We revisit semi-supervised learning on hypergraphs. Same as previous approaches, our method uses a convex program whose objective function is not everywhere differentiable. We exploit the non-uniqueness of the optimal solutions, and consider confidence intervals which give the exact ranges that unlabeled vertices take in any optimal solution. Moreover, we give a much simpler approach for solving the convex program based on the subgradient method. Our experiments on real-world datasets confirm that our confidence interval approach on hypergraphs outperforms existing methods, and our subgradient method gives faster running times when the number of vertices is much larger than the number of edges. Our experiments also support that using directed hypergraphs to capture causal relationships can improve the prediction accuracy. Furthermore, our model can be readily extended to capture multiclass learning.
12 June 2010 in Journal of Wuhan University of Technology-Mater. Sci. Ed.
Static and dynamic clustering algorithms are a fundamental tool in any machine learning library. Most of the efforts in developing dynamic machine learning and data mining algorithms have been focusing on the sliding window model (where at any given point in time only the most recent data items are retained) or more simplistic models. However, in many real-world applications one might need to deal with arbitrary deletions and insertions. We develop a $(2+\epsilon)$ -approximation algorithm for the k-center clustering problem with ``small'' amortized cost under the fully dynamic adversarial model. In such a model, points can be added or removed arbitrarily, provided that the adversary does not have access to the random choices of our algorithm. The amortized cost of our algorithm is poly-logarithmic when the ratio between the maximum and minimum distance between any two points in input is bounded by a polynomial, while k and ϵ are constant. Furthermore, we significantly improve the memory requirement of our fully dynamic algorithm, although at the cost of a worse approximation ratio of $4 +\epsilon$ . Our theoretical results are complemented with an extensive experimental evaluation on dynamic data from Twitter, Flickr, as well as trajectory data, demonstrating the effectiveness of our approach