**Introduction to Professor Siheng Chen and His Lab at Shanghai Jiao Tong University**

**1. Could you briefly introduce yourself (and your University/Lab)?**

Siheng Chen is a tenure-track associate professor at Shanghai Jiao Tong University and a PI at Shanghai Artificial Intelligence Laboratory. Before that, he was a research scientist at Mitsubishi Electric Research Laboratories (MERL) and an autonomy engineer at Uber Advanced Technologies Group, working on the perception and prediction systems of self-driving cars. Before that, he was a postdoctoral research associate at Carnegie Mellon University. He received the doctorate in Electrical and Computer Engineering from Carnegie Mellon University in 2016, with the thesis title “Data science with graphs: A signal processing”. He also received two masters degrees from Carnegie Mellon University in Electrical and Computer Engineering and Machine Learning, respectively. He received his bachelor’s degree in Electronics Engineering in 2011 from Beijing Institute of Technology, China. His paper “Diagnosis algorithms for indirect structural health monitoring of a bridge model via dimensionality reduction” published in Mechanical Systems and Signal Processing received ASME SHM/NDE 2020 Best Journal Paper Runner-Up Award. His paper “Discrete signal processing on graphs: Sampling theory” published in IEEE Transactions on Signal Processing won the 2018 IEEE Signal Processing Society Young Author Best Paper Award. His coauthored paper received the Best Student Paper Award at IEEE GlobalSIP 2018.

Siheng’s research mainly focuses on graph and collective intelligence, whose goal is to develop advanced graph-based theories and tools to analyze and optimize the collective efforts of multi-agent systems. The core techniques are rooted in graph signal processing and graph neural networks; and the applications range from autonomous driving, human behavior analysis, traffic monitoring to smart healthcare.

**2. What have been your most significant research contributions up to now?**

My most significant research contribution is probably a sampling theory of graph signals, or called graph sampling theory. Sampling is the process of collecting some data when collecting it all or analyzing it all would be unreasonable; and it is a key technique to handling massive data. As a generalization to the graph domain, graph sampling considers preserving the information of a graph signal by using its subsamples. The classical Nyquist-Shannon sampling theory instructs us how to capture information of a time-series signal by measuring signal values at a few time stamps. Graph sampling theory is a generalization of the classical Nyquist-Shannon sampling theory from the 1D time-series domain to the irregular graph domain. Graph sampling theory reveals the most influential nodes in an irregular graph. This theory can be used to active training sample selections for semi-supervised learning, active resampling of 3D point clouds and active monitoring of urban traffic.

From a theoretical aspect, we have achieved three fundamental theoretical findings: (1) perfect recovery is possible for graph signals bandlimited under the graph Fourier transform; (2) the sampled signal coefficients form a new graph signal, whose corresponding graph structure preserves the first-order difference of the original graph signal; (3) for approximately band-limited graph signals, experimentally designed sampling and active sampling have the same fundamental limitations, and can outperform uniform sampling on irregular graphs.

From a algorithmic aspect, we proposed three sampling algorithms: (1) a deterministic optimal sampling operator for bandlimited graph signals, which guarantees perfect recovery and robustness to noise; (2) a randomized sampling strategy for approximated bandlimited graph signals, which achieves the optimal convergence rate; (3) an end-to-end trainable graph pooling operator, called vertex infomax pooling (VIPool), which adopts the latest self-supervised learning techniques to select the most informative vertices.

From an application aspect, we applied the proposed graph sampling theory to three scenarios: (1) social network inferences. We query attributes from a few representative users and infer the attributes of all the other users; (2) 3D point cloud compression. We select a few representative 3D points to model the overall distribution of all 3D points; (3) urban traffic monitoring. We monitor the traffic conditions at a few intersections and infer the traffic conditions at all the other intersections.

**3. What problems in your research field deserve more attention (or what problems will you like to solve) in the next few years, and why?**

The problem I Would like to solve is to bridge graph signal processing and graph neural networks through scattering and unrolling techniques. As the extension of neural networks to the graph domain, graph neural networks have also received a lot of attention and achieved significant success in social network analysis and geometric data analysis. However, these neural network architectures are typically designed through trial and error. It is thus hard to explain the design rationale and further improve the architecture. I believe it is possible to use graph signal processing theory to analyze and guide the design of graph neural networks. Two promising technical approaches are graph scattering and graph unrolling.

Graph scattering are non-trainable graph convolutional networks comprising a cascade of graph

filter banks followed by nonlinear activation functions. The graph filter banks are mathematically

designed and are adopted to scatter an input graph signal into multiple channels. Graph scattering extracts scattering features that can be utilized towards graph learning tasks. This technique enables the generalization and stability analyses of graph neural networks.

Graph unrolling provides an interpretation of the design of graph neural network architecture from a signal processing perspective. It unrolls an iterative graph-based algorithm by mapping each iteration

into a single graph convolutional layer where the feed-forward process is equivalent to iteratively solving a graph-based optimization.

**4. What advice would you like to give to the young generation of researchers/engineers?**

(1) Solve real challenges, instead of just producing papers.

(2) Be open-minded to learn something new and different.

(3) Be persistent. It is the courage to continue that counts.

(4) Be optimistic. Research has too much uncertainty, get used to it.