Haishan Gao

Located in Mountain View, CA

Hi, my name is Haishan. I am a Software Engineer at Google Shopping developing ML infrastructure and information retrieval algorithms.

I finished my undergraduate study at Northwestern University with a double major in Computer Science and Economics. I was fortunate to work in the QRG lab under the supervision of Constantine Nakos and Prof. Ken Forbus on natural language understanding and question answering. I also worked with Prof. Özge Samancı on interactive art.

My research interest lies at the intersection of NLP, Computational Social Science and Interactive Media. My past research experience was focused on building interactive applications and using computational methods to analyze social issues.

I like reading and writing fiction outside of work.

selected projects

Replicated DNS Cache Service backed by Zookeeper-based Membership Management

2022

Abs PDF Code

We design and implement a distributed DNS cache service with replicated cache in a hierarchical structure. Our system demonstrates high availability, horizontal scalability, and robustness to withstand server node failures. It supports instant server membership management to add and remove nodes dynamically. We propose a tailored design of load balancing strategy and consensus protocol in our system that relaxes the consistency requirement to achieve high availability and partition tolerance. The system is deployed and fully tested on Amazon EC2. We measure its latency, throughput and ability to tolerate node failures and validate it outperforms Google DNS service (8.8.8.8) on the latency metric in a local area network environment.
Sales Prediction Based on Product Titles and Images with Deep Learning Approaches

2021

Abs PDF Code

Over the past few years, online shopping gradually became the mainstream shopping method. More and more local retailers chose to start their businesses on e-commerce platforms. However, few can survive due to the competitive pressure from the big companies and the entry barriers. The motivation of this project is to identify product listing strategies, primarily visual and textual presentation, that can help retailers to raise their product sales. To achieve that, we build a neural network architecture to predict product sales from text, image and other product listing features that retailers can control. In particular, we use pre-trained VGG16 model for image data and TF-IDF model for text data in our architecture. The accuracy of the model is 73.91%, far above the accuracy of human raters, which is 55.33%.
Political Polarization, Emotions and Engagement on Reddit

2021

Abs PDF

Trained transformer based NLP models to analyze the relationship between user engagement, political polarization and emotion intensity on Reddit to understand the echo chamber phenomenon of online forums.
Portal Snake (Unity 2D Cooperative Game)

2019

HTML