Research On Personalized Recommendation Methods For Mobile Applications

In recent years,the development of mobile Internet is in the ascendant,and mobile devices(such as smart phones,tablet computers,etc.)have become mainstream mobile network terminals.Mobile applications(hereinafter referred to as apps)deployed on mobile devices have brought great convenience to people's production and life.Although a large number of apps meet people's increasing daily life and work needs,due to the wide variety and large number of apps,it is also difficult for users to choose apps that truly meet their preferences.Therefore,there is an urgent need to explore scientific methods to help users accurately and quickly select apps that meet their preferences.The app recommender system leverages the historical data of users' use of mobile apps to mine users' preferences,and then recommends mobile apps that meet their preferences to users.Therefore,the implementation of personalized app recommendation can meet the needs of users for personalized and accurate selection of mobile applications.Personalized app recommendation for users will not only help users get a better experience in the mobile Internet Ecosystem,but also help mobile app providers promote their services and products.Therefore,mobile application recommendation system is getting the attention of academia and industry,and has become a hot research direction.The difference between app recommendation and recommendation of other items is that because the app is deployed on mobile devices and has the characteristics of "use anytime,anywhere",contextual information such as the time and location of the users' interaction with the apps has a significant impact on the recommendation result.In addition,app has functional exclusiveness features,which leads app recommendation to face different challenges from recommendation in the traditional field.Furthermore,the development of apps has the characteristics of low difficulty,short cycle,and version evolution,which makes the interaction data between users and apps become sparser than mass apps,which brings great difficulties to accurate recommendation.In order to cope with the above challenges,this research has combed the three levels of app context characteristics,user behavior trajectory and user-app interaction based on the interaction data between users and apps,and comprehensively explored personalized recommendation methods for mobile applications.The main research contributions of this study are summarized as follows.(1)The existing mainstream app methods focus on exploiting user characteristics to mine user preferences,while ignoring the unique temporal and spatial characteristics of the app when it is used.This research comprehensively considers the two characteristics of users and app context factors,and proposes an app recommendation method based on app context factors,called PCMARA.PCMARA generates an app recommendation list for users based on the user's current situational information.Specifically,(i)PCMARA uses the Gaussian mixture model to construct the app context factor,and based on the app context factor,designs an app similarity model construction method based on the temporal and spatial characteristics of the app.(ii)Secondly,PCMARA uses a tensor model to dynamically predict the user's app preference.(iii)Finally,PCMARA comprehensively considers the current situational characteristics of both the user and the app,and generates a recommendation list for the target user according to the current time and location of the target user.PCMARA is applied to real-world datasets,and large-scale experiments are carried out.The experimental results show that PCMARA has achieved satisfactory recommendation performance.(2)The contextual information has the characteristics of continuity and periodicity in the space-time dimension,so the users' current context is related to the previous context.Therefore,when mining users' preferences,it is one-sided to only consider the current contexts of target users and ignore the previous context.This also hinders the further improvement of the recommendation effect.Based on this fact,this research proposes a next app recommendation method based on user behavior trajectories and named CMARA.CMARA generates the next app recommendation list for target users according to users' historical behavior information and current contextual information.Specifically,(i)CMARA integrates heterogeneous information of target users(such as the apps,time and location used by the user at a certain context)into the users' behavior trajectories to model the users' app preference;(ii)CMARA leverages the users' contextual points to construct the contextual Voronoi diagram,and use the contextual Voronoi diagram to construct a novel user similarity model;(iii)CMARA uses the current context information of target users to generate the next app recommendation list that meets the users' preferences.Through experiments on large-scale real-world data,this study verifies that CMARA has better recommendation performance than other benchmark recommendation methods.(3)Whether a user and an app are fit under contextual conditions directly affects whether the user and app can interact.Therefore,it is worth studying to explore the deep-level contextual relationship between users and mobile apps.This research proposes an app recommendation framework based on contextual feature interaction,called CFDIL.CFDIL attempts to explore the users' app preferences in a specific context from the perspective of user-app interaction,and make recommendations.Specifically,first of all,CFDIL uses the contextual characteristics information and attributes information of users and apps to construct the feature portraits of users and apps.Then,CFDIL trained a deep learning framework based on feature portraits,thereby effectively mining the multi-level interaction features between users and apps under specific contextual conditions.Finally,CFDIL provides users with more accurate app recommendations through factorization machines and convolutional neural networks.In addition,the construction of feature portraits and the tensor factorization operation of labels effectively alleviate the adverse effects of sparse data on the recommendation model.Experimental results show that,compared with other recommendation methods,CFDIL has achieved better app recommendation performance.