However the paper also got me thinking about link prediction. Here's a quote from the paper:
Link prediction - the matrix $A$ is the adjacency matrix of a partially observed graph; entries are 0 for both not-existing and undiscovered links. The search space is unrestricted as before and the matrix $S$ contains the scores for link prediction; the ideal loss function is the empirical average of the zero-one loss for each coefficient, \[So I read that as, ``this is a P-U problem that we are reducing to pointwise classification.'' However my preferred method for P-U problems is to reduce to ranking (AUC loss). What would that look like for link prediction?
l_{E} (S, A) = \frac{1}{|E|} \sum_{(i,j) \in E} 1_{(A_{ij} - 1/2) \cdot S_{ij} \leq 0}.
\]
- Instances are edges (i.e., pairs of vertices plus dyadic specific information).
- Reduction of AUC is to pairwise classification, so pairs of edges, or pairs of pairs of vertices.
- Each positive (observed) edge in the adjacency graph would be paired with an unlabeled (unobserved or unexplored) edge, the latter perhaps drawn uniformly from all possible edges; or possibly from all possible edges given one vertex (``per-vertex AUC'').
- The final classification model could be purely latent (e.g., pure LFLL), purely explicitly feature driven (e.g., bilinear implemented with VW), or a combination (e.g., LFLL with side information).
- In my experience LLFL with side information is very tricky to train, unlike pure LLFL.
Next time I run into a link prediction problem I'm going to give this a whirl.
This (link prediction with latent factor models optimized for AUC) is basically what we did in "our" (St. and Chr. did most of the work) UAI 2009 paper: Bayesian Personalized Ranking from Implicit Feedback. Works pretty well. It was also used by some of the top teams in KDD Cup 2011 (Track 2).
ReplyDeleteThat's good to know. Thanks for the note!
Delete