User Understanding Based On Large-scale Smartpone-sensed Data

A smartphone is regarded as an essential component of modern life.Frequent use of smartphones generates massive personal historical information.Typical personal informa-tion includes:1)location signal,derived from GPS and cell towers;2)application usage signal,derived from what an app being used at what time and for what;3)social signal,derived from CDR(call detail record),GPS,WiFi/Bluetooth connection,and contacts;4)individual activity,derived from accelerometer,camera,gyroscope,etc.Since a smartphone is usually tightly associated with a same person,it reveals rich clues regarding his/her be-haviors,life style,preferences,and so on.Smartphones are providing us an opportunity to understand users well.Well understanding of users,on one hand,can help us to improve devices,services and applications.For example,it is very important for personalization of applications,such as personalized web search,personalized recommendation,targeted advertising,and smart environments.On the other hand,understanding users from smartphone data can help users understand themselves objectively and extensively.Behaviors recorded by smartphones can help users discover objective and unobservable information about themselves.Besides,people’s memory capacity is limited while mobile phones can continuously collect detailed information about their behaviors for a long duration.The detailed records in mobile phones are important for understanding users extensively,so that the users can break bad living habits to improve life quality.Based on real-world large-scale datasets of smartphones,we attempted to understand users in three fold in this thesis:(1)Understanding Individual Mobility and Life Style from Anonymized WiFi ScanlistsFirst,we were attempting.to discover people’s life style from the anonymized WiFi scanlists.We built a mobility graph for each user to depict his/her mobility after extracting stay places from WiFi scanlists.From the mobility graphs,we detected individual activity areas through community detection.Based on the detected activity areas,we defined two metrics of activeness and diversity to measure individual mobility.Besides,importance places such as home and work place were identified,and users’ life style related to home and work place were studied,such as how many hours in average one stays at home,activeness of going outside in the evening,average working hours at work places on weekdays and weekends.The proposed approach was verified using the Device Analyzer data,which contains records of smart phone usage of more than 17,000 volunteering participants.(2)Mining User Attributes Using Large-scale APP Lists of SmartphonesIn addition to mobility pattern,we mined user attributes from an APP list.We de-veloped an attribute-specific representation to describe user characteristics,and then mod-eled the relationship between an attribute and an APP list.A large-scale real-world dataset with APP lists of more than 100,000 smartphones was used for evaluation.Our approach achieves the average EER(equal error rate)of 16.4%for twelve predefined user attributes.(3)Discovering Different Kinds of Smartphone Users Through Their Applica-tion Usage BehaviorsFinally,we discovered hundreds of user groups through their application usage behav-iors on smartphones.We analyzed one month of application usage from 106,762 Android users and discovered 382 distinct types of users based on their application usage behav-iors,using our own two-step clustering and feature ranking selection approach.Our results have profound implications on the reproducibility and reliability of mobile computing stud-ies,design and development of applications,determination of which apps should be pre-installed on a smartphone and,in general,on the smartphone usage experience for different types of users.