Relational active learning for link-based classification

Abstract

Many information tasks involve objects that are explicitly or implicitly connected in a network (or graph), such as webpages connected by hyperlinks or people linked by "friendships" in a social network. Research on link-based classification (LBC) has shown how to leverage these connections to improve classification accuracy. Unfortunately, acquiring a sufficient number of labeled examples to enable accurate learning for LBC can often be expensive or impractical. In response, some recent work has proposed the use of active learning, where the LBC method can intelligently select a limited set of additional labels to acquire, so as to reduce the overall cost of learning a model with sufficient accuracy. This work, however, has produced conflicting results and has not considered recent progress for LBC inference and semi-supervised learning. In this paper, we evaluate multiple prior methods for active learning and demonstrate that none consistently improve upon random guessing. We then introduce two new methods that both seek to improve active learning by leveraging the link structure to identify nodes to acquire that are more representative of the underlying data. We show that both approaches have some merit, but that one method, by proactively acquiring nodes so as to produce a more representative distribution of known labels, often leads to significant accuracy increases with minimal computational cost.

Untitled

“…For instance, Bilgic et al [2010] studied active learning with sparsely labeled graphs using only CI. For this problem, the optimal label acquisition strategies could be quite different if RI or RCI were considered, since they could tolerate learning with fewer and more widely dispersed labels [McDowell 2015]. Similar changes, if neighbor attributes were included, may apply to other recent findings related to active exploration [Pfeiffer III et al 2014a] and active surveying [Namata et al 2012] in networks.…”

“…We call this approach SSL-EM. Xiang and Neville [2008] and Pfeiffer III et al [2014b, 2015 use this method with a fixed number of iterations (N Lrn = 10), whereas Lu and Getoor [2003b] repeat based on a convergence condition. 7 As described above, prior work has generally used either one or many learning iterations.…”

“…Finally, most SSL approaches for LBC have used a "hard" labeling approach: Steps 2 and 7 predict the single most likely label for each node, which is then used for feature computation (Step 5) and learning (Step 6). In contrast, some studies [Xiang and Neville 2008;Pfeiffer III et al 2014b, 2015 maintain "soft" probability estimates of the predicted labels for each node in V U and use these for learning (though they did not evaluate whether the "soft" learning was helpful). We thus evaluate additional "SOFTLRN" variants, which is the most consistent with the typical formulations of EM.…”

“…Section 7.3 and Appendix B considers the alternate forms of learning and inference, respectively. For SSL-EM, we mimic prior work [Xiang and Neville 2008;Pfeiffer III et al 2014b, 2015 by using N Lrn = 10 learning iterations. When using logistic regression, we also use "label regularization" [McDowell and Aha 2012] which biases the learning toward models that yield sensible label distributions (it does not apply to naive Bayes).…”