Vol.:(0123456789) 1 3 Social Network Analysis and Mining (2018) 8:39 https://doi.org/10.1007/s13278-018-0516-z ORIGINAL ARTICLE Learning to lurker rank: an evaluation of learning‑to‑rank methods for lurking behavior analysis Diego Perna 1  · Roberto Interdonato 2  · Andrea Tagarelli 1 Received: 8 December 2017 / Revised: 11 April 2018 / Accepted: 19 May 2018 © Springer-Verlag GmbH Austria, part of Springer Nature 2018 Abstract While being long researched in social science and computer–human interaction, lurking behaviors in online social networks (OSNs) have been computationally studied only in recent years. Remarkably, determining lurking behaviors has been mod- eled as an unsupervised, eigenvector-centrality-based ranking problem, and it has been shown that lurkers can efectively be ranked according to the link structure of an OSN graph. Although this approach has enabled researchers to overcome the lack of ground-truth data at a large scale, the complexity of the problem hints at the opportunity of learning from past lurking experiences as well as of using a variety of behavioral features, including any available, possibly platform-specifc information on the activity and interaction of lurkers in an OSN. In this paper, we leverage this opportunity in a principled way, by proposing a machine-learning framework which, once trained on lurking/non-lurking examples from multiple OSNs, allows us to predict the ranking of unseen lurking behaviors, ultimately enabling the prioritization of user engagement tasks. Results obtained on 23 network datasets by state-of-the-art learning-to-rank methods, using diferent optimization and evalu- ation criteria, show the signifcance of the proposed approach. Keywords Lurking · Learning-to-rank · Behavior analysis · User engagement 1 Introduction Lurking in an online social network (OSN) characterizes those users in the crowd who do not signifcantly take an active and tangible role in the interaction with other mem- bers (Nonnecke and Preece 2000; Preece 2004; Edelmann 2013; Sun et al. 2014). Such users are referred to as lurkers, since they gain beneft from information produced by oth- ers, though their presence is legitimated (Lave and Wenger 1991), expected and even welcome (Pan et al. 2014; Tsai and Pai 2014). More importantly, lurkers might hold potential social capital, because they acquire knowledge from the OSN: by observing user-generated communications, they can form their own opinions and even expertise, though they rarely will let other people know their “value”. Therefore, it might be desirable to make lurkers’ social capital available to other users (Farzan et al. 2010). This can be accomplished through mechanisms of engagement (Easley and Kleinberg 2010; Farzan and Brusilovsky 2011; Malliaros and Vazir- giannis 2013; Rowe 2013; Imlawi and Gregg 2014), ulti- mately encouraging lurkers to more actively participate in the OSN life. In Tagarelli and Interdonato (2013, 2014), the authors presented the frst tool, namely, LurkerRank, for automati- cally bringing order into the crowd of users that may show lurking behaviors at varying degrees. LurkerRank adopts a query-independent, eigenvector-centrality-based approach that utilizes the link graph structure underlying user relation- ships (e.g., followships, like/comment interactions). Its rank- ing solution can enable a way to prioritize the engagement of (top-ranked) lurkers. LurkerRank has also been adapted as query-dependent (e.g., trust-biased) ranking method (Interdonato and Tagarelli 2016; Tagarelli and Interdonato 2014), extended to handle time-evolving networks (Tagarelli An abridged version of this paper appeared in Perna and Tagarelli (2017). * Andrea Tagarelli andrea.tagarelli@unical.it Diego Perna d.perna@dimes.unical.it Roberto Interdonato roberto.interdonato@cirad.fr 1 DIMES, University of Calabria, Rende, Italy 2 Cirad, UMR Tetis, Montpellier, France