Exploring profile textual features for cross-network linkability: application to Quora and Twitter users

Content-based user identity linkage across different social platforms has been explored extensively during the last decade. Existing techniques investigated the use of personal discrete attributes such as user name, gender and email. Using discrete attributes for linking profiles has serious drawbacks, as these attributes are inconsistent and non-authentic in many platforms. In this paper, we suggest a matching approach that explores the textual information contained in user posts rather than their discrete profile information. However, we face major constraints as the profiles' textual representation can be very sparse. In addition, the absence of any discrete attributes in the matching scheme makes the problem quadratic, as no profile pairwise filtering can be done. We tackle this by suggesting a clustering method based on locality-sensitive hashing that first clusters users into sub-groups based on their topical representation and then selects the true matching pair based on their token signatures. Our solution significantly reduces the problem complexity by reducing the number of pairwise profiles to compare by two orders of magnitude, and maintains a high matching precision for users that have sufficient amount of textual data.