Application recommendation method and device, equipment, storage medium and computer program product

Abstract

The invention provides an application recommendation method and device, equipment, a storage medium and a computer program product, and the method comprises the steps: obtaining application preference information corresponding to a user according to an application situation factor and a user attribute feature through employing a CFGAN model based on a generative adversarial neural network; the application situation factor comprises at least one of application use time and application use position; and performing application recommendation according to the application preference information. According to the scheme, a CFGAN model of a deep learning model can be used for application recommendation, the expansibility of the scheme is improved, meanwhile, application preferences are generated according to application situation factors and user attribute features, application recommendation is carried out according to the application preferences, the user preferences can be mined as much as possible to a certain extent, and therefore the recommendation result is optimized; the problems that in the prior art, an application recommendation scheme is poor in recommendation result and low in scheme expandability are solved.

Description

Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to an application recommendation method, device, equipment, storage medium and computer program product. Background Art

[0002] In the prior art, some mobile application recommendation solutions have emerged in order to provide better services for users. However, since mobile applications are easy to develop and new mobile applications are constantly generated, these mobile application recommendation solutions have some defects, such as:

[0003] (1) It cannot fully explore user preferences, making model construction difficult, resulting in poor recommendation results;

[0004] (2) It is usually necessary to add new mobile applications to the solution matrix for model retraining, resulting in a lack of scalability of the solution.

[0005] As described above, the application recommendation scheme in the prior art has problems such as poor recommendation results and low scalability of the scheme. Summary of the invention

[0006] The purpose of the present application is to provide an application recommendation method, apparatus, device, storage medium and computer program product to solve the problems of poor recommendation results and low scalability of application recommendation solutions in the prior art.

[0007] In order to solve the above technical problems, the present application embodiment provides an application recommendation method, including:

[0008] The collaborative filtering recommendation framework CFGAN model based on generative adversarial neural network is used to obtain the application preference information corresponding to the user according to the application context factor and the user attribute characteristics; the application context factor includes at least one of the application usage time and the application usage location;

[0009] An application recommendation is performed according to the application preference information.

[0010] Optionally, the collaborative filtering recommendation framework CFGAN model based on generative adversarial neural network is used to obtain the application preference information corresponding to the user according to the application context factors and user attribute characteristics, including:

[0011] Encode application context factors and user attribute characteristics according to data types to obtain model input data; the data types include enumeration types and numerical variables;

[0012] The CFGAN model is used to obtain the application preference information corresponding to the user according to the model input data.

[0013] Optionally, also include:

[0014] The CFGAN model is trained by adopting a feature extraction capability enhancement operation;

[0015] The feature extraction capability enhancement operation includes: using an attention mechanism to enhance the initial mask vector in the initial model.

[0016] Optionally, the attention mechanism includes: at least one of a spatial attention SAM mechanism, a channel attention mechanism, and a permutation attention mechanism.

[0017] Optionally, the feature extraction capability enhancement operation is adopted to train the CFGAN model, including:

[0018] Inputting training data into the generator of the initial model to obtain predicted user application preferences; the training data includes: data on application context factors and user attribute characteristics;

[0019] The initial mask vector is enhanced by using an attention mechanism to obtain a target mask vector;

[0020] Obtaining a mask result according to the predicted user application preference and the target mask vector;

[0021] Inputting the real user application preference, the real reference data and the mask result into the discriminator of the initial model to obtain a discrimination result; the real reference data includes: data on application context factors and user attribute characteristics;

[0022] According to the discrimination result, the initial model is trained to obtain the CFGAN model.

[0023] Optionally, also include:

[0024] Using a factorization machine and a convolutional neural network, obtaining first preference reference information and second preference reference information of the user; the first preference reference information corresponds to a first level of user interaction features, and the second preference reference information corresponds to a second level of user interaction features;

[0025] The performing application recommendation according to the application preference information includes:

[0026] Application recommendation is performed according to the application preference information, the first preference reference information and the second preference reference information.

[0027] The present application also provides an application recommendation device, including:

[0028] The first processing module is used to obtain the application preference information corresponding to the user according to the application context factor and the user attribute characteristics by using the collaborative filtering recommendation framework CFGAN model based on the generative adversarial neural network; the application context factor includes at least one of the application use time and the application use location;

[0029] The first recommendation module is used to make application recommendations based on the application preference information.

[0030] Optionally, the collaborative filtering recommendation framework CFGAN model based on generative adversarial neural network is used to obtain the application preference information corresponding to the user according to the application context factors and user attribute characteristics, including:

[0031] Encode application context factors and user attribute characteristics according to data types to obtain model input data; the data types include enumeration types and numerical variables;

[0032] The CFGAN model is used to obtain the application preference information corresponding to the user according to the model input data.

[0033] Optionally, also include:

[0034] A first training module is used to train the CFGAN model by using a feature extraction capability enhancement operation;

[0035] The feature extraction capability enhancement operation includes: using an attention mechanism to enhance the initial mask vector in the initial model.

[0036] Optionally, the attention mechanism includes: at least one of a spatial attention SAM mechanism, a channel attention mechanism, and a permutation attention mechanism.

[0037] Optionally, the feature extraction capability enhancement operation is adopted to train the CFGAN model, including:

[0038] Inputting training data into the generator of the initial model to obtain predicted user application preferences; the training data includes: data on application context factors and user attribute characteristics;

[0039] The initial mask vector is enhanced by using an attention mechanism to obtain a target mask vector;

[0040] Obtaining a mask result according to the predicted user application preference and the target mask vector;

[0041] Inputting the real user application preference, the real reference data and the mask result into the discriminator of the initial model to obtain a discrimination result; the real reference data includes: data on application context factors and user attribute characteristics;

[0042] According to the discrimination result, the initial model is trained to obtain the CFGAN model.

[0043] Optionally, also include:

[0044] A first acquisition module is used to acquire first preference reference information and second preference reference information of a user by using a factorization machine and a convolutional neural network; the first preference reference information corresponds to a first-level user interaction feature, and the second preference reference information corresponds to a second-level user interaction feature;

[0045] The performing application recommendation according to the application preference information includes:

[0046] Application recommendation is performed according to the application preference information, the first preference reference information and the second preference reference information.

[0047] The embodiment of the present application also provides an application recommendation device, including: a processor;

[0048] The processor is used to obtain application preference information corresponding to the user according to application context factors and user attribute characteristics by using a collaborative filtering recommendation framework CFGAN model based on a generative adversarial neural network; the application context factors include: at least one of application usage time and application usage location;

[0049] An application recommendation is performed according to the application preference information.

[0050] Optionally, the collaborative filtering recommendation framework CFGAN model based on generative adversarial neural network is used to obtain the application preference information corresponding to the user according to the application context factors and user attribute characteristics, including:

[0051] Encode application context factors and user attribute characteristics according to data types to obtain model input data; the data types include enumeration types and numerical variables;

[0052] The CFGAN model is used to obtain the application preference information corresponding to the user according to the model input data.

[0053] Optionally, the processor is further configured to:

[0054] The CFGAN model is trained by adopting a feature extraction capability enhancement operation;

[0055] The feature extraction capability enhancement operation includes: using an attention mechanism to enhance the initial mask vector in the initial model.

[0056] Optionally, the attention mechanism includes: at least one of a spatial attention SAM mechanism, a channel attention mechanism, and a permutation attention mechanism.

[0057] Optionally, the feature extraction capability enhancement operation is adopted to train the CFGAN model, including:

[0058] Inputting training data into the generator of the initial model to obtain predicted user application preferences; the training data includes: data on application context factors and user attribute characteristics;

[0059] The initial mask vector is enhanced by using an attention mechanism to obtain a target mask vector;

[0060] Obtaining a mask result according to the predicted user application preference and the target mask vector;

[0061] Inputting the real user application preference, the real reference data and the mask result into the discriminator of the initial model to obtain a discrimination result; the real reference data includes: data on application context factors and user attribute characteristics;

[0062] According to the discrimination result, the initial model is trained to obtain the CFGAN model.

[0063] Optionally, the processor is further configured to:

[0064] Using a factorization machine and a convolutional neural network, obtaining first preference reference information and second preference reference information of the user; the first preference reference information corresponds to a first level of user interaction features, and the second preference reference information corresponds to a second level of user interaction features;

[0065] The performing application recommendation according to the application preference information includes:

[0066] Application recommendation is performed according to the application preference information, the first preference reference information and the second preference reference information.

[0067] An embodiment of the present application also provides an application recommendation device, including a memory, a processor, and a program stored in the memory and executable on the processor; when the processor executes the program, the above-mentioned application recommendation method is implemented.

[0068] The embodiment of the present application also provides a readable storage medium on which a program is stored. When the program is executed by a processor, the steps in the above-mentioned application recommendation method are implemented.

[0069] An embodiment of the present application also provides a computer program product, including computer instructions, which implement the steps of the above-mentioned application recommendation method when executed by a processor.

[0070] The beneficial effects of the above technical solution of the present application are as follows:

[0071] In the above scheme, the application recommendation method obtains the application preference information corresponding to the user according to the application context factors and user attribute characteristics by utilizing the collaborative filtering recommendation framework CFGAN model based on the generative adversarial neural network; the application context factors include: at least one of the application usage time and the application usage location; application recommendation is performed according to the application preference information; it can support the use of the CFGAN model of the deep learning model for application recommendation, thereby avoiding the problem of regenerating the application interaction matrix for training due to application iteration, thereby improving the scalability of the scheme. At the same time, considering the characteristics of mobile applications, this scheme generates application preferences according to the application context factors and user attribute characteristics and performs application recommendations accordingly, which can support the maximum mining of user preferences to a certain extent, thereby optimizing the recommendation results; and solves the problems of poor recommendation results and low scalability of the application recommendation scheme in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 A schematic diagram of a framework of a recommendation method based on MF (matrix decomposition) according to an embodiment of the present application;

[0073] Figure 2 A flowchart of a method for recommending the application of an embodiment of the present application;

[0074] Figure 3 A schematic diagram of a CFGAN framework based on contextual factors and user attributes according to an embodiment of the present application;

[0075] Figure 4 A schematic diagram of the model training process of an embodiment of the present application;

[0076] Figure 5 This is a schematic diagram of the SAM mechanism framework of an embodiment of the present application;

[0077] Figure 6 A schematic diagram of the structure of a recommended device for application of an embodiment of the present application;

[0078] Figure 7 This is a schematic diagram of the recommended device structure for the application of the embodiment of the present application. DETAILED DESCRIPTION

[0079] In order to make the technical problems, technical solutions and advantages to be solved by the present application clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0080] The following first introduces the relevant contents of this plan.

[0081] In recent years, the mobile Internet has developed rapidly, and mobile devices (such as smart phones, tablet computers, etc.) have become mainstream network terminals. Among them, mobile applications deployed on mobile devices have brought great convenience to people's production and life. However, although a large number of mobile applications meet people's growing daily life and work needs, due to the wide variety and large number of mobile applications, it is also difficult for users to conveniently choose mobile applications that truly meet their preferences; therefore, it is necessary to help users accurately and quickly choose mobile applications that meet their preferences. At present, the mobile application recommendation system uses the user's historical data to mine the user's preferences, and then recommends mobile applications that meet their preferences to the user; therefore, the implementation of personalized mobile application recommendations can meet the user's personalized needs for accurate selection of mobile applications. In this way, personalized mobile application recommendations for users not only help users get a better experience in the mobile Internet ecosystem, but also help mobile application providers promote their services and products.

[0082] Specifically, similar to item recommendations, current mobile application recommendation models are mainly divided into two categories from the perspective of recommendation methods, as follows:

[0083] Solution 1: Recommendation based on CF (collaborative filtering);

[0084] The problem of mobile application recommendation is similar to the problem of traditional item recommendation, so the ideas of recommendation methods in other fields can also be transplanted to mobile application recommendation. The most widely used method in the recommendation system method is the CF-based recommendation method. The CF-based recommendation method is widely used in many fields including mobile application recommendation because of its simple logic and easy implementation. The CF-based recommendation method is a similarity-oriented recommendation method based on the assumption that the target user has similar preferences with users with similar historical item experience. Therefore, the most important part of the CF-based recommendation method is to calculate the similarity between users or items. This method needs to record and calculate the similarity of users or items, so it is also called a memory-based recommendation method. In practical applications, the CF-based recommendation method requires explicit feedback information from users on items; if there is no explicit user feedback, the implicit user feedback information should be converted into explicit feedback information according to certain rules; the feedback information is used as the user's preference vector, and the user similarity model is obtained by calculating the similarity of preference vectors between different users.

[0085] Solution 2: Model-based recommendation;

[0086] Among the model-based recommendation methods, the most representative one is the recommendation method based on MF (matrix decomposition). The basic idea of ​​this method is as follows: First, an original sparse user-mobile application interaction matrix is ​​constructed based on the MF method, in which the elements are the feedback information of specific users on specific mobile applications, such as ratings, usage frequency, etc. Then, the sparse user-mobile application interaction matrix is ​​processed into a non-sparse matrix using matrix decomposition technology. The elements in this matrix are regarded as the user's preferences for mobile applications. Finally, a recommendation list is generated for the user based on the preference. Specifically, the framework of the recommendation method based on MF can be as follows: Figure 1 As shown (the two types of filling in the figure are just to illustrate different items), Figure 1 The interactive rating matrix X of the original user and mobile application on the far left has n users and m mobile applications (i.e., the items in the figure), and the elements in the matrix are the user's ratings on the mobile application; by decomposing the matrix, it is decomposed into two low-dimensional matrices on the right, which represent the user's feature vector U and the mobile application's feature vector V respectively (k in the figure represents the feature vector dimension); then the two low-dimensional matrices are multiplied to obtain a dense matrix with the same dimension as the original matrix and similar information. The elements in this similarity matrix (i.e., dense matrix) are the user's preference values ​​for mobile applications. The MF-based recommendation method recommends mobile applications to users based on the element values ​​in the decomposed matrix, which has a great advantage in combating data sparsity problems, and has therefore been widely studied and applied in the field of mobile application recommendation.

[0087] However, the above two solutions have their own defects, which are introduced as follows:

[0088] (1) The core idea of ​​CF-based mobile application recommendation is to establish a similarity model for users or mobile applications, and recommend mobile applications to users based on the similarity model. The recommendation results mainly depend on the construction of the similarity model. In this regard, the CF-based mobile application recommendation method has the following challenges: Faced with the extremely sparse interaction data between users and mobile applications in the real world, the CF-based mobile application recommendation method cannot fully explore the user's preferences. The elements in the two-dimensional interaction matrix between users and mobile applications in the real world represent the user's rating or preference for the mobile application. If the user has rated the mobile application, the element is the rating value, and if there is no rating, the element is set to 0. However, in the real world, users often only come into contact with a small number of mobile applications, resulting in most of the elements in the above interaction matrix being 0, that is, this two-dimensional matrix is ​​extremely sparse. Faced with such sparse user mobile application data, the CF-based recommendation method cannot easily find similar vectors of the target user or mobile application feature vector, making it difficult to build a similarity model, resulting in poor recommendation results.

[0089] (2) One of the main goals of MF-based recommendation methods is to combat data sparsity, and thus they have achieved better results than collaborative filtering-based recommendation methods in some areas. However, since the low-dimensional vectors generated by this method during the matrix decomposition process have no clear physical meaning, the interpretability of this method is poor, so there are still some challenges in its application in the field of mobile application recommendation. In addition, because the results of matrix decomposition are greatly affected by the initial user-item interaction matrix, the recommendation results may be non-personalized for a single target user. In addition, since mobile applications are easy to develop and new mobile applications are constantly generated, it is usually necessary to add new mobile applications to the user-mobile application interaction matrix for retraining, which also leads to the lack of scalability of recommendation methods based on matrix decomposition or latent semantic models.

[0090] Furthermore, the recommendation of mobile applications is different from the recommendation of other items in that: since mobile applications are deployed on mobile devices and can be used anytime, anywhere, the contextual information such as the time and location of the user's interaction with the mobile application has a great impact on the recommendation results. In addition, mobile applications are also functionally exclusive, which leads to challenges different from those faced by recommendations in traditional fields. Furthermore, the development of mobile applications is easy, has a short cycle, and has continuously updated versions, making the interaction data between users and mobile applications more sparse compared to the massive number of mobile applications, which brings great difficulties to accurate recommendations.

[0091] Based on the above, this application provides an application recommendation method to solve the problems of poor recommendation results and low scalability of application recommendation solutions in existing technologies, such as Figure 2 As shown, including:

[0092] Step 21: using the collaborative filtering recommendation framework CFGAN model based on generative adversarial neural network, according to the application context factors and user attribute characteristics, obtaining the application preference information corresponding to the user; the application context factors include: at least one of the application usage time and the application usage location;

[0093] Step 22: Recommend applications based on the application preference information.

[0094] Among them, the application preference information may include: at least one of the information such as the preference probability for each application, the interaction probability with each application, etc.; and / or the user attribute characteristics include: at least one of gender, age, occupation, income and interests, but are not limited to this.

[0095] Furthermore, in this solution, feature vectors of other applications (such as relevant information vectors of time dimension, space dimension, etc.) can also be considered to obtain the application preference information, that is, step 21 may include: obtaining the application preference information corresponding to the user according to the application context factor, user attribute characteristics and feature vectors of other applications; the other application may also be referred to as the first application, etc., and the other application may include: at least one of the applications that the user has interacted with, downloaded, etc., which is not limited here. This can support application recommendation by using the complementarity between applications. Complementarity refers to the phenomenon of using multiple application functions together to meet user needs, and the feature vector of the application can feedback a certain complementarity.

[0096] The application recommendation method provided in the embodiment of the present application obtains the application preference information corresponding to the user according to the application context factors and user attribute characteristics by utilizing the collaborative filtering recommendation framework CFGAN model based on the generative adversarial neural network; the application context factors include: at least one of the application usage time and the application usage location; application recommendation is performed according to the application preference information; it can support the use of the CFGAN model of the deep learning model for application recommendation, thereby avoiding the problem of regenerating the application interaction matrix for training due to application iteration, thereby improving the scalability of the solution. At the same time, considering the characteristics of mobile applications, this solution generates application preferences according to the application context factors and user attribute characteristics and performs application recommendations accordingly, which can support the maximum mining of user preferences to a certain extent, thereby optimizing the recommendation results; and solves the problems of poor recommendation results and low scalability of the application recommendation solution in the prior art.

[0097] The method of using the collaborative filtering recommendation framework CFGAN model based on generative adversarial neural network to obtain the application preference information corresponding to the user according to the application context factors and user attribute characteristics includes: encoding the application context factors and user attribute characteristics according to the data type to obtain model input data; the data type includes enumeration type and numerical variable; using the CFGAN model to obtain the application preference information corresponding to the user according to the model input data. In this way, the above application preference information can be accurately obtained.

[0098] Furthermore, the application recommendation method further includes: adopting a feature extraction capability enhancement operation to train the CFGAN model; wherein the feature extraction capability enhancement operation includes: adopting an attention mechanism to enhance the initial mask vector in the initial model. This can enhance the model's recommendation capability in the face of sparse data set feature scenarios, achieve accurate and personalized recommendation business output, and avoid the reduction of recommendation effect due to data sparsity.

[0099] The attention mechanism includes at least one of a spatial attention SAM mechanism, a channel attention mechanism, and a permutation attention mechanism. In this way, the attention mechanism can be clarified.

[0100] In the embodiment of the present application, the feature extraction capability enhancement operation is adopted to train the CFGAN model, including: inputting the training data into the generator of the initial model to obtain the predicted user application preference (such as the predicted user interaction vector); the training data includes: data for application context factors and user attribute features; using the attention mechanism to enhance the initial mask vector to obtain the target mask vector; according to the predicted user application preference and the target mask vector, the mask result is obtained; the real user application preference (such as the real user interaction vector), the real reference data and the mask result are input into the discriminator of the initial model to obtain the discrimination result; the real reference data includes: data for application context factors and user attribute features; according to the discrimination result, the initial model is trained (for example, if the discrimination result indicates true, the training is terminated, and if the discrimination result indicates false, the generator is instructed to continue training) to obtain the CFGAN model. In this way, the CFGAN model can be accurately obtained. Among them, "using the attention mechanism to enhance the initial mask vector to obtain the target mask vector" may include: using the attention mechanism (such as the SAM mechanism) to perform feature focus on the interaction vector (enhancing the initial mask corresponding to the predicted user application preference), specifically, for example: processing the initial mask vector to generate a spatial attention feature, determining the application corresponding to the top Z features according to the spatial attention feature, updating the initial mask vector according to the determined application, and obtaining the target mask vector, but not limited to this. Among them, in the case of instructing the generator to continue training, at least one of the following information may be fed back to the generator: training loss value, training instruction (which can be used to instruct to continue training), but not limited to this.

[0101] Furthermore, the application recommendation method further includes: using a factor decomposition machine and a convolutional neural network to obtain the user's first preference reference information and second preference reference information (for example, using a factor decomposition machine to obtain the user's first preference reference information; using a convolutional neural network to obtain the user's second preference reference information); the first preference reference information corresponds to the first level of user interaction features, and the second preference reference information corresponds to the second level of user interaction features; the application recommendation based on the application preference information includes: based on the application preference information, the first preference reference information and the second preference reference information, the application recommendation. This can further improve the application recommendation effect. Among them, "using a factor decomposition machine and a convolutional neural network to obtain the user's first preference reference information and the second preference reference information", for example: by using a factor decomposition machine and a convolutional neural network, respectively learning the low-order and high-order interaction features between the user and the application (for example, using a factor decomposition machine to learn the low-order interaction features between the user and the application; using a convolutional neural network to learn the high-order interaction features between the user and the application), multi-order learning is performed to effectively mine the user's mobile application usage preferences. Among them, "making application recommendations based on the application preference information, the first preference reference information and the second preference reference information" may include: making application recommendations based on the application preference information, the first preference reference information, the second preference reference information and the weights (which may be preset values ​​or determined based on input information) corresponding to (each information, i.e., these three information), but is not limited to this.

[0102] The application recommendation method provided in the embodiment of the present application is illustrated below with examples, and the attention mechanism takes the spatial attention SAM mechanism as an example.

[0103] In view of the above technical problems, the embodiment of the present application provides an application recommendation method, which can be specifically implemented as an application recommendation method based on contextual factors and generative adversarial networks to realize a personalized recommendation method for mobile applications; it involves: using the CFGAN model of the deep learning model for application recommendation, which can avoid the dilemma of the MF-based recommendation algorithm needing to recalculate the matrix when the application is rapidly iterated; at the same time, taking into account the characteristics of mobile applications, on the basis of collecting and sorting out the interaction data between users and mobile applications, the mobile application contextual factors (corresponding to the above application contextual factors) and the user attribute angle (corresponding to the above user attribute characteristics) are taken into consideration to generate a mobile application recommendation list for users (corresponding to the above application recommendation); on the other hand, in order to strengthen the recommendation ability of this solution in the face of sparse data set feature scenarios and realize accurate personalized recommendation business output, the SAM (spatial attention) mechanism can be integrated with CFGAN. The following is a specific example of the content involved in this solution.

[0104] 1. Introduce situational factors;

[0105] In this solution, the combination of time and geographic location can be called a contextual factor. In real life, mobile applications often show a phenomenon of spatiotemporal clustering. For example, users use video mobile applications to watch videos at home during non-working hours, but rarely use video mobile applications during working hours. This phenomenon occurs because the function of video mobile applications is mainly entertainment and is suitable for leisure environments. This phenomenon reflects the strong correlation between mobile application functions and contextual factors, that is, mobile applications with specific functions are suitable for situations that meet the function. Therefore, considering this spatiotemporal clustering of mobile applications, the contextual factors of mobile applications can be used as a basis for recommendation.

[0106] Unlike items in the traditional recommendation field, mobile applications also have the feature of functional exclusivity; that is, if a user downloads a mobile application with specific functions to meet his specific needs under specific situational conditions, then the user may not be interested in other mobile applications with similar functions. The functional exclusivity of mobile applications makes the similarity calculation method of items based on the CF method invalid for mobile applications. For example, if a user downloads a mobile application with a weather forecast function, according to the idea based on the CF method, the recommendation system will predict that mobile applications with similar weather forecast functions also meet the user's preferences; however, it is obvious that by downloading the mobile application, the user's demand for the weather forecast function has been met; this user is no longer interested in other mobile applications with weather forecast functions. Ignoring the exclusivity of mobile application functions in the process of building a mobile application recommendation model is the reason why the recommendation effect is difficult to further improve.

[0107] However, it is difficult to construct a model directly from the perspective of exclusivity. Therefore, this solution takes into account the particularity of mobile applications and uses complementarity for recommendation. Complementarity refers to the phenomenon of using multiple mobile application functions together to meet user needs. For example, users like to use camera mobile applications to take photos at tourist attractions, and use photo editing mobile applications to beautify photos, and then share photos on social mobile applications. After introducing contextual factors into the model, complementarity is usually manifested as the phenomenon of overlapping clustering of contextual factors between mobile applications with different functions, reflecting the functional complementarity of mobile applications under contextual conditions. In real life, the clustering of contextual information of different applications is very common.

[0108] 2. CFGAN framework based on contextual factors and user attributes;

[0109] In recent years, deep learning technology has achieved great success in image recognition, natural language processing, speech recognition and other fields because of its powerful nonlinear expression ability and ability to automatically learn the potential relationships in features. Among them, deep learning technology emphasizes learning from massive data, solving the problem that traditional machine learning algorithms are difficult to handle with high-dimensional, heterogeneous, and noisy data. Due to the excellent nonlinear expression ability of deep learning technology, deep learning technology can be applied in recommendation systems; however, the loss function design of deep learning models is quite difficult for researchers who lack strong domain knowledge. Therefore, the GAN (generative adversarial network) model can be introduced to replace the design of complex loss functions and place the problem in a minimal-maximum two-person zero-sum game framework. Among them, CFGAN, as a novel GAN-based CF recommendation technology, has better recommendation accuracy and is also optimized for collaborative filtering. Based on this, the embodiment of the present application proposes a CFGAN based on contextual factors and user attributes, which can be used as Figure 3 As shown in the figure, the CFGAN model based on contextual factors and user attributes proposed in this scheme mainly involves the following two aspects:

[0110] (1) Considering the impact of application characteristics on recommendation results, the contextual factors and user attributes are encoded and used as the input of the CFGAN model for recommendation, thereby improving the accuracy of recommendation.

[0111] (2) Considering the sparsity of user-application interaction data in actual application scenarios, the SAM mechanism is introduced to enhance its feature focusing capability (corresponding to enhancing Figure 3 Zhonge u ).

[0112] In this application, the training process of the entire model may involve the following steps: Figure 4 The content shown includes: the generator generates an interaction vector, and uses SAM to optimize the hidden vector (which may correspond to the following according to the mask vector e u Get vector e u '), and thus the interaction vector is hidden (corresponding to the following and e u ′ performs mask operation); specifically, the training process of the entire model may include the following four parts:

[0113] Part 1. Information extraction and processing;

[0114] This section mainly extracts two types of information from mobile applications and users: contextual factors and user attributes (features).

[0115] Among them, the contextual factors of the application include the time and location information of using the application, which can be transmitted back by the mobile device, and may include the time of use, the longitude and latitude of the use location, etc. (which may correspond to the above-mentioned application contextual factors including: at least one of the application use time and the application use location); user attribute characteristics may include: gender, age, occupation, income, interests, etc., and then the two extracted information can be divided into enumeration type and numerical variable according to the data type for encoding; corresponding to the above-mentioned application context factors and user attribute characteristics, they can be encoded separately according to the data type to obtain model input data; the data type includes enumeration type and numerical variable.

[0116] For enumeration type variables (such as user gender, location type, or user occupation), one-hot encoding (unique hot encoding or dummy encoding) can be used for encoding, and multiple features will be generated after mapping. One-hot encoding, also known as one-bit effective encoding, mainly uses N-bit state registers to encode N states. Each state has an independent register bit, and only one bit is valid at any time.

[0117] For numerical types, the data can be standardized by subtracting the mean and then dividing by the variance (or standard deviation), so that the processed data conforms to the standard normal distribution; among them, numerical attributes such as age and income can be characterized by this method.

[0118] Finally, all processed attribute features of various types can be concatenated as the user's attribute feature vector. Each item of the feature vector has a clear meaning and there is no ambiguity.

[0119] Part 2. The generator generates an interaction vector;

[0120] Different from the original GAN ​​model that can generate different results according to different random noise variables, the CFGAN proposed in this scheme based on application context factors and user attributes aims to generate the most reasonable recommendation result for users instead of multiple results, so random noise variables are not used.

[0121] Among them, the feature vector (context factor and user attribute vector) obtained in part one can be input into the generator G composed of a multi-layer perceptron to generate a forged (or predicted) interaction vector (which can correspond to the above-mentioned use of the CFGAN model, according to the model input data, to obtain the user's corresponding application preference information), which can be as follows Figure 3 As shown, the generator G can obtain users u1···u m The corresponding user interaction vectors i1, i2, ···i n; The circles of different sizes are used to represent the interaction probability of users with the corresponding applications. The larger the circle, the greater the interaction probability. When the probability of user-item interaction reaches the maximum value of 1, it is represented by a solid circle. For the real user interaction vector, a solid circle indicates interaction behavior, and a dotted hollow circle indicates no interaction behavior. uj is the interaction vector of the jth (1≤j≤m) user, i k is the probability that the user interacts with the kth item (1≤k≤n, n is the size of the item set). Masking the generated interaction vector is the core part of CFGAN. GAN aims to make the generator generate interaction vectors as similar as possible to the real user interaction vector. Only the output of G on the interaction term can help the generator G and the discriminator D learn. Therefore, the mask can be used to generate the interaction vector. A user interaction vector will be generated (corresponding to i k ) and the real user interaction vector r u The terms corresponding to the uninteracted terms are hidden, where e u is an n-dimensional indicator vector that specifies whether the user interacts with the application. Represents element-wise product.

[0122] Among them, for the mask vector e u ={s1,s2,...,s j},s j ={q0,...,q n},q n ∈(0,1), if there is an interaction record between the user and the item, the corresponding position of the mask vector is 1, otherwise it is set to 0. The mask operation is to and e u By doing the inner product, the CFGAN model can discard the items that the user has not interacted with during training, and focus more on the items that the user has interacted with.

[0123] The above involves generating user interaction vectors It can correspond to the above predicted user application preference, the real user interaction vector r u May correspond to the above-mentioned real user application preferences.

[0124] Part 3. Introducing SAM mechanism (SAM optimization);

[0125] Considering that the absence of interaction does not mean that the user is not interested in the item, there may be other situations, such as the user has not yet observed and discovered the item; and using the model to recommend mobile applications requires the user's interaction vector as an additional prior condition, but the user interaction vector is usually very sparse. The lack of interaction and semantic ambiguity will cause the performance of the discriminator to degrade, which in turn affects the accuracy of the generator and recommendations. It also does not consider the situation where the user is interested in the application but has no interaction, and lacks attention to the applications that the user may be interested in. At the same time, there are fewer applications with user interaction in sparse scenarios, and the model's ability to extract features from them is weak. In this regard, considering that the sparsity of user-item interaction data is relatively high in most application scenarios, in order to further improve the model's ability to focus on features of sparse data, this solution can use the following Figure 5 The SAM mechanism shown (corresponding to the spatial attention module in the figure) optimizes the model, where:

[0126] This mechanism can be used to e the (initial) mask vector in the model. u to vector e′ u During the training process, the constraints on the applications that users may be interested in are relaxed, and the gradient of this part is passed back to the generator. The enhancement of the initial mask vector may include: first, the input feature e u (correspond Figure 5 The feature F' in the channel is subjected to global maximum pooling based on the channel. max and the global average pooling operation f avg Get f max (e u ) and f avg (e u ), and then the results of the two pooling operations f max (e u ) and f avg (e u ) is spliced ​​based on the channel dimension, and the convolution operation is performed through the convolution layer to reduce the dimension, and then the spatial attention Ms feature is generated through the activation function. Based on this, the formula is obtained based on the generated spatial attention Ms feature and the input feature: spatial attention Ms feature = f SAM (e u )=σ(Conv(concat(f max (e u ),f avg (e u )))). For the feature vector after spatial enhancement, select the applications corresponding to the first Z features of each user after feature aggregation (that is, select the applications corresponding to the features ranked close to the first Z, which may include applications that have actually interacted and applications that have not been interacted with) as the applications that the user may interact with, and set e u Set the corresponding position in to 1, and get e′ u(which may correspond to the above-mentioned target mask vector), so that the gradient of the corresponding part can be returned and participate in the model training; in this scheme, the SAM optimization operation can be performed in each training, but it is not limited to this.

[0127] This part three may correspond to the above-mentioned feature extraction capability enhancement operation to train the CFGAN model; wherein the feature extraction capability enhancement operation includes: using an attention mechanism to enhance the initial mask vector in the initial model; the attention mechanism includes: a spatial attention SAM mechanism.

[0128] Part 4. Discriminator judges learning (the discriminator determines whether to iterate);

[0129] In the discriminator D, the result of the mask operation (which may correspond to the above mask result), the real user interaction vector, the context factor and the user attribute vector participate in the training together, make a true or false judgment and return the gradient (indicate whether to iterate) for further learning of the generator; for example, if the discriminator determines that it is "false", it determines to continue iterating and instructs the generator to continue training and generating; if the discriminator determines that it is "true", it determines to end. Among them, D() represents the probability that the input comes from the real user interaction vector. The discriminator hopes to distinguish the real data from the data generated by the generator as much as possible, that is: make D(r u |c u ) as large as possible, As small as possible; the generator G hopes that the generated interaction vector is close to the real interaction vector, that is, As large as possible. For the sake of model simplicity, both objective functions can be transformed into minimization problems. For this, the objective function J of the discriminator D is D It can be expressed as:

[0130]

[0131] The objective function of the generator J G It is expressed as: Model by J G and J D Perform adversarial learning, where c u Represents situational information.

[0132] The content of the model training involved above may correspond to the above: inputting the training data into the generator of the initial model to obtain the predicted user application preference; the training data includes: data for application context factors and user attribute characteristics; using the attention mechanism to enhance the initial mask vector to obtain the target mask vector; obtaining the mask result based on the predicted user application preference and the target mask vector; inputting the real user application preference, the real reference data and the mask result into the discriminator of the initial model to obtain the discrimination result; the real reference data includes: data for application context factors and user attribute characteristics; according to the discrimination result, training the initial model to obtain the CFGAN model.

[0133] Afterwards, when the adversarial training is completed, the user's preference scores for all applications will be generated, and the top-K can be returned as the result (i.e., recommended); corresponding to the above, application recommendations can be made based on the application preference information.

[0134] Based on the above, this plan mainly involves:

[0135] 1. Taking contextual factors (such as the time and space information of mobile applications) as the entry point, we introduce the exclusivity, complementarity and clustering of contextual factors of mobile applications, and provide a basis for the establishment of a recommendation model. Based on the collection and sorting of user-mobile application interaction data, we take into account the contextual factors of mobile applications and user attributes to generate a mobile application recommendation list for users. This can generate a user preference distribution that is as close to the real one as possible, and improve the accuracy of recommendations.

[0136] 2. The development of mobile applications is easy, has a short cycle, and has continuous version updates, which makes the interaction data between users and mobile applications more sparse compared to massive mobile applications, bringing great difficulties to accurate recommendations. In this regard, the SAM mechanism is integrated into the recommendation model to enhance the feature extraction capability for sparse data scenarios (that is, considering the sparsity of user-application interaction data in actual application scenarios, the SAM mechanism is introduced to enhance its feature focusing capability), thereby enhancing its recommendation capability in sparse data set feature scenarios and achieving accurate personalized recommendation business output.

[0137] In summary, the present application provides an application recommendation method based on contextual factors and generative adversarial networks: taking mobile application contextual factors and user attributes into consideration, integrating the SAM mechanism with CFGAN for application recommendation; specifically, it may include: extracting two types of information, contextual factors and user attributes, from both mobile applications and users, wherein contextual factors may include: time and location information of application use; user attribute characteristics may include: gender, age, occupation, income, interests, etc.; then encoding these two types of information according to data types, divided into enumeration types and numerical variables; after encoding the extracted contextual factors and user attributes, input them into a generator G composed of a multi-layer perceptron to generate an interaction vector and use the SAM mechanism to perform feature focus on the interaction vector (using the SAM mechanism to optimize the mask corresponding to the interaction vector); in the discriminator D, the results of the mask operation, the real user interaction vector, the contextual factors and the user attribute vector are jointly trained to perform true and false discrimination and return the gradient for further learning of the generator; finally, after the training is completed, the user's preference scores for all applications can be generated, and the top-K can be taken for application recommendation.

[0138] Based on the above, this solution has the following advantages:

[0139] 1. Compared with the CF-based mobile application recommendation method, this scheme proposes to integrate the SAM mechanism into the recommendation model to enhance the feature extraction capability for sparse data scenarios and avoid the reduction of recommendation effect due to data sparsity.

[0140] 2. Compared with the MF-based recommendation method, this solution uses deep neural networks for training, which avoids the problem of regenerating the application interaction matrix for training due to application iteration and improves scalability.

[0141] 3. Taking into account the characteristics of mobile applications, on the basis of collecting and sorting out the interaction data between users and mobile applications, mobile application contextual factors and user attributes can be taken into consideration to generate a mobile application recommendation list for users.

[0142] Furthermore, this solution can be expanded as follows:

[0143] 1. Neural networks can be combined with other attention mechanisms, such as channel attention mechanism, permutation attention mechanism, etc., which may further improve the recommendation results.

[0144] 2. In addition to introducing the attention mechanism, the low-order and high-order interaction features between users and mobile applications can be learned respectively by using factor decomposition machines and convolutional neural networks, and multi-order learning can be performed to effectively explore users' mobile application usage preferences; corresponding to the above-mentioned use of factor decomposition machines and convolutional neural networks, the user's first preference reference information and second preference reference information can be obtained; the first preference reference information corresponds to the first level of user interaction features, and the second preference reference information corresponds to the second level of user interaction features; application recommendations are made based on the application preference information, the first preference reference information, and the second preference reference information.

[0145] It is hereby explained that the embodiments of the present application have a broad market application prospect in the field of application recommendation, and provide an effective solution for the industry on how to provide personalized application recommendations to users.

[0146] The present application also provides an application recommendation device, such as Figure 6 As shown, including:

[0147] The first processing module 61 is used to obtain the application preference information corresponding to the user according to the application context factor and the user attribute characteristics by using the collaborative filtering recommendation framework CFGAN model based on the generative adversarial neural network; the application context factor includes at least one of the application use time and the application use location;

[0148] The first recommendation module 62 is used to make application recommendations based on the application preference information.

[0149] The application recommendation device provided in the embodiment of the present application obtains the application preference information corresponding to the user according to the application context factors and user attribute characteristics by utilizing the collaborative filtering recommendation framework CFGAN model based on the generative adversarial neural network; the application context factors include: at least one of the application usage time and the application usage location; application recommendation is performed according to the application preference information; it can support the use of the CFGAN model of the deep learning model for application recommendation, thereby avoiding the problem of regenerating the application interaction matrix for training due to application iteration, thereby improving the scalability of the solution. At the same time, considering the characteristics of mobile applications, this solution generates application preferences according to the application context factors and user attribute characteristics and performs application recommendations accordingly, which can support the maximum mining of user preferences to a certain extent, thereby optimizing the recommendation results; and solves the problems of poor recommendation results and low scalability of the application recommendation solution in the prior art.

[0150] Among them, the collaborative filtering recommendation framework CFGAN model based on generative adversarial neural network is used to obtain the application preference information corresponding to the user according to the application context factors and user attribute characteristics, including: encoding the application context factors and user attribute characteristics according to the data type to obtain model input data; the data type includes enumeration type and numerical variable; using the CFGAN model, according to the model input data, to obtain the application preference information corresponding to the user.

[0151] Furthermore, the application recommendation device also includes: a first training module, which is used to adopt a feature extraction capability enhancement operation to train the CFGAN model; wherein the feature extraction capability enhancement operation includes: adopting an attention mechanism to enhance the initial mask vector in the initial model.

[0152] Among them, the attention mechanism includes: at least one of: spatial attention SAM mechanism, channel attention mechanism and permutation attention mechanism.

[0153] In an embodiment of the present application, the feature extraction capability enhancement operation is used to train the CFGAN model, including: inputting training data into the generator of the initial model to obtain predicted user application preferences; the training data includes: data for application context factors and user attribute features; using an attention mechanism to enhance the initial mask vector to obtain a target mask vector; obtaining a mask result based on the predicted user application preferences and the target mask vector; inputting real user application preferences, real reference data and the mask result into the discriminator of the initial model to obtain a discrimination result; the real reference data includes: data for application context factors and user attribute features; based on the discrimination result, training the initial model to obtain the CFGAN model.

[0154] Furthermore, the application recommendation device also includes: a first acquisition module, used to obtain the user's first preference reference information and second preference reference information by using a factor decomposition machine and a convolutional neural network; the first preference reference information corresponds to the first level of user interaction features, and the second preference reference information corresponds to the second level of user interaction features; the application recommendation based on the application preference information includes: application recommendation based on the application preference information, the first preference reference information and the second preference reference information.

[0155] Among them, the implementation embodiments of the above-mentioned application recommendation method are all applicable to the embodiments of the application recommendation device and can achieve the same technical effects.

[0156] The present application also provides an application recommendation device, such as Figure 7 As shown, it includes: a processor 71;

[0157] The processor 71 is used to obtain application preference information corresponding to the user according to application context factors and user attribute characteristics by using a collaborative filtering recommendation framework CFGAN model based on a generative adversarial neural network; the application context factors include: at least one of application usage time and application usage location;

[0158] An application recommendation is performed according to the application preference information.

[0159] The application recommendation device provided in the embodiment of the present application obtains the application preference information corresponding to the user according to the application context factors and user attribute characteristics by utilizing the collaborative filtering recommendation framework CFGAN model based on the generative adversarial neural network; the application context factors include: at least one of the application usage time and the application usage location; application recommendation is performed according to the application preference information; it can support the use of the CFGAN model of the deep learning model for application recommendation, thereby avoiding the problem of regenerating the application interaction matrix for training due to application iteration, thereby improving the scalability of the solution. At the same time, considering the characteristics of mobile applications, this solution generates application preferences according to the application context factors and user attribute characteristics and performs application recommendations accordingly, which can support the maximum mining of user preferences to a certain extent, thereby optimizing the recommendation results; and solves the problems of poor recommendation results and low scalability of the application recommendation solutions in the prior art.

[0160] Among them, the collaborative filtering recommendation framework CFGAN model based on generative adversarial neural network is used to obtain the application preference information corresponding to the user according to the application context factors and user attribute characteristics, including: encoding the application context factors and user attribute characteristics according to the data type to obtain model input data; the data type includes enumeration type and numerical variable; using the CFGAN model, according to the model input data, to obtain the application preference information corresponding to the user.

[0161] Furthermore, the processor is also used to: adopt a feature extraction capability enhancement operation to train the CFGAN model; wherein the feature extraction capability enhancement operation includes: adopting an attention mechanism to enhance the initial mask vector in the initial model.

[0162] Among them, the attention mechanism includes: at least one of: spatial attention SAM mechanism, channel attention mechanism and permutation attention mechanism.

[0163] In an embodiment of the present application, the feature extraction capability enhancement operation is used to train the CFGAN model, including: inputting training data into the generator of the initial model to obtain predicted user application preferences; the training data includes: data for application context factors and user attribute features; using an attention mechanism to enhance the initial mask vector to obtain a target mask vector; obtaining a mask result based on the predicted user application preferences and the target mask vector; inputting real user application preferences, real reference data and the mask result into the discriminator of the initial model to obtain a discrimination result; the real reference data includes: data for application context factors and user attribute features; based on the discrimination result, training the initial model to obtain the CFGAN model.

[0164] Furthermore, the processor is also used to: use a factor decomposition machine and a convolutional neural network to obtain a user's first preference reference information and a second preference reference information; the first preference reference information corresponds to a first-level user interaction feature, and the second preference reference information corresponds to a second-level user interaction feature; and the application recommendation based on the application preference information includes: application recommendation based on the application preference information, the first preference reference information and the second preference reference information.

[0165] Among them, the implementation embodiments of the above-mentioned application recommendation method are all applicable to the embodiments of the application recommendation device and can achieve the same technical effects.

[0166] An embodiment of the present application also provides an application recommendation device, including a memory, a processor, and a program stored in the memory and executable on the processor; when the processor executes the program, the above-mentioned application recommendation method is implemented.

[0167] Among them, the implementation embodiments of the above-mentioned application recommendation method are all applicable to the embodiments of the application recommendation device and can achieve the same technical effects.

[0168] The embodiment of the present application also provides a readable storage medium on which a program is stored. When the program is executed by a processor, the steps in the above-mentioned application recommendation method are implemented.

[0169] Among them, the implementation embodiments of the above-mentioned application recommendation method are all applicable to the embodiments of the readable storage medium and can achieve the same technical effect.

[0170] The embodiment of the present application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, the various processes of the method embodiment of the above-mentioned application recommendation method are implemented, and the same technical effect can be achieved. To avoid repetition, they are not repeated here.

[0171] It should be noted that many functional components described in this specification are referred to as modules in order to more particularly emphasize the independence of their implementation methods.

[0172] In the embodiment of the present application, module can be implemented with software so that it can be executed by various types of processors. For example, an executable code module of an identification can include one or more physical or logical blocks of computer instructions, for example, it can be constructed as an object, process or function. Nevertheless, the executable code of the identified module does not need to be physically located together, but can include different instructions stored in different positions, and when these instructions are logically combined together, it constitutes a module and realizes the specified purpose of the module.

[0173] In fact, executable code module can be a single instruction or many instructions, and can even be distributed on a plurality of different code segments, distributed among different programs, and distributed across a plurality of memory devices. Similarly, operating data can be identified in the module, and can be implemented and organized in the data structure of any appropriate type according to any appropriate form. The operating data can be collected as a single data set, or can be distributed in different locations (including on different storage devices), and can only be present on a system or network as an electronic signal at least in part.

[0174] When a module can be implemented by software, considering the level of existing hardware technology, a person skilled in the art can build a corresponding hardware circuit to implement the corresponding function of the module that can be implemented by software without considering the cost. The hardware circuit includes a conventional very large scale integration (VLSI) circuit or gate array and existing semiconductors such as logic chips, transistors, or other discrete components. The module can also be implemented by a programmable hardware device, such as a field programmable gate array, a programmable array logic, a programmable logic device, etc.

[0175] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary personnel in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. An application recommendation method, characterized in that: include: Using the collaborative filtering recommendation framework CFGAN model based on generative adversarial neural networks, the user's corresponding application preference information is obtained according to application context factors and user attribute characteristics; The application context factor includes: at least one of application usage time and application usage location; An application recommendation is performed according to the application preference information.

2. The application recommendation method according to claim 1, characterized in that: The collaborative filtering recommendation framework CFGAN model based on generative adversarial neural network is used to obtain the application preference information corresponding to the user according to the application context factors and user attribute characteristics, including: Encode application context factors and user attribute characteristics according to data types to obtain model input data; the data types include enumeration types and numerical variables; The CFGAN model is used to obtain the application preference information corresponding to the user according to the model input data.

3. The application recommendation method according to claim 1 or 2, characterized in that: Also includes: The CFGAN model is trained by adopting a feature extraction capability enhancement operation; The feature extraction capability enhancement operation includes: using an attention mechanism to enhance the initial mask vector in the initial model.

4. The application recommendation method according to claim 3, characterized in that: The attention mechanism includes: at least one of a spatial attention SAM mechanism, a channel attention mechanism, and a permutation attention mechanism.

5. The application recommendation method according to claim 3, characterized in that: The feature extraction capability enhancement operation is adopted to train the CFGAN model, including: Inputting training data into the generator of the initial model to obtain predicted user application preferences; the training data includes: data on application context factors and user attribute characteristics; The initial mask vector is enhanced by using an attention mechanism to obtain a target mask vector; Obtaining a mask result according to the predicted user application preference and the target mask vector; Inputting the real user application preference, the real reference data and the mask result into the discriminator of the initial model to obtain a discrimination result; the real reference data includes: data on application context factors and user attribute characteristics; According to the discrimination result, the initial model is trained to obtain the CFGAN model.

6. The application recommendation method according to claim 1, characterized in that: Also includes: Using a factorization machine and a convolutional neural network, obtaining the user's first preference reference information and the second preference reference information; The first preference reference information corresponds to a first level of user interaction characteristics, and the second preference reference information corresponds to a second level of user interaction characteristics; The performing application recommendation according to the application preference information includes: Application recommendation is performed according to the application preference information, the first preference reference information and the second preference reference information.

7. An application recommendation device, characterized in that: include: The first processing module is used to obtain the application preference information corresponding to the user according to the application context factors and user attribute characteristics by using the collaborative filtering recommendation framework CFGAN model based on the generative adversarial neural network; The application context factor includes: at least one of application usage time and application usage location; The first recommendation module is used to make application recommendations based on the application preference information.

8. An application recommendation device, characterized in that: include: processor; The processor is used to obtain application preference information corresponding to the user according to application context factors and user attribute characteristics by using a collaborative filtering recommendation framework CFGAN model based on a generative adversarial neural network; The application context factor includes: at least one of application usage time and application usage location; An application recommendation is performed according to the application preference information.

9. An application recommendation device, comprising a memory, a processor, and a program stored in the memory and executable on the processor; characterized in that: When the processor executes the program, the application recommendation method according to any one of claims 1 to 6 is implemented.

10. A readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the application recommendation method according to any one of claims 1 to 6 are implemented.

11. A computer program product, characterized in that The method comprises computer instructions, which, when executed by a processor, implement the steps of the application recommendation method according to any one of claims 1 to 6.

Images