Finding Local Experts on Twitter Zhiyuan Cheng, James Caverlee, Himanshu Barthwal, and Vandana Bachani Department of Computer Science and Engineering Texas A&M University College Station, TX, USA {zcheng, caverlee, barthwal, vbachani}@cse.tamu.edu ABSTRACT We address the problem of identifying local experts on Twitter. Specifically, we propose a local expertise framework that integrates both users’ topical expertise and their local authority by leverag- ing over 15 million geo-tagged Twitter lists. We evaluate the pro- posed approach across 16 queries coupled with over 2,000 indi- vidual judgments from Amazon Mechanical Turk. Our initial ex- periments find significant improvement over a naive local expert finding approach, suggesting the promise of exploiting geo-tagged Twitter lists for local expert finding. Categories and Subject Descriptors H.2.8 [Database Applications]: Data Mining Keywords Twitter, expert finding, local expert, social tagging 1. INTRODUCTION We tackle the problem of finding local experts in social media systems like Twitter. Local experts bring specialized knowledge about a particular location and can provide insights that are typi- cally unavailable to more general topic experts. A recent Yahoo! Research survey found that 43% of participants would like to di- rectly contact local experts for advice and recommendations online, while 39% would not mind being contacted by others [1]. And yet finding local experts is challenging. Traditional expert finding has focused on either small-scale, difficult-to-scale curation of experts (e.g., a magazine’s list of the “Top 100 Lawyers in Houston”) or on automated methods that can mine large-scale information sharing platforms. These approaches, however, have typically focused on finding general topic experts, rather than local experts. We present here our initial framework for local expert finding – LocalRank – that integrates both a person’s topical expertise and local authority. Our approach is motivated by the widespread adop- tion of GPS-enabled tagging of social media content via smart- phones and social media services (e.g., Facebook, Twitter, Foursquare) that provide a geo-social overlay of the physical environment. This massive scale geo-social resource provides unprecedented oppor- tunities to study the connection between people’s expertise and lo- cations. Concretely, LocalRank views a local expert as someone Copyright is held by the International World Wide Web Conference Com- mittee (IW3C2). IW3C2 reserves the right to provide a hyperlink to the author’s site if the Material is used in electronic media. WWW’14 Companion, April 7–11, 2014, Seoul, Korea. ACM 978-1-4503-2745-9/14/04. http://dx.doi.org/10.1145/2567948.2577354. (a) @BBQsnob (b) @JimmyFallon Figure 1: Heatmap of the location of Twitter users who have listed @BBQsnob or @JimmyFallon who is well recognized by the local community, where we estimate this local recognition via a novel spatial proximity expertise ap- proach that leverages over 15 million geo-tagged Twitter lists. To illustrate, Figure 1(a) shows a heatmap of the locations of Twitter users who have labeled Daniel Vaughn (@BBQsnob) on Twitter. As one of the foremost barbecue experts in Texas, Vaughn’s ex- pertise is recognized regionally in Texas, and more specifically by local barbecue centers in Austin and Dallas. In contrast, late-night host Jimmy Fallon’s heatmap suggests he is recognized nationally, but without a strong local community. Intuitively, Daniel Vaughn is recognized as a local expert in Austin in the area of Barbecue; Jimmy Fallon is certainly an expert (of comedy and entertainment), but his expertise is diffused nationally. 2. PROBLEM STATEMENT AND SOLUTION We are interested to find local experts with particular expertise in a specific location. We assume there is a pool of expert candidates V = {v1,v2, ..., vn}, that each candidate vi has an associated lo- cation l(vi ) and a set of areas of expertise described by a feature vector  vi . Each element in the vector is associated with an expertise topic word tw (e.g., “technology”), and the element value indicates to what extent the candidate is an expert in the corresponding topic. We define the Local Expert Finding problem as: DEFINITION 1. (Local Expert Finding) Given a query q that includes a query topic t(q), and a query location l(q), find the set of k candidates with the highest local expertise in query topic t(q) and location l(q). Topical vs. Local Authority: Identifying a local expert requires that we can accurately estimate not only the candidate’s expertise on a topic of interest (e.g., how much does this candidate know about barbecue), but also that we can identify the candidate’s local authority (e.g., how well does the local community recognize this candidate’s expertise). Hence, we propose to decompose the local expertise for a candidate vi into two related dimensions: (i) Topical Authority: which captures the candidate’s expertise on the topic area t(q); and (ii) Local Authority: which captures the candidate’s local authority in query location l(q). The local experts we are trying to identify should have both great topical authority and local authority: e.g., Daniel Vaughn (@bbqsnob) is an example of an expert with high topical authority (on barbecue), as well as high local authority (in Texas). 241