Citation Context

...ights result in higher autocorrelation and lead to improved classification accuracy. 1 Introduction Recent research on analyzing social networks has demonstrated that relational patterns of homophily =-=[13]-=- can be exploited to improve predictive models of both link structure and behavior. For example, researchers have found that network connectivity and attribute similarity can improve link prediction m...

Citation Context

...e exploited to improve predictive models of both link structure and behavior. For example, researchers have found that network connectivity and attribute similarity can improve link prediction models =-=[12, 17]-=-. Also, researchers have found that relational ties can improve behavior prediction in tasks such as as fraud detection [14] and viral marketing [5]. However, much of this past work has focused on soc...

Citation Context

...ntiate between the two ends of the spectrum. Since pairs of individuals with strong ties (e.g., close friends) are likely to exhibit greater similarity than those with weak ties (e.g., acquaintances) =-=[7]-=-, treating all relationships equally will increase the level of noise in the learned models and likely lead to degradation in performance. Fortunately, online social networks (OSNs) often consist of m...

Citation Context

... profile attributes, we considered a binary classification task based on its most frequent value. We apply a widely-used semi-supervised classification algorithm—the Gaussian random field (GRF) model =-=[18]-=-, which assumes autocorrelation is present in the graph and propagates information from the labeled portion of the graph to infer the values for unlabeled nodes. Since the GRF is designed for applicat...

Citation Context

... similarity can improve link prediction models [12, 17]. Also, researchers have found that relational ties can improve behavior prediction in tasks such as as fraud detection [14] and viral marketing =-=[5]-=-. However, much of this past work has focused on social networks with binary relational ties (e.g., friends or not). These binary indicators provide only a coarse indication of the nature of the relat...

Citation Context

...pace, and LinkedIn has lead to a surge in research focused on modeling networks and their properties. Much of this work has focused on the analysis of network structure and growth patterns (see e.g., =-=[2, 3, 16]-=-). However, nearly all these methods focus on descriptive analysis and generative models of link structure, based on the observed structure in the network—they do not attempt to model the latent prope...

Citation Context

...ile similarity and interaction activity, with the goal of automatically distinguishing strong relationships from weak ones. 1Recently, interaction data has been used to predict relationship strength =-=[10, 6]-=- but this work only considered two levels of relationship strength, namely weak and strong relationships. In addition, this past work focused on supervised learning methods, which requires human annot...

Citation Context

...pace, and LinkedIn has lead to a surge in research focused on modeling networks and their properties. Much of this work has focused on the analysis of network structure and growth patterns (see e.g., =-=[2, 3, 16]-=-). However, nearly all these methods focus on descriptive analysis and generative models of link structure, based on the observed structure in the network—they do not attempt to model the latent prope...

Citation Context

...process of social influence—when people who interact frequently become more similar—is another cause of relational autocorrelation that may affect link formation differently than homophily (see e.g., =-=[2]-=-). An important direction for future work is to model these two effects in a joint model of link strength, particularly in domains where the attribute values change over time. Acknowledgments This res...

Citation Context

...e exploited to improve predictive models of both link structure and behavior. For example, researchers have found that network connectivity and attribute similarity can improve link prediction models =-=[12, 17]-=-. Also, researchers have found that relational ties can improve behavior prediction in tasks such as as fraud detection [14] and viral marketing [5]. However, much of this past work has focused on soc...

Citation Context

... [15] also model interaction events, but they formulate a temporal link prediction task which tries to predict the occurrence of events (e.g., co-authorship) in future time intervals. Adamic and Adar =-=[1]-=- also investigate the use of ancillary network information but with the goal of predicting social ties, instead of tie strength. More recently, interaction data has been utilized in predicting relatio...

Citation Context

...s. This network can also be viewed as an interaction network. Autocorrelation improvement. In relational data, autocorrelation is the statistical dependency of the same attribute on related instances =-=[9]-=-. To measure the autocorrelation of a categorical attribute 5(a) gender (b) relationship status (c) political views (d) religious views Figure 2: Autocorrelation on various graphs, as link density is...

Citation Context

...o model the latent properties of the networks. Another direction of related research has focused on link prediction—which is a formulization of the problem of predicting future links in a social network, given a snapshot of the network at the current time step. This is the area of research that is most relevant to our work in this paper. Link prediction methods can be generally grouped into two approaches: those that use topological features to capture the link structure of the network (e.g., [13, 12]) and those that use attribute similarity features in addition to topological features (e.g., [19, 9]). We differ from past work on link prediction in that we focus on modeling link strength rather than link existence. We also aim to exploit interaction information among nodes in order to improve model accuracy. O’Madadhain et al. [16] model interaction events, but they formulate a temporal link prediction task which tries to predict the occurrence of events (e.g., co-authorship) in future time intervals. Adamic and Adar [1] also investigate the use of ancillary network information but with the goal of predicting social ties, instead of tie strength. More recently, interaction data has been u...

Citation Context

...pace, and LinkedIn has lead to a surge in research focused on modeling networks and their properties. Much of this work has focused on the analysis of network structure and growth patterns (see e.g., =-=[2, 3, 16]-=-). However, nearly all these methods focus on descriptive analysis and generative models of link structure, based on the observed structure in the network—they do not attempt to model the latent prope...

Citation Context

...onnectivity and attribute similarity can improve link prediction models [12, 17]. Also, researchers have found that relational ties can improve behavior prediction in tasks such as as fraud detection =-=[14]-=- and viral marketing [5]. However, much of this past work has focused on social networks with binary relational ties (e.g., friends or not). These binary indicators provide only a coarse indication of...

Citation Context

...ink prediction in that we focus on modeling link strength rather than link existence. We also aim to exploit interaction information among nodes in order to improve model accuracy. O’Madadhain et al. =-=[15]-=- also model interaction events, but they formulate a temporal link prediction task which tries to predict the occurrence of events (e.g., co-authorship) in future time intervals. Adamic and Adar [1] a...

Citation Context

...ile similarity and interaction activity, with the goal of automatically distinguishing strong relationships from weak ones. 1Recently, interaction data has been used to predict relationship strength =-=[10, 6]-=- but this work only considered two levels of relationship strength, namely weak and strong relationships. In addition, this past work focused on supervised learning methods, which requires human annot...

Citation Context

... most relevant to our work in this paper. Link prediction methods can be generally grouped into two approaches: those that use topological features to capture the link structure of the network (e.g., =-=[12, 11]-=-) and those that use attribute similarity features in addition to topological features (e.g., [17, 8]). We differ from past work on link prediction in that we focus on modeling link strength rather th...

Citation Context

...finite amount of resources (e.g., time) to use in the formation and maintenance of relationships, it is likely that they direct these resources towards the relationships that they deem more important =-=[4]-=-. Such interactions could be, for example, profile viewing activities between the pair of users, connection establishment, picture tagging, etc. The stronger the relationship, the higher likelihood th...

Citation Context

...dation in performance. Indeed, recent research that has attempted to prune away spurious relationships and highlight stronger relationships has been shown to improve the accuracy of relational models =-=[17]-=-. Fortunately, online social networks (OSNs) often consist of more than just a record of social network ties. Typically online communities contain ancillary interaction information among the users tha...

Citation Context

...proaches: those that use topological features to capture the link structure of the network (e.g., [12, 11]) and those that use attribute similarity features in addition to topological features (e.g., =-=[17, 8]-=-). We differ from past work on link prediction in that we focus on modeling link strength rather than link existence. We also aim to exploit interaction information among nodes in order to improve mod...