We design a hierarchical graph contraction data structure that maintains the cluster structure of an evolving graph with much faster query time than previous dynamic spectral clustering algorithms.

We introduce an algorithm that can dynamically maintain a sparse cluster-preserving approximation of a dense kernel/similarity graph. A key building block for this algorithm is a dynamic kernel density estimation algorithm.

We introduce an algorithm that can dynamically maintain the clusters of a graph as new edges and vertices are added.

We present algorithms for constructing nearly-optimal hierarchical clustering trees with respect to Dasgupta’s cost function on well-clustered graphs.

We introduce a neural network pruning algorithm using spectral graph sparsification, empirically demonstrating its effectiveness in preserving performance at higher sparsity levels and providing theoretical analyses of local and global network connectivity.

We develop simple theory and use experiments to show that the episodic training setup for Matching Networks and Prototypical Networks is inefficient.

We show that a path-like cluster structure can be found in directed graphs, by performing spectral clustering on the first eigenvector of a newly defined Hermitian adjacency matrix.

As one of the first groups worldwide to do so, we use drones equipped with infrared and RGB cameras to map geothermal surface manifestations.

Using theory from sociology, we develop bodily cues in virtual agents which effectively communicate social engagement.