Few-shot Video Action Recognition System
Mentors: Biplob Debnath, Oliver Po, Srimat Chakradhar, Asim Kadav, Farley Lai

• Worked on developing a few-shot video action recognition system that leverages temporal information to improve recognition performance for longer, sequential activity types.
• One paper under submission and one patent application pending.
  

Characterizing and Modeling Temporal Dynamics of User Behavior on Social Platforms
Mentors: Maarten W. Bos, Yozen Liu, Neil Shah, Leonardo Neves

• Worked in the Computational Social Science research team to study and understand how user behavior on social platforms is driven by regularities and momentary cognition & feelings. The work is aimed towards better and privacy-preserving modeling of user behavior for dynamic and personalized content recommendation.
• Published one paper in ICWSM, one paper in CHI and two patent applications pending.
  

Characterization and Detection of Malicious User Behavior in Social Media

• Developed a Twitter data crawler; streamlined data collection, filtering, and storage process.

• Crawled 560M Twitter user info & 300M Tweets and performed analysis (tweet content, user info & activity pattern) to characterize and model malicious vs. non-malicious users.(manuscript under review)

• Currently developing a real-time, adaptive and scalable algorithm for malicious users and coordinated malicious activity detection on social media.

A Real-time Twitter Analytics Dashboard

• Functionality: Hashtag & User Activity Tracking, Identifying frequent Word/Hashtag/URL & Influential users, Tweet filtering & classification (i.e. sentiment, intent, spam), Info-graphic visualization.

• Tools & Framework: Flask, Django, PostgreSQL, Elasticsearch, Docker, Heroku, AWS, Chart.js.

Seismic Phase Classification for Automated Monitoring

• Developed a deep-learning model for fine-grained phase classification from single-station seismic time-series. The model utilizes a Convolutional Neural Network and Long-Short Term Memory Network in combinationto exploit both short- and long-term temporal patterns in seismic data to improve classification accuracy.

Event and Anomaly Detection in Pressure Sensor Data

• Developed a semi-supervised algorithm using Singular Spectrum Analysis for Structured Noise detection in Oil/Gas well pressure data to automate Pressure Transient Analysis; created a user interface.
  

Conducted Theory and Lab Courses of Undergraduate Computer Science students

Developed a novel algorithm for automatic breast lesion segmentation from B-mode ultrasound images