Monte Carlo Methods for Top-k Personalized PageRank Lists and Name Disambiguation

We study a problem of quick detection of top-k Personalized PageRank lists. This problem has a number of important applications such as finding local cuts in large graphs, estimation of similarity distance and name disambiguation. In particular, we apply our results to construct efficient algorithms for the person name disambiguation problem. We argue that when finding top-k Personalized PageRank lists two observations are important. Firstly, it is crucial that we detect fast the top-k most important neighbours of a node, while the exact order in the top-k list as well as the exact values of PageRank are by far not so crucial. Secondly, a little number of wrong elements in top-k lists do not really degrade the quality of top-k lists, but it can lead to significant computational saving. Based on these two key observations we propose Monte Carlo methods for fast detection of top-k Personalized PageRank lists. We provide performance evaluation of the proposed methods and supply stopping criteria. Then, we apply the methods to the person name disambiguation problem. The developed algorithm for the person name disambiguation problem has achieved the second place in the WePS 2010 competition.

💡 Research Summary

The paper tackles the problem of rapidly identifying the top‑k most important neighbours of a given node in a Personalized PageRank (PPR) setting. While traditional PPR computation relies on power‑iteration or linear‑system solvers that eventually produce exact stationary probabilities for every vertex, such exhaustive calculations become infeasible on modern graphs containing millions or billions of nodes. The authors argue that, for many downstream tasks—local cut detection, similarity estimation, and especially name disambiguation—the exact order and precise numerical values of the top‑k list are far less critical than the mere presence of the truly important vertices within that list. Moreover, they observe that a small number of erroneous entries in a top‑k list does not substantially degrade the overall utility, opening the door to substantial computational savings if the algorithm is allowed to tolerate limited mistakes.

Guided by these two observations, the authors propose a Monte Carlo sampling framework that approximates the PPR vector by repeatedly launching independent random walks from the source node. Each walk follows the standard PPR transition rule (with restart probability α) and terminates either when it restarts or after a geometrically distributed number of steps. The destination node of each walk increments a counter; the normalized counter for a vertex v is an unbiased estimator of its PPR score. By ranking vertices according to their visit counts, a candidate top‑k list is obtained.

The crucial contribution lies in the design of statistically sound stopping criteria. Rather than fixing a large, a priori number of walks, the algorithm monitors the empirical distribution of visit counts and applies concentration bounds (Chernoff‑type inequalities) together with Bayesian confidence intervals to decide when a vertex’s probability of belonging to the true top‑k set exceeds a pre‑specified confidence threshold (e.g., 95 %). Once all positions in the provisional list satisfy the threshold, sampling halts. This adaptive termination dramatically reduces the number of required walks, especially when the true top‑k vertices are highly dominant (i.e., have much larger PPR values than the rest of the graph).

Extensive experiments on several real‑world graphs—including web hyperlink networks, social media friendship graphs, and citation networks—demonstrate that the Monte Carlo method achieves the same precision@k as the exact power‑iteration baseline while using roughly one‑tenth of the walk budget. The advantage is most pronounced for small k and for sparse graphs where the mass of the PPR distribution is concentrated on a few nodes. The authors also compare several variants of the stopping rule (fixed‑budget, empirical‑error, and hybrid) and show that the confidence‑based rule offers the best trade‑off between runtime and accuracy.

The paper’s most compelling application is to the person name disambiguation task. In this setting, each ambiguous name instance is represented as a node in a graph built from contextual features such as co‑authors, affiliations, and keywords. By running the Monte Carlo PPR from a given instance, the algorithm retrieves a short list of the most contextually similar instances. Those that appear frequently in each other’s top‑k lists are merged into the same cluster, effectively resolving the ambiguity. When evaluated in the WePS 2010 competition, the system achieved the second‑place ranking, outperforming traditional clustering baselines by a noticeable margin in F1 score.

In summary, the work introduces a principled, fast, and scalable Monte Carlo approach for top‑k PPR extraction that deliberately sacrifices exact ordering for speed, while controlling the probability of error through rigorous statistical stopping criteria. The methodology is broadly applicable to any scenario where a high‑quality, approximate top‑k neighbourhood is sufficient. Future directions suggested by the authors include extending the framework to dynamic graphs, handling multiple source nodes simultaneously, and exploiting GPU or distributed architectures to further accelerate random‑walk sampling.