Mobile-User-Behavior-Based Application Recommendation Algorithm And Framework

The Mobile Application Market is enduring exponential expansion. And the information overload problem arised in mobile application field. Recommender system, as the classic information overload solution, is scuessful in other area. However, specialized in the mobile application scenario, traditional recommender system suffers from two main obstacles. For one thing, with the rigid privacy restriction,user laziness and commercial restriction, it's intractable to collect high quality rating data for collaborative filtering recommender system in most cases. For another, the absence of describing mechanism for users clustering level evolving in long period application usage leads to the inefficacy for improving recommeder system results with the present user behavior analysis algorithms.To curb these drawbacks, this dissertation provides an implicit rating algorithm based on user behavior logs and a user behavior transition analysis algorithm based on self-organizing map. Besides, this thesis also implemented a prototype mobile application recommeder system framework to provide the big data computing and storage support for these models and algorithms.In this dissertation, the first contribution is the implicit rating generation algorithm based on mobile user behavior logs files. This model utilizes user behavior logs to generate user-appliction ratings which can be analyzed effectively by exstiting recommeder systems. Given the user behavior logs are friendly to data collecting program, the protential collectable behavior logs datasets are massive and the insufficient data problem can be eased. The implicit rating algorithm is composed by behavior model and rating model which are both based on Gaussian Mixture Model.The behavior model captures the users' application usage pattern and the rating model generates user-application ratings.Another focus for this thesis is the user behavior transition analysis model based on self-organizing map coordinates.The purpose of this algorithm is to offer a representation for user cluster level behavior transition. Based on such representation, mature sequence mining algorithms can be applied to provide user application usage transition analysis, and the analysis result can be utilized by recommender system to improving recommendations.Besides, the mobile-app recommender system prototype implemented by this article consists storage model and recommender module. Such prototype provides necessary support for implicit rating model and user transition analysis algorithm mentioned above.Finally, this study designed a series of experiments to verify the performance of defined rating model and user transition analysis algorithm. The storage model and recommender moule have also been tested. The experiment results indicate that on the unrated dataset the implicit rating algorithm provided average 13% improvement compared with supervised rating model in F1 value. Given the same presicion lost, the user behavior analysis outperformed other random remcommder algorithms(GRank and RRank) in novelity gain. The storage model and hybrid recommeder moudule can provide slightly higer performance than similar implementations.