Machine Learning

Feb 2020 – Apr 2020
Discovered a better embedding method for nodes from a large-scale, real-world Knowledge Graph named YAGO, and completed various downstream tasks such as link prediction and node classification using PyTorch framework.
Initialized nodes with TransE embedding, sampled balanced subgraph based on edge-type and neighborhood, and fed subgraphs into mini batches to train a Relational Graph Convolutional Network (RGCN) with embedding output.

Apr 2019 - June 2019
Developed an automatic segmentation solution for brain MRI to keep the essential tissue with deep learning models.
Employed OSVOS model with Tensorflow using fully convolutional neural network (FCN): pre-trained parent network for basic foreground segmentation; fine-tuned the network on ground truth image pairs to minimize pixel-wise cross entropy loss.
Used less than 20 sets of segmented MRI, and got over 90% accuracy in pixel-wise comparison on testing sets.

Jan 2019 - Mar 2019
Built personalized recommendation system using user movie ratings with multiple machine learning methods in Python.
After experiments on different model combinations, used ensemble of two best models to predict user rating: SVD model on user movie pairs’ ratings and user-based Linear Regression with movie tags information.
Achieved root MSE of 0.822 on over 4,000,000 user-movie pairs in final rating prediction.

Apr 2018 - June 2018
Aggregated people’s attitudes towards the two Parties and Donald Trump by NLP on Reddit posts and comments.
Fit tokenized and lemmatized sentences from Reddit text into Machine Learning model (Logistic Regression) in Python, which learns to label sentiments of positive/negative towards two parties and Donald Trump.
Combined queries to MySQL database, and visualized clear political sentiments fluctuation over states in time series graph with R.