: Schematic diagram of the connections between modules in the mobile APP advertising fake traffic identification system;

[0022] Figure 4 Example 5: Schematic diagram of the connection between modules of the mobile APP advertising fake traffic identification system. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are only used to explain the invention and are not intended to limit the scope of protection of the present invention. In order to better illustrate the present invention, numerous specific details are given in the following specific implementation methods. Those skilled in the art should understand that the present invention can be implemented even without certain specific details. For methods, means, components and circuits well known to those skilled in the art, and for technical features that do not make creative contributions to the technical problem to be solved by the present invention, no detailed description is given. The "ad clicks" mentioned in the present invention refer to clicks on mobile APP advertisements.

[0024] Example 1

[0025] A method for identifying fake ad traffic in mobile apps, such as Figure 1 As shown, the method may include the following steps:

[0026] S101. Obtain the click records of a user clicking on advertisements through a mobile APP, and obtain indicator information of different dimensions within a preset time period, such as the user's mobile device machine code, the number of clicks on advertisements by the user, the click-through rate of advertisements by the user, the short-term fluctuation rate of ID of the same mobile device machine code, and the total number of advertisement clicks of the ID of the user's mobile device.

[0027] As mentioned above, the user's mobile device ID serves as the unique identifier for the phone. The number of times a user clicks on ads within a preset time period is counted. For example, if a user clicks on an ad for APP1 twice, an ad for APP2 twice, an ad for APP3 once, and an ad for APP4 once within a preset time period, the user's total clicks are 6. Generally, a normal user's click count will not exceed a certain threshold. The click-through rate (CTR) is the number of times the same ad is clicked repeatedly. For example, if the number of times an ad for APP1 is clicked repeatedly within a preset time period is 5, a normal user is unlikely to click the same ad repeatedly in a short period. The short-term fluctuation rate of the ID for the same mobile device ID is the rate at which the ID used by the same phone changes. For example, if the ID of the same mobile device changes 5 times within 10 minutes, a rapid rate of change increases the probability of the user being a fake user. The total number of ad clicks under the user's mobile device ID is also considered. If the total number of ad clicks under a single ID is too high, the probability of the user being a fake user increases.

[0028] S102. Input at least one of the above indicator information into the trained fake data identification model and output the probability A that the user is a fake data.

[0029] S103. Compare probability A with the preset probability range (A1-A2). When A exceeds A2, determine that the user is a fake user. The probability range (A1-A2) is not a fixed range. It depends on the indicator information input into the fake user identification model. Since the input indicator information is different, its value range will be different. Even for the same indicator, its probability range value can be dynamically adjusted.

[0030] To output the probability that a user is a fake user, the aforementioned indicator information needs to be input into the identification model. Therefore, before use, step S100 is required: collect a large amount of the aforementioned indicator information to train and optimize the fake user identification model so that its identification accuracy reaches a usable preset accuracy. For model training, multiple indicator information can be used together, or only one indicator information can be used. Therefore, multiple indicator information can be obtained and input into the fake user identification model for probability judgment, or only one indicator information can be used for probability judgment. Not all judgment results need to exceed a preset threshold; as long as one exceeds a certain threshold, it can be set as a user to be judged, improving the accuracy of the judgment and reducing the false judgment rate. For model selection, any existing model that can achieve the identification function can be used, such as autogradient regression trees, logistic regression models, or random forests. After determining that a user is a fake user, each indicator information of that user is marked and can be stored as training samples for application.

[0031] Example 2

[0032] like Figure 2 As shown, the mobile app ad fake traffic identification method provided in this embodiment differs from Embodiment 1 in that it further specifies that when A1 < A < A2, the user is defined as a user to be judged, and then further judgment operations are performed. The same user can have multiple fake traffic probability values ​​for multiple indicators. As long as one of the probabilities is within the above range, the user can be defined as a user to be judged. Of course, as long as one of the probabilities exceeds the maximum value of the probability range, it can be judged as fake traffic.

[0033] S201. Obtain the ad click records of the user to be judged, analyze the ad click records and the ad placement apps associated with them, classify the ad placement apps according to the ad type, and obtain the ad placement apps with information flow ads and other ad placement apps with non-information flow ads.

[0034] S202. Monitor the user's actions on the advertising app. When the user opens the information flow advertising app, proceed to step S203. When the user opens other advertising apps and clicks on ads, proceed to step S206.

[0035] S203. Divide the display interface into a normal information flow area and an advertising information flow area, obtain the average browsing speed V1 of the normal information flow area and the average browsing speed V2 of the advertising information flow area, and calculate the browsing speed ratio V1 / V2.

[0036] For apps that display news feeds, such as WeChat, Weibo, and Toutiao, the user interface typically displays news feeds. The normal news feed area is defined as the continuous, normal news feed area without ads, while the ad news feed area is defined as the area with native news feed ads. The ad news feed area can be defined based on the size of a typical mobile phone screen and the size of the ad news feed area; alternatively, it can be defined by extending upwards and downwards by a certain distance. User monitoring can be any existing monitoring solution, such as installing app plugins or client software on the user's device.

[0037] S204. When it is detected that a user to be judged has clicked on an advertisement, the obtained browsing speed ratio V1 / V2 is input into the trained speed judgment fake data identification model, and the probability B of judging the user as fake data is output. The speed judgment fake data identification model is a fake data identification model obtained by training the model using the browsing speed ratio V1 / V2 as training samples. Logistic regression model can be selected for training.

[0038] S205. Compare the probability B with the preset threshold B1. When B is greater than B1, determine that the user is a fake user.

[0039] S206. Obtain N Uniform Resource Locator (URLs) for accessed pages at preset time intervals, parse each URL according to preset parsing rules, and obtain the access level X for each URL.

[0040] S207. Input the obtained N access layer results X into the trained hierarchical judgment fake quantity identification model, and output the probability C of the user being judged as a fake quantity; the hierarchical judgment fake quantity identification model is the fake quantity identification model obtained after training the model using the access layer results as training samples. Positive samples and negative samples can be selected, and the gradient boosting regression tree model can be selected for training.

[0041] The preset time interval and number of retrievals N can be customized according to the nature of the APP and the fake traffic identification system. For example, retrieval can be performed once every 2 minutes, for a total of 5 times. Each time, the APP's URL is crawled, and then the Uniform Resource Locator (URL) is parsed. All APP web pages have a hierarchical design. Generally, the homepage is level 0. Further analysis can be performed to the true level of the web page, such as level 1, level 2, level 3, etc. During the URL analysis process, search keyword analysis is also performed. If the URL of a certain level has different search keywords compared with the previous level, its access level is set to the true level plus 1.

[0042] As an example, taking the mobile version of Tmall as an example, the main page is set as level 0, the next level page entered from the main page is level 1, the actual level of the page obtained by actively entering the search term is level 1, the page level is defined as level 2, the next level page entered after clicking on the page after the search is level 2, and if the search is further performed, the page is level 3.

[0043] S208. Compare the probability C with the preset threshold C1. When C is greater than C1, determine that the user is a fake user.

[0044] The mobile app ad fake traffic identification method provided in this embodiment can locate a user as a user to be judged if any probability A falls into the grayscale value range. Then, the user is further judged. Different indicator information is selected for different ad formats to make further judgments, which greatly improves the accuracy of fake traffic identification, reduces false judgments and false omissions, and is more suitable for fake traffic identification of mobile app ads.

[0045] Example 3

[0046] The method for identifying fake ad traffic in mobile apps provided in this embodiment differs from that in Embodiment 2 in that the method further includes the following steps:

[0047] The short-term volatility of the ID of the same device machine code is compared with a preset threshold. If it exceeds the preset threshold, other device machine codes with the same ID are obtained, and the users corresponding to the other device machine codes are defined as users to be judged.

[0048] The total number of ad clicks for the user's mobile device ID is compared with a preset threshold. If the total number of clicks exceeds the preset threshold, the machine codes of other devices using the corresponding ID are obtained, and the users corresponding to the other device machine codes are defined as users to be judged.

[0049] Once a user is defined as a user to be judged, step S201 is performed. The ad fraud detection method in this embodiment is based on the examination of fraudulent groups. When a user's ID changes frequently or the total number of ad clicks for an ID is too large, there is a high probability that there is an ad fraudulent group. Therefore, the corresponding user is found without the need to perform the initial probability A judgment. The user to be judged is directly judged, which further improves the accuracy of ad fraud detection.

[0050] Example 4

[0051] like Figure 3 As shown, this embodiment provides a mobile APP advertising fake traffic identification system, including a sample acquisition module 401, a model training module 402, a record acquisition module 404, and a first fake traffic judgment module 403;

[0052] The sample acquisition module 401 is used to extract at least one of the following indicators from M user information, of which at least N user information has been judged as fake data, within a preset time period: user's mobile device machine code, user's clicks on advertisements, user's click-through rate on advertisements, short-term volatility of the same mobile device machine code, and total number of advertisement clicks for the user's mobile device ID.

[0053] The model training module 402 is used to input M training samples into the model to be trained for training, and obtain the trained spurious sample recognition model.

[0054] The record acquisition module 404 acquires the user's mobile APP ad click records and obtains at least one of the following indicators within a preset time period: the user's mobile device machine code, the number of clicks on the ad by the user, the click-through rate of the ad by the user, the short-term fluctuation rate of the ID of the same mobile device machine code, and the total number of ad clicks by the ID of the user's mobile device.

[0055] The first fake data judgment module 403 inputs at least one of the above-mentioned indicator information into the trained fake data recognition model and outputs the probability A that the user is a fake data; compares the probability A with the preset probability range (A1-A2), and judges the user as a fake data when A exceeds A2.

[0056] Example 5

[0057] like Figure 4 As shown, the mobile APP advertising fake traffic identification system provided in this embodiment differs from that in embodiment 4 in that it further includes an advertising APP classification module 501, a monitoring module 502, a browsing speed calculation module 503, a second fake traffic judgment module 504, a URL acquisition module 505, an access layer parsing module 506, and a third fake traffic judgment module 507.

[0058] The first fake quantity judgment module 403 is also used to define the user as the user to be judged when it is judged that A1 < A < A2, and send a classification instruction to the advertising APP classification module 501.

[0059] The advertising APP classification module 501 is used to receive classification instructions, obtain the advertising click records of the user to be judged, analyze the advertising APPs associated with the advertising click records, classify the advertising APPs according to the advertising type, obtain the information flow advertising APPs with information flow ads and other advertising APPs with non-information flow ads, and send monitoring instructions to the monitoring module 502.

[0060] The monitoring module 502 is used to receive monitoring instructions. When it detects that the user to be judged opens the information flow advertising app, it sends a calculation instruction to the browsing speed calculation module 503. When it detects that the user to be judged clicks on an ad in the corresponding information flow advertising app, it sends a judgment instruction to the second fake data judgment module 504. When it detects that the user to be judged opens another advertising app and clicks on an ad, it sends a retrieval instruction to the URL retrieval module 505.

[0061] The browsing speed calculation module 503 is used to receive calculation instructions, divide the display interface into a normal information flow area and an advertising information flow area, obtain the average browsing speed V1 of the normal information flow area and the average browsing speed V2 of the advertising information flow area, and calculate the browsing speed ratio V1 / V2.

[0062] The second fake data judgment module 504 is used to receive judgment instructions, input the obtained browsing speed V1 / V2 into the trained speed judgment fake data recognition model, and output the probability B of judging the user as fake data; compare the probability B with the preset threshold B1, and judge the user as fake data when B is greater than B1.

[0063] The URL acquisition module 505 is used to receive acquisition instructions, acquire N Uniform Resource Locator URLs of access pages according to a preset time interval, and then send a parsing instruction to the access layer parsing module 506.

[0064] The access layer parsing module 506 is used to receive parsing instructions, parse each URL according to preset parsing rules, obtain the access layer X of each URL, and then send the identification instructions to the third fake data identification module.

[0065] The third fake user identification module 507 is used to receive identification instructions, input the obtained N access layer results X into the trained layer identification fake user identification model, output the probability C of the user being identified as a fake user, compare the probability C with the preset threshold C1, and when C is greater than C1, the user is identified as a fake user.

[0066] The mobile app advertising fake traffic identification system provided in this embodiment can locate a user as a user to be judged if any probability A falls into the grayscale value range. Then, the user is further judged. Different indicator information is selected for different advertising formats to make further judgments, which greatly improves the accuracy of fake traffic identification, reduces false judgments and false omissions, and is more suitable for identifying fake traffic in mobile app advertising.

Claims

1. A method for identifying fake ad impressions in mobile apps, characterized in that, The method includes the following steps: Obtain the user's mobile app ad click records and get at least one of the following metrics within a preset time period: user's mobile device machine code, number of clicks on the ad, click-through rate of the ad, short-term fluctuation rate of the ID of the same mobile device machine code, and total ad clicks of the ID of the user's mobile device. Input at least one of the above-mentioned indicator information into the trained fake data identification model, and output the probability A that the user is a fake data; The probability A is compared with the preset probability range (A1-A2). When A exceeds A2, the user is judged to be a fake user. Before obtaining a user's mobile app ad click history, the following steps are also included: Select M user information entries, of which at least N entries have been identified as fake entries. Extract at least one of the following metrics from different dimensions within a preset time period: user's mobile device machine code, user's clicks on ads, user's click-through rate on ads, short-term volatility of the same mobile device machine code, and total ad clicks for the user's mobile device ID. Set up a training set, each training set has M training samples, each training sample includes at least one of the aforementioned indicator information, input the M training samples into the model to be trained for training, and obtain the trained spurious quantity identification model. When A1 < A < A2, the user is defined as the user to be judged, and the following steps are performed: Obtain the ad click records of the user to be judged, analyze the ad click records and the ad placement apps associated with them, and classify the ad placement apps according to the ad type to obtain the ad placement apps with information flow ads and other ad placement apps with non-information flow ads; Monitor the user's actions on the ad-serving APP. When the user opens the ad-serving APP, the display interface is divided into a normal information flow area and an ad information flow area. The average browsing speed V1 of the normal information flow area and the average browsing speed V2 of the ad information flow area are obtained, and the browsing speed ratio V1 / V2 is calculated. When the user to be judged is detected to have clicked on an advertisement, the obtained browsing speed V1 / V2 is input into the trained speed judgment fake data identification model, and the probability B of the user being judged to be fake data is output. The probability B is compared with a preset threshold B1. When B is greater than B1, the user is determined to be a fake user. When it is detected that a user to be judged opens another advertising app and clicks on an ad, N Uniform Resource Locator (URLs) of the access pages are obtained at preset time intervals. Each URL is parsed according to preset parsing rules to obtain the access level X of each URL. i Where i = 1, ..., N, the N access layer results X are input into the trained layer judgment fake quantity identification model, and the probability C of the judged user is a fake quantity is output; The probability C is compared with a preset threshold C1. When C is greater than C1, the user is determined to be a fake user.

2. The method for identifying fake ad impressions in mobile apps as described in claim 1, characterized in that, When any of the URLs contains a search keyword that is different from the previous level, the access level X is determined. i It equals the actual number of accessed levels plus 1.

3. The method for identifying fake ad impressions in mobile apps as described in claim 2, characterized in that, The method further includes the following steps: The short-term volatility of the ID of the same device machine code is compared with a preset threshold. If it exceeds the preset threshold, other device machine codes with the same ID are obtained, and the users corresponding to the other device machine codes are defined as users to be judged. The total number of ad clicks for the user's mobile device ID is compared with a preset threshold. If the total number of clicks exceeds the preset threshold, the machine codes of other devices using the corresponding ID are obtained, and the users corresponding to the other device machine codes are defined as users to be judged.

4. A mobile app advertising fake traffic identification system, characterized in that, The system includes: The sample acquisition module (401) is used to extract at least one of the following indicators from M user information, of which at least N user information has been judged as fake data, within a preset time period: user's mobile device machine code, user's clicks on advertisements, user's click-through rate on advertisements, short-term volatility of the same mobile device machine code, and total number of advertisement clicks for the user's mobile device ID. The model training module (402) is used to input M training samples into the model to be trained for training, and obtain the trained spurious quantity recognition model. The record acquisition module (404) acquires the user's mobile APP ad click records and obtains at least one of the following indicators within a preset time period: the user's mobile device machine code, the number of clicks on the ad by the user, the click-through rate of the ad by the user, the short-term fluctuation rate of the ID of the same mobile device machine code, and the total number of ad clicks of the ID of the user's mobile device. The first fake data judgment module (403) inputs at least one of the above-mentioned indicator information into the trained fake data recognition model and outputs the probability A that the user is a fake data; compares the probability A with the preset probability range (A1-A2), and judges the user as a fake data when A exceeds A2; The system also includes an advertising APP classification module (501), a monitoring module (502), a browsing speed calculation module (503), and a second fake volume judgment module (504). The first fake quantity judgment module (403) is also used to define the user as the user to be judged when it is judged that A1 < A < A2, and send a classification instruction to the advertising APP classification module (501); The advertising APP classification module (501) is used to receive classification instructions, obtain the advertising click records of the user to be judged, analyze the advertising APP associated with the advertising click records, classify the advertising APP according to the advertising type, obtain the information flow advertising APP with information flow ads and other advertising APP with non-information flow ads, and send monitoring instructions to the monitoring module (502). The monitoring module (502) is used to receive monitoring instructions. When it is monitored that the user to be judged opens the information flow advertising APP, it sends a calculation instruction to the browsing speed calculation module (503); and when it is monitored that the user to be judged generates an advertising click behavior in the corresponding information flow advertising APP, it sends a judgment instruction to the second fake quantity judgment module (504). The browsing speed calculation module (503) is used to receive calculation instructions, divide the display interface into a normal information flow area and an advertising information flow area, obtain the average browsing speed V1 of the normal information flow area and the average browsing speed V2 of the advertising information flow area, and calculate the browsing speed ratio V1 / V2. The second fake data judgment module (504) is used to receive judgment instructions, input the obtained browsing speed V1 / V2 into the trained speed judgment fake data recognition model, and output the probability B of judging the user as fake data; compare the probability B with the preset threshold B1, and judge the user as fake data when B is greater than B1. The system also includes a URL acquisition module (505), an access layer parsing module (506), and a third fake quantity judgment module (507). The monitoring module (502) is also used to send a retrieval instruction to the URL retrieval module (505) when it detects that the user to be judged opens the other advertising APP and generates an advertising click behavior; The URL acquisition module (505) is used to receive acquisition instructions, acquire N Uniform Resource Locator URLs of access pages according to a preset time interval, and then send a parsing instruction to the access layer parsing module (506). The access layer parsing module (506) is used to receive parsing instructions, parse each URL according to preset parsing rules, and obtain the access layer X of each URL. i , where i = 1, ..., N, and then send the identification instruction to the third spurious quantity identification module; The third spurious quantity judgment module (507) is used to receive the identification instruction and process the obtained N access layer results X1, X2, ..., X n The input is fed into the trained hierarchical judgment fake data identification model, which outputs the probability C of judging the user as a fake data. The probability C is compared with the preset threshold C1. When C is greater than C1, the user is judged to be a fake data.

Images