- Automated Pipeline Selection: Developing a mechanism to combine primitives, provided by Data-Driven Discovery of Models (D3M) library, automatically into pipelines on given machine learning datasets. The framework is developed in Docker, supported by MongoDB and paralleled by Ray.- Primitives Development: Abstract general pattern in feature engineering and implement them as reusable primitives.

- Online News Recommender System: Developed an unsupervised algorithm which can (1) handle 120 million dimensional feature in real time; (2) extract valuable information from unlabeled text, news figures, and user personas. The system is currently supporting 10 millions level of page-views per day.- Multimodal Learning framework: Developed a novel multimodal learning framework to fuse heterogeneous information from text, images, audios, etc. The framework, implemented by Tensorflow, is 100 times more efficient than Deep Canonical Correlation Analysis (Deep CCA), and can be steadily optimized and generalized.- Graph Embeddings: Implemented Structural Deep Network Embeddings (SDNE), Graph Convolutional Networks (GCN) and GraphSAGE by Pytorch and apply them to social networks and knowledge graphs. Proposed a novel loss function to replace the negative sampling strategy which reduces the complexity in optimization.

- Teaching Assistant for Learning from Data on Machine Learning topics. - Designed and graded the coding and written assignments