No abstract

Abstract-Many information tasks involve objects that are explicitly or implicitly connected in a network, such as webpages connected by hyperlinks or people linked by "friendships" in a social network. Research on link-based classification (LBC) has studied how to leverage these connections to improve classification accuracy. This research broadly falls into two groups. First, there are methods that use the original attributes and/or links of the network, via a link-aware supervised classifier or via a nonlearning method based on label propagation or random walks. Second, there are recent methods that first compute a set of latent features or links that summarize the network, then use a (hopefully simpler) supervised classifier or label propagation method. Some work has claimed that the latent methods can improve accuracy, but has not adequately compared with the best non-latent methods. In response, this paper provides the first substantial comparison between these two groups. We find that certain non-latent methods typically provide the best overall accuracy, but that latent methods can be competitive when a network is densely-labeled or when the attributes are not very informative. Moreover, we introduce two novel combinations of these methods that in some cases substantially increase accuracy.

Data describing networks such as social networks, citation graphs, hypertext systems, and communication networks is becoming increasingly common and important for analysis. Research on link-based classification studies methods to leverage connections in such networks to improve accuracy. Recently, a number of such methods have been proposed that first construct a set of latent features or links that summarize the network, then use this information for inference. Some work has claimed that such latent methods improve accuracy, but has not compared against the best non-latent methods. In response, this article provides the first substantial comparison between these two groups. Using six real datasets, a range of synthetic data, and multiple underlying models, we show that (non-latent) collective inference methods usually perform best, but that the dataset's label sparsity, attribute predictiveness, and link density can dramatically affect the performance trends. Inspired by these findings, we introduce three novel algorithms that combine a latent construction with a latent or non-latent method, and demonstrate that they can sometimes substantially increase accuracy.