An automatic content rating method for Android Chinese application market

Abstract

The application discloses an automatic content classification method for an Android Chinese application market, and comprises the following steps: acquiring application source data; wherein the application source data is used for representing attribute information of an application; extracting features of the application source data; inputting the features into a trained fusion classification model, outputting a predicted category probability set through the fusion classification model, and obtaining a classification result according to the predicted category probability set. The application extracts multi-dimensional features of the application source data, so that the extracted feature information is very rich, and the features are input into the trained fusion classification model to obtain an accurate predicted category probability set, and finally, the classification result is obtained, so that the Android Chinese application market can be classified according to content.

Description

Technical Field

[0001] This invention relates to the field of machine learning software engineering technology, and in particular to an automatic content classification method for the Android Chinese application market. Background Technology

[0002] Smartphones play a vital role in people's daily lives. Through a variety of mobile applications with specific functions, people can engage in social interaction, entertainment, shopping, work, and more. While smartphones can be used by people of any age, the content or functions of each application may not be suitable for everyone. For example, applications containing inappropriate content such as pornography, gambling, or drugs should not be provided to children or teenagers. Therefore, mobile applications should be appropriately classified into content rating systems. Existing rating models are either incompatible with the Android Chinese app market or suffer from low accuracy because the extracted features are derived from the app market rather than the applications themselves.

[0003] Therefore, existing technologies still need improvement and development. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an automatic content classification method for the Android Chinese application market, which addresses the above-mentioned deficiencies of the prior art. This method aims to solve the problem that the existing model classification methods are either not applicable to the Android Chinese application market or have low content classification accuracy because the extracted features come from the application market rather than from the application itself.

[0005] The technical solution adopted by this invention to solve the problem is as follows:

[0006] In a first aspect, embodiments of the present invention provide an automatic content rating method for the Android Chinese application market, wherein the method includes:

[0007] Obtain application source data; wherein the application source data is used to characterize the attribute information of the application;

[0008] Extract features from the application source data;

[0009] The features are input into a trained fusion classification model, which outputs a predicted category probability set. The classification result is obtained based on the predicted category probability set.

[0010] In one implementation, the features of extracting the application source data include:

[0011] When the source data of the application is text, the text is input into a preset word vector model, and the word vector model outputs features corresponding to the text.

[0012] When the application source data is an image, the image is filtered, and image features are extracted from the filtered image to obtain features corresponding to the image.

[0013] In one implementation, the generation process of the fusion classification model includes:

[0014] Obtain several training feature types; where each training feature type contains several training feature data.

[0015] For each training feature type, several initial classification models are obtained, and several training feature data from each training feature type are input into each initial classification model. The accuracy of several models corresponding to the training feature type is obtained through the initial classification models.

[0016] Based on the accuracy of several models, a target classification model corresponding to the training feature type is obtained;

[0017] By fusing the target classification models corresponding to several training feature types, a fused classification model is obtained.

[0018] In one implementation, the step of inputting several training feature data from each training feature type into each initial classification model, and obtaining several model accuracies corresponding to the training feature type through the initial classification model includes:

[0019] Obtain the true classification results of several training feature data in each training feature type;

[0020] Input several training feature data into each initial classification model to obtain the predicted classification result corresponding to several training feature data in each training feature type;

[0021] The number of cases where the true classification results of several training feature data in each training feature type are the same as the predicted classification results of several training feature data in each training feature type is obtained by counting the number of cases where the true classification results are the same as the predicted classification results of several training feature data in each training feature type.

[0022] Based on the number of accurately predicted classifications after each training feature type is classified by each initial classification model, the accuracy of several models corresponding to the training feature type is obtained.

[0023] In one implementation, obtaining the accuracy of several models corresponding to the training feature type based on the number of accurately predicted classifications after each initial classification model for each training feature type includes:

[0024] The accuracy of each training feature type after being classified by each initial classification model is obtained by calculating the number of accurately predicted classifications after each training feature type is classified by each initial classification model.

[0025] Based on the accuracy of each training feature type after being classified by each initial classification model, the accuracy of several models corresponding to the training feature type is obtained.

[0026] In one implementation, the step of obtaining the accuracy of each training feature type after classification by each initial classification model based on the accurately predicted number of classifications for each training feature type includes:

[0027] Obtain the total number of training feature data in each training feature type;

[0028] The accuracy of each training feature type after classification by each initial classification model is obtained by dividing the number of accurate predicted classifications of each training feature type by the total number of training feature data in each training feature type.

[0029] In one implementation, obtaining the target classification model corresponding to the training feature type based on the accuracy of several of the models includes:

[0030] Select the model with the highest accuracy from the aforementioned model accuracies;

[0031] The initial classification model with the highest model accuracy is used as the target classification model corresponding to the training feature type.

[0032] In one implementation, fusing the target classification models corresponding to several of the training feature types to obtain a fused classification model includes:

[0033] Based on a preset algorithm, the target classification models corresponding to several training feature types are accumulated to obtain a fusion classification model.

[0034] Secondly, embodiments of the present invention also provide an automatic content rating device for the Android Chinese application market, wherein the device includes:

[0035] An application source data acquisition module is used to acquire application source data; wherein, the application source data is used to characterize the attribute information of the application;

[0036] The extraction module is used to extract features from the application source data;

[0037] The predicted category probability set output module is used to input the features into the trained fusion classification model, output the predicted category probability set through the fusion classification model, and obtain the classification result based on the predicted category probability set.

[0038] Thirdly, embodiments of the present invention also provide a smart terminal, including a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs including an automatic content rating method for the Android Chinese application market as described in any of the above.

[0039] Fourthly, embodiments of the present invention also provide a non-transitory computer-readable storage medium, wherein when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is able to perform the automatic content rating method for the Android Chinese application market as described in any of the above.

[0040] The beneficial effects of this invention are as follows: First, application source data is acquired; wherein, the application source data is used to characterize the attribute information of the application; then, features of the application source data are extracted; finally, the features are input into a trained fusion classification model, the fusion classification model outputs a predicted category probability set, and a classification result is obtained based on the predicted category probability set. It is evident that this invention, by extracting features from multi-dimensional application source data, yields very rich feature information, and by inputting the features into a trained fusion classification model, an accurate predicted category probability set is obtained, enabling content classification for the Android Chinese application market. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of the automatic content classification method for the Android Chinese application market provided in an embodiment of the present invention.

[0043] Figure 2 This is an overall flowchart of an implementation method for automatic content classification provided in an embodiment of the present invention.

[0044] Figure 3 This is an overall flowchart of a training classifier implementation method provided in an embodiment of the present invention.

[0045] Figure 4 This is a schematic diagram of an automatic content rating device for the Android Chinese application market provided in an embodiment of the present invention.

[0046] Figure 5 This is a block diagram illustrating the internal structure of a smart terminal provided in an embodiment of the present invention. Detailed Implementation

[0047] This invention discloses an automatic content rating method for the Android Chinese application market. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention.

[0048] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0049] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0050] With the increasing prevalence of younger mobile phone users, mobile applications need to be appropriately age-ratinged to protect children's online safety. Android and iOS are the mainstream smartphone operating systems, and their official app stores, Google Play and the App Store, have age rating policies, displaying the rating on the app's details page to provide appropriate information to suitable users. The App Store's age ratings are generally considered more accurate because they have dedicated staff who review each submitted app and rate it according to their policies. However, manual rating requires significant time and effort and errors can still occur. On the other hand, age ratings on Google Play are determined by the developers themselves, making inappropriate ratings even more common.

[0051] Unfortunately, there is currently very little research in academia on automatic age rating technology for mobile applications. Chen et al. paired applications in the App Store and Google Play, using the age ratings provided by the App Store as the ground truth labels to train a classifier ALM to verify the accuracy of the age ratings for corresponding applications on Google Play. ALM is based on keyword matching, with keywords selected manually from application descriptions and user reviews. The model is trained separately for each rating and automatically adjusts the keyword weights, achieving an accuracy of 93.9%-100% on 1464 game applications. Hu et al. added two types of text features to ALM: a word vector model to capture synonyms for keywords and a bag-of-words model to describe the context. They proposed a two-stage multi-label machine learning framework, AAMR, which first predicts inappropriate content in applications based on rating policies and applications, and then predicts age ratings from inappropriate content. The model achieved an accuracy of 79% in tests on Google Play and the App Store. Liu et al. designed a machine learning model to predict whether mobile applications are designed for children. The model uses features involving more data types, such as application categories, application icons, and screenshots. They obtained apps suitable for children and adults on Google Play by searching keywords on search engines, and the model eventually achieved 95% accuracy on 1,728 apps.

[0052] Existing technologies share two main drawbacks: 1) the models are only applicable to English-language app stores; 2) the features extracted from apps only focus on information obtainable from the app store, without delving into the app itself. However, for Android phone users, official app stores are not the only way to obtain apps. Numerous third-party app stores offer Android app downloads, and in some countries, Google Play is blocked, forcing users to obtain apps only through unofficial channels. China, as a populous country with a considerable number of mobile phone users, lacks access to Google Play, making third-party app stores the primary app download channel for Chinese mobile phone users. These app stores themselves lack unified standards; for example, among common Chinese Android app stores, only one provides age ratings for apps. Furthermore, the quality of apps offered by these app stores varies greatly, and the actual behavior of the apps may differ from their appearance in the app stores. Therefore, underage mobile phone users in China are more likely to be exposed to inappropriate content in mobile apps. We aim to propose a technology that can identify whether apps in the Chinese Android app store are suitable for children.

[0053] To address the problems of existing technologies, this embodiment provides an automatic content classification method for the Android Chinese application market. This invention extracts features from multi-dimensional application source data, resulting in rich feature information. These features are then input into a trained fusion classification model to obtain an accurate predicted category probability set, enabling content classification for the Android Chinese application market. Specifically, the method first acquires application source data, which represents the application's attribute information. Then, features are extracted from the application source data. Finally, these features are input into a trained fusion classification model, which outputs a predicted category probability set. The classification result is then obtained based on this predicted category probability set.

[0054] Exemplary methods

[0055] This embodiment provides an automatic content rating method for the Android Chinese application market. This method can be applied to smart terminals using machine learning software engineering. Specifically, as follows... Figure 1 As shown, the method includes:

[0056] Step S100: Obtain application source data; wherein, the application source data is used to characterize the attribute information of the application;

[0057] Specifically, the application's attribute information includes the application category, ad tags, application description, application icon, and screenshots from the Android Chinese app market; image resources in the application resource file; application permissions in the application manifest file; and application programming interfaces (APIs) in the application source code. Existing technologies match applications from unrated app markets with those from rated ones to obtain their existing rating. However, this method is not feasible in the Chinese Android app market because different app markets offer significantly different application sets, and application repackaging is common, meaning that apps with identical appearances may have internal differences. To address these issues and minimize the deceptiveness of information provided by app markets, this application obtains application source data not only from the Android Chinese app market but also from the application itself. Furthermore, this invention uses static analysis, processing all APKs with the APK decompilation tool AndroGuard to obtain various types of features, such as application usage permissions from the application manifest file and APIs used by the application from the application source code. In practice, application usage permissions refer to data permissions involving money or privacy, such as payment data. Application programming interfaces (APIs) are used to filter out applications that do not contribute to the categorization. If 80% of the application source code uses the same API, the application is considered a generic application and is not categorized. To enrich the source data, the following are obtained: application categories, ad tags, application descriptions, application icons, application screenshots, image resources in application resource files, application permissions in application manifest files, and APIs in application source code from the Android Chinese application market. In this embodiment, the Android Chinese application market includes Google Play Store, Tencent App Store, Huawei AppGallery, and 360 Mobile Assistant. Among them, Google Play is the official Android application market, while Tencent App Store, Huawei AppGallery, and 360 AppGallery are mainstream third-party Android application markets in China; Google Play and Huawei AppGallery provide age ratings for applications, while Tencent App Store and 360 AppGallery do not. For each Android Chinese application market, this invention uses a customized crawler to crawl the basic information of the applications in the application market and the download links of application resources (APK, icon, screenshots), and then downloads the application resources to the local machine.For Huawei and 360 Mobile Assistant, the crawler retrieves information on all accessible applications by category. App Store only retrieves a small number of applications by category. The crawled application names are used as keywords to search on App Store to access more applications, and the newly added application names are used as new keywords for searching. This process is repeated until all application names have been searched. Google Play also returns only a small number of applications by category. However, the third-party service AndroZoo continuously collects applications from several Android app markets, including Google Play. Therefore, this invention uses the application list provided by AndroZoo to obtain information on the latest versions of all applications from Google Play, identifying over two million different applications. This invention samples from these, taking 1000 applications from each application category based on download volume.

[0058] Once the application source data is obtained, it can be executed as follows: Figure 1 The following steps are shown: S200, extracting features from the application source data; correspondingly, extracting features from the application source data includes the following steps:

[0059] S201. When the application source data is text, the text is input into a preset word vector model, and the word vector model outputs features corresponding to the text.

[0060] S202. When the application source data is an image, the image is filtered, and the filtered image is subjected to image feature extraction to obtain the features corresponding to the image.

[0061] Specifically, the characteristics are shown in Table 1:

[0062] Table 1. Overview of Features

[0063]

[0064] When the application source data is text, the text is input into a preset word vector model, and the word vector model outputs features corresponding to the text. The word vector model includes Word2Vec and Bert. In one implementation, the text can also be input into a preset bag-of-words model, and the bag-of-words model outputs features corresponding to the text. When the application source data is an image, the image is filtered, and image features are extracted from the filtered image to obtain features corresponding to the image. In this embodiment, images, as complex data, require additional processing: filtering the image and then extracting image features from the filtered image. For example, there are three types of images related to each application: icons, screenshots, and image resources. Icons and screenshots are obtained from the application market, and image resources are image format files found by uncompressing the APK and traversing the file directory. Both the jar tool and the unzip command can be used to uncompress APKs, but the unzip command cannot handle multi-part compressed packages and garbled characters, so this invention chooses the jar tool. An application may contain a large number of image resources, but not all of them are helpful for extracting application features in this invention. For example, a solid-color image cannot provide any useful information to the classifier; the more complex the image, the more information it contains, and the more worthy of further analysis; images that are too small are also unlikely to contain useful information; two images that are too similar do not need to be retained, only one needs to be kept. Therefore, in image filtering, 1) for image resources obtained from each application, this invention first filters by size, retaining only images larger than 144x144 (pixels) (the size of application icons in mainstream screen densities of xxhdpi); 2) then selects images in descending order of their Shannon entropy (which describes the average information content of the image); 3) calculates the hash value of the image; if the current image is too similar to any of the selected images (i.e., the hash value difference is less than a threshold, which can be 20%), then the current image is not retained; 4) this process continues until the total number of selected images reaches 20, or all image resources have been processed. In practice, application icons and screenshots are selected by default. Therefore, after image filtering, for each application, including application icons and screenshots, this invention retains a maximum of 20 sufficiently large and dissimilar complex images. After filtering, image features are extracted from these retained images. These image features are categorized into various types, such as tag attributes and text attributes. Image tags can be used to annotate images using the Google Vision API, and text in the images can be extracted using Optical Character Recognition (OCR) technology.In practice, OCR technology's accuracy in recognizing Chinese characters is slightly lower than its accuracy in recognizing English characters because it has a lower accuracy rate for Chinese characters, especially complex Chinese characters or fonts. It also separates each Chinese character and can only process images horizontally, leading to the loss of existing word segmentation information, especially when the text is not arranged horizontally. Therefore, for the original text obtained by OCR technology, this invention requires a preprocessing step: retaining only the Chinese characters, English letters, and numbers, then concatenating them, and adding spaces on both sides of each string of numbers and letters. Each application should only perform feature extraction once, after which various features are grouped and stored in a structured text document for efficient reading by the program. After feature extraction, feature selection and preprocessing are also required to improve the classifier's performance. For application categories, we customized a mapping for each Android Chinese app market to unify their different categories; for application permissions, we only considered permissions that Android classified as "dangerous"; for application programming interfaces (APIs), we only considered APIs that were highly relevant to this classification question (which can be measured by the chi-square test), and filtered out the commonly used APIs (which can be done with the help of term frequency-inverse document frequency (TF-IDF)).

[0065] After obtaining the characteristics of the application source data, the following can be executed: Figure 1 The following steps are shown: S300, input the features into the trained fusion classification model, output the predicted category probability set through the fusion classification model, and obtain the classification result based on the predicted category probability set.

[0066] Since the extracted features are very rich, inputting these features into the trained fusion classification model allows the model to output an accurate predicted category probability set. In this embodiment, the trained classification model can be tested first or used in practice. Because App Store and 360 App Market do not provide age ratings, datasets from these platforms can be input into the fusion classification model of this invention to obtain the predicted category probability set. The predicted category probability set can be divided into the probability that the app is suitable for children under 12 years old and the probability that the app is not suitable for children under 12 years old. That is, this embodiment uses binary classification, but in practice, multi-classification can be performed according to actual needs. Finally, the rating result is obtained based on the predicted category probability set. Thus, when the probability that the app is suitable for children under 12 years old is greater than or equal to the probability that the app is not suitable for children under 12 years old, the rating result is that the app is suitable for children under 12 years old; otherwise, the rating result is that the app is not suitable for children under 12 years old. For some apps that exist in Tencent App Store and 360 App Store but not in Huawei App Store, dynamic analysis and manual labeling can be used: first, use an automated testing tool (such as Monkey) to run the app and generate several runtime screenshots, and then have people with relevant domain knowledge view these screenshots. If an app has at least one screenshot labeled as unsuitable for children under 12, then the app is considered unsuitable for children under 12. The reliability of this test depends on whether the screenshots generated by the testing tool are representative.

[0067] In one implementation, the generation process of the fusion classification model includes the following steps: obtaining several training feature types; wherein each training feature type contains several training feature data; for each training feature type, obtaining several initial classification models, and inputting several training feature data from each training feature type into each initial classification model, obtaining several model accuracies corresponding to the training feature type through the initial classification models; obtaining a target classification model corresponding to the training feature type based on the several model accuracies; and fusing the target classification models corresponding to the several training feature types to obtain the fusion classification model.

[0068] Specifically, such as Figure 2-3As shown, since Huawei AppGallery provides age ratings, this embodiment is based on Huawei AppGallery for training. In practice, Google Play Store can also be used for model training, as Google Play Store also provides age ratings. First, several training feature types are obtained. In this embodiment, the training feature types are: 7 types including: app category mappings to Huawei app categories (an ad tag is a binary value indicating whether an ad exists); word vectors extracted from app descriptions; image annotations and text word vectors extracted from app icons, app screenshots, and image resources; Android danger level permissions extracted from app permissions; and uncommon APIs extracted from app APIs (Application Programming Interfaces). For each of the 7 training feature types, each training feature type contains several training feature data. In this embodiment, each training feature type contains 1000 positive and 1000 negative examples of app feature data. Positive examples refer to data on apps suitable for children in the rating results, and negative examples refer to data on apps unsuitable for children in the rating results. For each training feature type, several initial classification models are obtained. In this embodiment, the several initial classification models are decision tree, logistic regression, support vector machine, K-nearest neighbor, random forest and neural network. In practice, several training feature data from each training feature type are input into each initial classification model, and the accuracy of several models corresponding to the training feature type is obtained through the initial classification model. The process of inputting several training feature data from each training feature type into each initial classification model and obtaining the accuracy of several models corresponding to the training feature type includes the following steps: obtaining the true classification results of several training feature data in each training feature type; inputting several training feature data into each initial classification model to obtain the predicted classification results corresponding to several training feature data in each training feature type; counting the number of instances where the true classification results and the predicted classification results are the same for several training feature data in each training feature type, thus obtaining the number of accurately predicted classifications for each training feature type after classification by each initial classification model; and obtaining the accuracy of several models corresponding to the training feature type based on the number of accurately predicted classifications for each training feature type after classification by each initial classification model.

[0069] In practice, the true classification results of several training feature data in each training feature type are known. That is, the classification corresponding to each training feature data in each training feature type is known and determined as to be applicable to or not applicable to children. When several training feature data are input into each initial classification model, the predicted classification results corresponding to several training feature data in each training feature type are obtained. At this time, the number of instances where the true classification results and the predicted classification results of several training feature data in each training feature type are the same can be counted to obtain the accurate predicted classification number of each training feature type after classification by each initial classification model. For example, when the true classification result of one training feature data in each training feature type is applicable to children (in this embodiment, the age range of children is under 12 years old), and the predicted classification result of one training feature data in each training feature type is also applicable to children... For children; or when the true classification result of a training feature data in each training feature type is "not applicable to children," and the predicted classification result of a training feature data in each training feature type is also "not applicable to children," this indicates that the true classification result of a training feature data in each training feature type and the predicted classification result of a training feature data in each training feature type are the same. By counting the number of such identical results, the number of accurately predicted classifications after each initial classification model is obtained for each training feature type. Then, based on the number of accurately predicted classifications after each initial classification model is used for each training feature type, the accuracy of several models corresponding to the training feature type is obtained. Accordingly, obtaining the accuracy of several models corresponding to the training feature type based on the number of accurately predicted classifications after each initial classification model includes the following steps: obtaining the total number of training feature data in each training feature type; dividing the number of accurately predicted classifications after each initial classification model is used for each training feature type by the total number of training feature data in each training feature type to obtain the accuracy of each training feature type after each initial classification model; and obtaining the accuracy of several models corresponding to the training feature type based on the accuracy of each training feature type after each initial classification model.

[0070] Specifically, the total number of training feature data in each training feature type is first obtained. In this embodiment, since each training feature type contains 1000 positive and 1000 negative application feature data, the total number of training feature data in each training feature type is 2000. The number of accurate predicted classifications after each training feature type is classified by each initial classification model is divided by the total number of training feature data in each training feature type to obtain the accuracy of each training feature type after classification by each initial classification model. For example, when the number of accurately predicted categories for each training feature type after classification by a decision tree is 1500, dividing 1500 by 2000 yields 75%, so the accuracy of each training feature type after classification by a decision tree is 75%. Similarly, we can obtain the accuracy of each training feature type after classification by logistic regression, support vector machine, K-nearest neighbors, random forest, and neural network. This yields the accuracy of several models corresponding to the training feature types. In this embodiment, the accuracy of each model is obtained after parameter tuning using several training feature data. Besides accuracy, evaluation metrics can also include f1 or AUC.

[0071] After obtaining several model accuracies, a target classification model corresponding to the training feature type can be obtained based on these accuracies. Accordingly, obtaining the target classification model corresponding to the training feature type based on the several model accuracies includes the following steps: selecting the model with the highest accuracy from the several model accuracies; and using the initial classification model corresponding to the highest model accuracy as the target classification model corresponding to the training feature type. For example, from the following obtained accuracy rates: 75% after decision tree classification, 70% after logistic regression classification, 65% after support vector machine classification, 60% after K-nearest neighbor classification, 55% after random forest classification, and 50% after neural network classification, the one with the highest accuracy is selected, and the initial classification model (such as a decision tree) corresponding to the one with the highest accuracy is used as the target classification model corresponding to the training feature type. For each training feature type, a corresponding target classification model is obtained. In practice, since there are 7 training feature types, a total of 7 target classification models can be obtained from the 7 training feature types.

[0072] After obtaining the target classification model, the target classification models corresponding to several training feature types are fused to obtain a fused classification model. Accordingly, fusing the target classification models corresponding to several training feature types to obtain the fused classification model includes the following steps: based on a preset algorithm, the target classification models corresponding to several training feature types are accumulated to obtain the fused classification model.

[0073] In this embodiment, the preset algorithm is a soft voting algorithm. The soft voting algorithm can be used to accumulate the target classification models corresponding to several training feature types; or the soft voting algorithm can be used to calculate the weighted average of the target classification models corresponding to several training feature types to obtain a fusion classification model. The final fusion classification model will select the target classification model with higher accuracy. In practice, the target classification model with shorter acquisition time or lower cost will be given priority.

[0074] Exemplary device

[0075] like Figure 4 As shown in the figure, this embodiment of the invention provides an automatic content classification device for the Android Chinese application market. The device includes an application source data acquisition module 401, an extraction module 402, and a predicted category probability set output module 403, wherein: the application source data acquisition module 401 is used to acquire application source data; wherein, the application source data is used to characterize the attribute information of the application;

[0076] Extraction module 402 is used to extract features from the application source data;

[0077] The predicted category probability set output module 403 is used to input the features into the trained fusion classification model, output the predicted category probability set through the fusion classification model, and obtain the classification result based on the predicted category probability set.

[0078] Based on the above embodiments, the present invention also provides a smart terminal, the principle block diagram of which can be as follows: Figure 5As shown, the smart terminal includes a processor, memory, network interface, display screen, and temperature sensor connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements an automatic content rating method for the Android Chinese application market. The display screen can be an LCD screen or an e-ink screen. The temperature sensor is pre-installed within the smart terminal to detect the operating temperature of internal devices.

[0079] Those skilled in the art will understand that Figure 5 The schematic diagram in the figure is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the smart terminal to which the present invention is applied. A specific smart terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0080] In one embodiment, a smart terminal is provided, including a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors. The one or more programs include instructions for performing the following operations:

[0081] Obtain application source data; wherein the application source data is used to characterize the attribute information of the application;

[0082] Extract features from the application source data;

[0083] The features are input into a trained fusion classification model, which outputs a predicted category probability set. The classification result is obtained based on the predicted category probability set.

[0084] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0085] In summary, this invention discloses an automatic content classification method for the Android Chinese application market. The method includes: acquiring application source data; wherein the application source data is used to characterize the attribute information of the application; extracting features from the application source data; inputting the features into a trained fusion classification model; outputting a predicted category probability set through the fusion classification model; and obtaining a classification result based on the predicted category probability set. This invention extracts features from multi-dimensional application source data, resulting in rich feature information. By inputting these features into a trained fusion classification model, an accurate predicted category probability set is obtained, ultimately leading to a classification result, enabling content classification for the Android Chinese application market.

[0086] Based on the above embodiments, the present invention discloses an automatic content classification method for the Android Chinese application market. It should be understood that the application of the present invention is not limited to the above examples. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An automatic content rating method for Android Chinese application market, characterized in that, The method comprises: obtaining application source data; wherein the application source data is used to represent attribute information of an application, including application categories, advertising labels, application descriptions, application icons, application screenshots, image resources in application resource files, application permissions in application manifest files, and application program interfaces in application source codes in an Android Chinese application market; extracting features of the application source data; inputting the features into a trained fusion classification model, outputting a set of predicted category probabilities through the fusion classification model, and obtaining a classification result according to the set of predicted category probabilities, the set of predicted category probabilities including a probability that an application is suitable for children under 12 years old and a probability that an application is not suitable for children under 12 years old; the extracting features of the application source data comprises: when the application source data is text, inputting the text into a preset word vector model to output features corresponding to the text through the word vector model; when the application source data is an image, filtering the image, and performing image feature extraction on the filtered image to obtain features corresponding to the image; the generation process of the fusion classification model comprises: obtaining a plurality of training feature types; wherein each training feature type includes a plurality of training feature data; for each training feature type, obtaining a plurality of initial classification models, inputting the plurality of training feature data in each training feature type into each initial classification model, and obtaining a plurality of model accuracies corresponding to the training feature type through the initial classification model, the initial classification model including decision tree, logistic regression, support vector machine, K-nearest neighbor, random forest, and neural network; obtaining a target classification model corresponding to the training feature type according to a plurality of the model accuracies; based on a soft voting algorithm, accumulating a plurality of target classification models corresponding to the training feature types to obtain a fusion classification model.

2. The method for automatic content rating for Android Chinese application market of claim 1, wherein, the inputting the plurality of training feature data in each training feature type into each initial classification model and obtaining a plurality of model accuracies corresponding to the training feature type through the initial classification model comprises: obtaining true classification results of the plurality of training feature data in each training feature type; inputting the plurality of training feature data into each initial classification model to obtain predicted classification results corresponding to the plurality of training feature data in each training feature type; counting the number of true classification results of the plurality of training feature data in each training feature type and the predicted classification results corresponding to the plurality of training feature data in each training feature type that are the same to obtain the number of accurate prediction classifications of each training feature type after classification by each initial classification model; obtaining a plurality of model accuracies corresponding to the training feature type according to the number of accurate prediction classifications of each training feature type after classification by each initial classification model.

3. The method for automatic content rating for Android Chinese application market of claim 2, wherein, the obtaining a plurality of model accuracies corresponding to the training feature type according to the number of accurate prediction classifications of each training feature type after classification by each initial classification model comprises: According to the accurate prediction classification number of each training feature type after classification by each initial classification model, the accuracy of each training feature type after classification by each initial classification model is obtained. According to the accuracy of each training feature type after classification by each initial classification model, a plurality of model accuracies corresponding to the training feature type are obtained.

4. The method for automatic content rating for Android Chinese application market of claim 3, wherein, The accuracy of each training feature type after classification by each initial classification model is obtained according to the accurate prediction classification number of each training feature type after classification by each initial classification model, and includes: The total number of training feature data in each training feature type is obtained. The accurate prediction classification number of each training feature type after classification by each initial classification model is divided by the total number of training feature data in each training feature type to obtain the accuracy of each training feature type after classification by each initial classification model.

5. The method for automatic content rating for Android Chinese application market of claim 1, wherein, The target classification model corresponding to the training feature type is obtained according to the plurality of model accuracies, and includes: The maximum model accuracy is selected from the plurality of model accuracies. The initial classification model corresponding to the maximum model accuracy is taken as the target classification model corresponding to the training feature type.

6. A smart terminal, characterized by A computer program product, comprising a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors to perform the method of any one of claims 1-5.

7. A non-transitory computer-readable storage medium, comprising: When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device can perform the method of any one of claims 1-5.

Images