Projects | Academic

Question Answering on Long Context

Language model pretraining has led to significant performance on question answering purpose but answering questions when the context is the whole document is challenging. Our project is to answer the question given an unseen wikipedia article. We also enhance model automation by finding out an appropriate confidence threshold.

Question Answering on Long Context

Portfolio Management Using Reinforcement Learning

In this project, we used DDPG with LSTM and CNN to resarch on the portfolio management problem and explore the impact of observation window size.

Portfolio Management Using Reinforcement Learning

Anomaly Detection for Web Traffic Data

In this project, we experimented with ARIMA, C-LSTM and CNN to perform anomaly detection on web traffic data. Experiments demonstrate that ARIMA performs differently on different types of anomalies as ARIMA focuses only on local data rather than a full picture. On the other hand, C-LSTM outperforms CNN, reaching a recall rate of 79.1%.

Anomaly Detection for Web Traffic Data

Self/Semi-Supervised Image Classification

Competed on finding the best Self/Semi-Supervised Learning (SSL) algorithm for a 512,000-images classification task with only 5% labelled from 800 classes. Researched and trained contrastive learning models (SimCLR, SimSiam, BarlowTwins), and pseudo labeling models (FixMatch, CoMatch) via PyTorch on Greene Cluster. Achieved 51% accuracy on the testing set and got 2nd place.

Self/Semi-Supervised Image Classification

Pyspark Song Recommender System

Trained alternating least squares (ALS) model via PySpark on the Million Song Dataset to get latent representation for users and songs. Used PCA to reduce dimension and visualize music genres.

Pyspark Song Recommender System

Aspect-Based Sentiment Analysis

Predicted aspect-level sentiment polarities on a restaurant reviews dataset (MAMS) that contains multiple aspects with multiple sentiment polarities. Implemented LSTM model and fine-tuned BERT based model. Achieved 84% accuracy on MAMS and performed error analysis.

Aspect-Based Sentiment Analysis

Predict number of views for videos on Bilibili

Bilibili is the most popular creator-driven content video play platform in China. To help vloggers better understand drivers of popular videos, we predicted the number of views for new videos based on video titles and labels, vlogger’s and previous video’s information using models include Ridge, Random Forest, XGBoost. We found features like video length, video posting frequency and time also play an important role in video popularity except previous video plays.

Predict number of views for videos on Bilibili

Bear Maps

With tile images and map feature data was downloaded, build a web mapping application using java. The data structures & alogrithms maily used are Graph, A*.

Bear Maps

Suggested Store Locations

Suggested new store locations using K-means and DBSCAN clustering with parsed geolocation data including public transport locations, E-commerce delivery addresses, competitors' store locations.

Suggested Store Locations