Research On The Prediction Of Mobile App Usage

Recent burst of mobile devices, mobile operating systems and online app stores hashelp to generate a prospering mobile app market. Mobile apps have taken up much timeof the users, bringing new problems and challenges. For example, quantities of installedmobile apps make it a headache for the users to find the app they want, deterioratingthe user experience. Meanwhile, the special features of mobile app usage(MAU), suchas invariant gestures and unitary subject, disable the traditional research methodology ofhuman behavior understanding. However, if MAU is successfully predicted, intelligentrecommendations can be made to reduce the cost of finding an app. Being combined withthe context of mobile usage, the prediction can help to achieve better understanding ofhuman behavior.By proving the predictability, researching on the predicting algorithm, collecting us-agedataandconductingaseriesofexperiments, thispaperfundamentallysolvestheprob-lem of MAU prediction. This paper originally abstracts the user as a discrete stationaryinformation source with memoryeffectand MAUas corresponding outputseries, theoret-ically proves the predictability of MAU by analyzing the entropy rate. Furthermore, thispaper gives the methods to calculate the upper and lower bounds of predictability. Fol-lowing that, predicting algorithms respectively based on Support Vector Machine(SVM)and 3-Layer Back Propagation Neural Network(3LBPNN) have been discussed. We de-signed and developed AppMagic on the Android platform to collect MAU data. Ensuringan average disturbance of 3.8‰, AppMagic has collected usage data from 25 volunteer-s over 25 days, generating a data set with 122547 log items. Experiments conductedover the data set show that, the upper bound of predictability is stable, indicating thepredictability is a general feature of MAU. The average upper bound is 82.4%, whichreveals the feasibility of predicting MAU. Varying the relative parameters, we find that,SVM-based algorithm has achieved the best predicting accuracy when the smoothing pa-rameter is augmented to an appropriate value, and for the 3LBPNN-based algorithm, thepeak accuracy is with training times set to 150,000. The SVM-based algorithm outper-forms the 3LBPNN-based algorithm both in general and at the peak point, fitting betterfor the prediction of MAU. After the noise-app filter is added, both algorithms gain ansignificant improvement, which indicates the existence of faulty and unmeant app usage. Average prediction accuracy of improved SVM-based algorithm has reached 73.4% inthe end, basically satisfying the requirements of prediction. At last, the time consumptionof 3LBPNN-based algorithm is less, makes it more suitable for the scenario that imposesrigid requirement on real-time property.