Research for XDATA Program at JHU

Our group (PI Carey E. Priebe) has several XDATA related research efforts, with strong emphasis on statistical inference for graph data.

1. Scan Statistics

Applying the idea of “moving window'' analysis from time series to graph neighborhoods, local connectivity statistics (number of edges in a neighborhood) are computed. These are used to detect chatter anomalies in time series of graphs. Example results for CharityNet dataset are here.

2. Model Selection Using Non-negative Factorization for a Time Series of Graphs

We are interested in model selection techniques that can be used to augment existing non-negative factorization algorithms. Preliminary results based on a standard Non-negative Matrix Factorization algorithm are presented here.

3. Adjacency Spectral Embedding with Out-of-Sample Extension

By using out-of-sample extension for spectral embedding, it is possible to embed graphs where parts of the adjacency matrix might be missing, or time varying. Expository results for this method using simulated data are here.

4. Dissimilarity Representation for Time Series of Graphs and Clustering of their embeddings

In certain cases, instead of feature representation where each observation is a finite list of features, a dissimilarity representation for data is preferable. We have worked on a specific data setting where the observations are multivariate time series. Results for Akamai CIDR data can be found in the JHU presentations found in XDATA wiki.

5. Collaboration with Harvard

We are working with Edo Airoldi's team to extend R's igraph package by implementing the following functions:

adjacency spectral embedding: adjacency.spectral.embedding
seeded graph matching: sgm
hierarchical stochastic block model: hsbm.game
"elbow” finding: dimSelect
scan statistics (“us” & “them”): local.scan
design data structure for graphs that change in time (in progress)

6. Publications

The complete list is here!