I. Political Ideology Prediction Through Twitter Data (2020-2021)

Working with Prof. Christopher Bail and Prof. Alex Volfovsky at Polarization Lab, we predicted political ideology of Twitter users by extracting information from followee and follower relationships on Twitter through Laplacian Eigenmap.

We implemented and hosted our algorithm on a public server to provide API service. You can use this front-end page here to try and test your political ideology. Don’t hesitate to tell us if you think our prediction is accurate (or horribly incorrect)!

We also have another ideology predictor based on your tweets. Feel free to try it and tell us how you like it!

II. Graph Neural Networks (GNNs) and Their Applications Review/Presentation (2020)

I have gained tremendous interest in machine learning on network-structured data from the above project. So, you could imagine how excited I was when I found Prof. Jure Leskovec’s lectures about machine learning on graphs. Of all the topics I was most fascinated by the graph neural networks (GNNs). They extend deep learning to Non-Euclidean data (i.e. network data) to tackle all sorts of graph-related problems including node classification, graph classification, link prediction, etc.

So I chose GNN to be the topic of my project for deep learning class, and reviewed its fundamentals and applications. The readers are assumed to be familiar with machine learning and deep learning but interested about basics of GNN and its real-life applications. I have also prepared a one-hour tutorial (as there is no better to learn than to teach) and presented it. You can find my final review paper and slides in my Github repository.

I have used GNN to improve performance of the model in the Political Ideology Prediction project. My interest only gets stronger after this experience. Looking forward to conducting researches on this topic during my Ph.D. journal!

III. Interactive Spectrum Fitting Software (2021)

If I hadn’t found my passion in machine learning and research, I believe I would chase a career in front-end SDE. I have spent quite some time in undergraduate to study assorted front-end tech stacks. And later I found them to be handy when you in fast need of some interactive yet cool-looking tools. Here is one that I recently built.

For a meta-material, we can measure its spectrum T. But what we are really interested, given T, is its permittivity EPSILON and its permeability MU. EPSILON and MU are both complex numbers modeled by Lorentzian functions. So, in essense, if we pick the correct Lorentzian parameters for EPSILON and MU, we can have the correct EPSILON and MU that give us the observed T.

There is an analytical way to go from EPSILON/MU to T. But the process includes an optimization problem that is very sensitive on the initialization of the Lorentzian parameters. It would be very nice if we can actually see how the current guess of Lorentzian parameters fit and interact with the parameters to manually pick a good initialization.

If you find this app interesting, you can have the source code in my Github repository.