
I'm currently a secondyear computer science PhD student at Stanford University, advised by Jure Leskovec. My research generally focuses on developing machine learning algorithms applied to graphstructured data.
I have worked on developing generalized graph convolutional networks that are scalable to webscale datasets, with applications in recommender systems, anomaly dection and biology. I'm also interested in deep generative models for graphs and multihop knowledge graph reasoning. I graduated from Duke University in 2016 with the highest distinction. I majored in computer science and mathematics. I did research in geometric algorithms, computational topology, numerical analysis and signal processing. Email: rexying AT stanford.edu 
News
 May 2018: Gave a poster presentation at SDSI workshop.
 May 2018: GraphRNN code is made public at https://github.com/JiaxuanYou/graphgeneration . Feel free to make issues if there's any question:)
 Looking forward to an internship in Facebook AI Research, NYC in summer 2018!
 April 2018: Gave a tutorial "Representation Learning on Networks" at WWW 2018 with Jure Leskovec, William L. Hamilton and Rok Sosic.
Website: http://snap.stanford.edu/proj/embeddingswww/  Dec 2017: One maintrack poster presentation at NIPS 2017 and an invited talk at the NIPS MLTrain Workshop on GraphSAGE.
PhD Research
Graph Representation Learning
When applying machine learning techniques to graphstructured data, an important step is to map nodes in the graph to dense vector representations (embeddings) that describe the neighborhood structure of the nodes, so that machine learning algorithms can take these node embeddings as input features. This is analogous to word embeddings in NLP, where words are mapped to embeddings and used for various downstream tasks.
My work primarily focuses on graph neural networks (GNN) architectures for learning graph and node embeddings, in the following aspects
Common applications include machine learning for social networks, recommender systems, biological networks, knowledge graph, physical systems and molecular chemistry.
My work primarily focuses on graph neural networks (GNN) architectures for learning graph and node embeddings, in the following aspects
 We can leverage embeddings in inductive setting, scalable to extremely large graphs with billions of nodes (GraphSAGE, PinSage).
 Effectively leverage hierarchical structures in graphs for learning both graphlevel (DiffPool) and nodelevel (HGCN) representations.
 Build generative models (GraphRNN, GCPN) for graphs, that allows us to model distributions of graphs, or generate graphs with optimal properties.
Common applications include machine learning for social networks, recommender systems, biological networks, knowledge graph, physical systems and molecular chemistry.
Inductive Representation Learning on Large GraphsNIPS 2017
William L. Hamilton*, Rex Ying*, Jure Leskovec

Representation Learning on Graphs:

Rotations
 In spring 2016, I rotated with Professor Greg Valiant, working on efficient approximate maximum inner product search.
 In fall 2016, I rotated with Professor Leo Guibas, working on inferring driving state (velocity, acceleration, steering angle etc.) based on cars' front camera videos.
Undergraduate Research
My undergrad research topics include geometric algorithms, numerical analysis and signal processing. I also did several other projects in image processing, computational topology during my senior year. During my summer internships at Google Ads, I worked on using machine learning to detect API queries using malicious script and debiasing user rating.
Here are some of my researches at Duke.
Here are some of my researches at Duke.
A Simple Efficient Approximation Algorithm for Dynamic Time WarpingACM SIGSPATIAL 2016
Rex Ying, Jiangwei Pan, Kyle Fox, Pankaj K. Agarwal
Dynamic time warping (DTW) is a widely used curve similarity measure. We present a simple and efficient (1 + eps)
approximation algorithm for DTW between a pair of point sequences. Although an algorithm with similar asymptotic time complexity was presented in our previous paper, our algorithm is considerably simpler and more efficient in practice. Our experiments on both synthetic and realworld data sets (GeoLife trajectories) show that it is an order of magnitude faster than the standard exact DP algorithm on point sequences of length 5,000 or more while keeping the approximation error within 5–10%. For input size of 50,000, our algorithm is more than 200 times faster.

Approximating Dynamic Time Warping

Paper.pdf  
File Size:  549 kb 
File Type: 
EMGJournal of Biomechanics 2016Rex Ying, Christine E. Wall
We present a new statistical method and MATLAB script (EMGExtractor) that includes an adaptive algorithm to discriminate noise regions from electromyography(EMG) that improves the labeling accuracy.

Locally Adaptive Spline InterpolationRex Ying, Xiaobai Sun, AlexandrosStavros Iliopoulos
In this paper, I give a framework and an efficient algorithm for constructing a high order, nonstandard spline from low order splines.
Using this technique, we provide an improvement over the traditional spline technique such that the result can be more efficiently adaptive to local updates, resilient to outliers, naturally parallel and has less error propagation, while still preserving the smoothness conditions of the spline. An inhomogeneous spline refers to a spline whose piecewise functions do not have the same order. 
Create a free web site with Weebly