UMBC ebiquity
Social Network Analysis and Mining
Other Times, Other Values: Leveraging Attribute History to Link User Profiles across Online Social Networks
September 1, 2016
Profile linking is the ability to connect profiles of a user on different social networks. Linked profiles can help companies to build psychographics of its potential customers and segment them for targeted marketing in a cost-effective way, can help advertisers target personalized ads and can help security practitioners capture detailed characteristics of malicious / fraudulent users. Existing methods link profiles by observing high similarity between most recent (current) values of the attributes like name and username. However, for a section of users who are observed to evolve their attributes over time and choose dissimilar values across their profiles, these current values have low similarity. Existing methods then falsely conclude that profiles refer to different users. To reduce such false conclusions, we suggest to gather rich history of values assigned to an attribute over time and compare attribute histories to link user profiles across networks. We believe that attribute history highlights user preferences and behavior while creating attribute values on a social network. Co-existence of these preferences across profiles on different social networks result in alike attribute histories that suggests profiles potentially refer to a single user. Through this study, we quantify the importance of attribute history for profile linking on a dataset of real-world users with profiles on Twitter, Facebook, Instagram and Tumblr. We show that attribute history correctly links 48% more profile pairs with non-matching current values that are incorrectly unlinked by existing methods. We further explore if factors such as longevity and availability of attribute history on either profiles affect linking performance. To the best of our knowledge, this is the first study that explores viability of using attribute history to link profiles on social networks.
Article
Springer
Downloads: 1507 downloads