Mining Smartphone Data to Classify Life-Facets of Social Relationships Jun-Ki Min, Jason Wiese, Jason I. Hong, John Zimmerman Human-Computer Interaction Institute, Carnegie Mellon University [email protected], {jwwiese, jasonh, johnz}@cs.cmu.edu ABSTRACT Even social devices and services like smartphones and People engage with many overlapping social networks and social networking services (SNS) operate with a enact diverse social roles across different facets of their tremendously limited understanding of social role, often lives. Unfortunately, many online social networking classifying everyone as “friend” or “friend of friend.” services reduce most people’s contacts to “friend.” A richer Systems with a richer understand of role could assist people computational model of relationships would be useful for a in a variety of ways. At the highest level, these systems number of applications such as managing privacy settings could engage in much more situationally appropriate and organizing communications. In this paper, we take a behavior. More pragmatically, these systems could help by step towards a richer computational model by using call and organizing and prioritizing communications, by working to text message logs from mobile phones to classifying prevent unwanted self-disclosure or socially inappropriate contacts according to life facet (family, work, and social). behavior when sharing on SNSs, by reminding people of We extract various features such as communication the role they should be enacting before taking a phone call intensity, regularity, medium, and temporal tendency, and or engaging in other mediated communication, and by classify the relationships using machine-learning mining situational enactments of different roles to help techniques. Our experimental results on 40 users showed systems better understand the meaning of a place, of a that we could classify life facets with up to 90.5% accuracy. situation, or of the services people might most desire. The most relevant features include call duration, channel selection, and time of day for the communication. Some computational systems provide tools for users to manually label groups and assign their contacts to these Author Keywords groups; however, people do not appear to use these tools. Mobile social network; interpersonal relationships mining; One recent study reported that only 16% of people create life-facets; smartphone. any contact groups on their mobile phones [20]. In addition, ACM Classification Keywords Facebook reported that less than 5% of users create groups H.5.m. Information interfaces and presentation (e.g., HCI): within their set of friends [23]. We suspect that these Miscellaneous. features are rarely used because people do not perceive enough value in improved services to invest the time and General Terms attention necessary to categorize the hundreds of contacts Algorithms; Experimentation; Human Factors. they digitally maintain. The challenge is greater than simply INTRODUCTION classifying each contact once. Research shows relationships People enact many different social roles as they move are dynamic [35,3]. Kelley et al. argue that groups created between contexts and interact with different people. A for privacy purposes need to be periodically updated woman might enact the role of mother, wife, daughter, because of changes in relationships [25]. sister, neighbor, supervisor, colleague, teammate, Our goal is to develop methods for systems to infer social subordinate, chairperson, coach, and music patron all within role at a level of granularity that allows significantly a single day. Social roles provide a sort of invisible improved service offerings by mining logs of electronic structure, guiding people in their choice of actions in communications and sensor data from smartphones and various social situations. Computational systems, social media. Data from email and online SNSs provide rich interestingly, often have little or no understanding of the details on the communications between groups of people. many roles a person might play, of the behavioral Data from smartphones provide a person’s locations as well expectations associated with these roles, or of the specific as proximity to and co-location with others. For example, role a person is enacting when interacting with a system. previous work demonstrated how information on co- location patterns could be used to predict if two people are Permission to make digital or hard copies of all or part of this work for friends [10]. In addition, phones provide call and text personal or classroom use is granted without fee provided that copies are message (SMS) logs, capturing the: who, when, not made or distributed for profit or commercial advantage and that copies initiator/receiver, and the duration/length of many bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior communication events. The integration of all of this data specific permission and/or a fee. offers the opportunity to model human behavior and social CSCW ’13, February 23–27, 2013, San Antonio, Texas, USA. interactions at a scale and fidelity not previously possible. Copyright 2013 ACM 978-1-4503-1331-5/13/02...$15.00. This paper documents our investigation of one part of this Our work is an effort to help address these concerns by bigger picture, namely using smartphone communication providing a mechanism for systems to gain more detailed data (contact list, call logs, and SMS data) to classify the information about the specific relationship between people. life facet of a user’s contact as family, work, or social. Life Social Network Analysis and Group Mining facets are distinct domains within which people enact Researchers have addressed social network (SN) analysis different roles. We chose these facets based on previous on smartphones and SNS using both supervised and social networking research by Ozenc and Farnham [36]. unsupervised approaches [3,6,32]. Many of them have While these facets are broad and coarse-grained, they focused on tie strength, defining its properties as four provide initial evidence that concepts such as social role dimensions: amount of time, intimacy, intensity, and can be learned. Based on communication log data from 40 reciprocal services [19]. Studies have shown that the vast participants, we used machine-learning techniques to majority of interaction on SNS is with small numbers of classify life facets with 90% accuracy for contacts users had strong ties. For example, recent work by Burke suggested communicated with. that the average number of friends on Facebook is around This paper provides two contributions. First, it provides a 180 [5] (though many users have many more), while most set of extracted smartphones features that help to mine life people on Facebook only interact regularly with 4 to 6 facets. Second, it provides details of models built to classify people [42]. A different study examined people who posted life facets and an evaluation of these models based on a and tagged pictures of each other on Facebook, and found data set of 16,940 calls and 63,900 SMS messages. that on average people had 6.6 such “friends” [8]. BACKGROUND Based on network information like tie strength, researchers Kinds of Relationships have tried to analyze distinction of groups within a SN. Both social identity theory (from social psych) and identity Skeels and Grudin [38] found that people faced many theory (from sociology) agree that social role is a critical tensions when they tried to manage the co-presence of component in understanding human relationships multiple groups within their network. Lampinen et al. [26] [17,39,40]. People enact many different social roles and showed that people address group co-presence based on form relationships/affiliations with groups based on these behavioral strategies such as dividing the platform into roles. The degree that social role structure and granularity separate spaces, using suitable channels of communication, vary depends on individuals; however, social constructs like and performing self-censorship. In addition, several studies the life facets of family, work, and social are much more have demonstrated users’ desire to create groups of contacts universal [33,36]. In exploring her work/family border for practical applications like multi-level access control theory, Clark suggested that the greater the differences when sharing content [12,25,38]. Olson et al. [34] found across role domains, the less people engage in across-the- that people decide with whom to share information based border communication [9]. Similarly, Farnham and on the type of relationship, such as family and coworker. Churchill found that one approach people use to manage For example, Gilbert and Karahalios suggested that privacy roles and facets is to match them to specific communication controls based on tie strength might help to segment a mediums, such as only using the phone to communicate user’s SN into meaningful groups, and achieved 85% with your mom [17]. accuracy on binary classification of people’s strong or weak tie [18]. With their study on people’s privacy concern, Relative to offline social contexts (such as face-to-face Jones and OʼNeill found six criteria of grouping that people interaction), people experience new challenges in online commonly considered: social circles and cliques, tie contexts, which tend to be more broadcast oriented. This strength, temporal