Group Affiliation Detection Using Model Divergence for Wearable Devices Dawud Gordon TwoSense Labs Karlsruhe, Germany dawud@twosense-labs.com Martin Wirz ETH Z ¨ urich Z¨ urich, Switzerland wirz@ife.ee.ethz.ch Daniel Roggen University of Sussex Sussex, United Kingdom daniel.roggen@ieee.org Gerhard Tr¨ oster ETH Z ¨ urich Z¨ urich, Switzerland troester@ife.ee.ethz.ch Michael Beigl KIT / TecO Karlsruhe, Germany michael.beigl@kit.edu ABSTRACT Methods for recognizing group affiliations using mobile de- vices have been proposed using centralized instances to ag- gregate and evaluate data. However centralized systems do not scale well and fail when the network is congested. We present a method for distributed, peer-to-peer (P2P) recog- nition of group affiliations in multi-group environments, us- ing the divergence of mobile phone sensor data distributions as an indicator of similarity. The method assesses pairwise similarity between individuals using model parameters in- stead of sensor observations, and then interprets that informa- tion in a distributed manner. An experiment was conducted with 10 individuals in different group configurations to com- pare P2P and conventional centralized approaches. Although the output of the proposed method fluctuates, we can still correctly detect 93% of group affiliations by applying a fil- ter. We foresee applications in mobile social networking, life logging, smart environments, crowd situations and possibly crowd emergencies. Author Keywords Group affiliation detection; computational social sciences; peer-to-peer; wearable computing; mobile computing; ACM Classification Keywords I.2.11 Distributed Artificial Intelligence: Miscellaneous 1. INTRODUCTION AND MOTIVATION Around 70% of the time we spend in public areas is done to- gether with other people [16]. In general we are social crea- tures and spend a great deal of our time in groups of one form or another [9]. Groups are better than individuals at accomplishing tasks, which is often why they are formed in the first place [9]. Understanding group behavior and context is then crucial for systems which are trying to assist these Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from Permissions@acm.org. ISWC’14, September 13-17 2014, Seattle, WA, USA Copyright is held by the owner/author(s). Publication rights licensed to ACM. ACM 978-1-4503-2969-9/14/09. . . $15.00. http://dx.doi.org/10.1145/2634317.2634319 groups in some fashion [10]. Before an understanding of the group’s context can be reached, group and individual affilia- tions must be identified through the precess of group affilia- tion detection (GAD). Often times several groups can occupy the same space at once [16], making it important to detect non-affiliation as well as affiliation. A group is two or more individuals who are connected to each other by social relationships [9]. Humans have an in- nate ability to visually recognize these groups quickly [16], using unconscious processes which can be described using the Gestalt Laws [9]. Our minds automatically observe and group objects together based on proximity, similarity and in- teraction. It is this perception process of detecting groups and affiliations which GAD proposes to emulate [17]. Since human-like perception is the goal, we are therefore bound to that perception as it defines correct and incorrect affiliation decisions. The problem is then to differentiate inter-group similarity from intra-group similarity. Members of the same group have similar physical behav- ior because group members often perform activities together [17], adopt behavioral norms of the group [9], and mimic be- havior of interaction partners [7]. By sensing these behavioral similarities, or “social proximity,” we can effectively detect groups, and group affiliations [14]. Information from wear- able sensors is centralized, features indicative of affiliation are extracted, and the result is clustered to identify groups and affiliations [17]. However in situations where centralized ag- gregation is not practical [11], such as emergencies [4], new methods for evaluating group affiliation using P2P analysis systems must be explored. We present a method for P2P assessment of group affiliation by modeling the data as a distribution and then calculating the disparity (or similarity) as the Jeffrey’s divergence be- tween models from different individuals. We call this method divergence-based affiliation detection (DBAD). We compare DBAD with centralized and distributed approaches using sig- nal correlation which is the basis for previous approaches [17, 14]. We present 2 methods for accomplishing GAD, one where nodes exchange Gaussian probability density func- tions (DBAD-P) of sensor data, and another where they ex- change histograms of observations (DBAD-H). We evaluate these methods with an experiment involving 10 individuals