Data Science and AI
1. Deep Q-learning with Model-based Exploration: Efficient Learning on Environments with Sparse Rewards
[paper]
In this paper, we proposed an improved DQN algorithm that utilizes learned environment dynamics to guide exploration. We demonstrated that it outperformed the original DQN on the classic environment with sparse rewards, Mountain Car. Our algorithm was able to explore a wider range of states, and increased the learning speed.
[paper]
In this paper, we proposed an improved DQN algorithm that utilizes learned environment dynamics to guide exploration. We demonstrated that it outperformed the original DQN on the classic environment with sparse rewards, Mountain Car. Our algorithm was able to explore a wider range of states, and increased the learning speed.
2. Predicting Crashing Patients with Informal Clinical Notes by Ensemble of Neural Document Embeddings
[paper]
In this collaboration with St. Michael's Hospital's analytics team (LKS-CHART), we developed an ensemble of deep models to generate document level embeddings of unstructured clinical notes. We showed that classifiers trained with our embedding performed better than doc2vec, LDA and fine-tuned BERT models in the task of predicting patients' outcome (will crash or not).
[paper]
In this collaboration with St. Michael's Hospital's analytics team (LKS-CHART), we developed an ensemble of deep models to generate document level embeddings of unstructured clinical notes. We showed that classifiers trained with our embedding performed better than doc2vec, LDA and fine-tuned BERT models in the task of predicting patients' outcome (will crash or not).
3. Predictive Modelling of Revenues of Modern American Movies
[jupyter notebook] [report]
This project builds a model that predicts a movie's total revenue, given certain traits and facts about the movie. It aims to provide an effective prediction as soon as the movies are released, which means that data like opening weekend box office, IMDb rating, and social media sentiments cannot be used as features in the models.
[jupyter notebook] [report]
This project builds a model that predicts a movie's total revenue, given certain traits and facts about the movie. It aims to provide an effective prediction as soon as the movies are released, which means that data like opening weekend box office, IMDb rating, and social media sentiments cannot be used as features in the models.
4. MOOClet Engine for Dynamic AB-Testing (Contributor)
[Repository]
The MOOClet Framework is a web engine, which supports wide range of dynamic AB experimentations. I implemented a policy for dynamic experiments with contextual user informations through a contextual multi-armed bandit model using Thompson Sampling.
[Repository]
The MOOClet Framework is a web engine, which supports wide range of dynamic AB experimentations. I implemented a policy for dynamic experiments with contextual user informations through a contextual multi-armed bandit model using Thompson Sampling.
5. Understanding and Correcting Inaccurate Calorie Estimations on Amazon Mechanical Turk
[Accepted SIG-CHI 2019]
We conducted a study on interventions to improve people's knowledge of food calories through experiments using interactive Qualtrics survey on Amazon Mechanical Turk workers.
[Accepted SIG-CHI 2019]
We conducted a study on interventions to improve people's knowledge of food calories through experiments using interactive Qualtrics survey on Amazon Mechanical Turk workers.
6. Analysis of the Use of Tobacco Products in Subgroups of American School Children
[Report]
Test hypotheses regarding the use of tobacco products among American school children and estimate effects of demographic factors using generalized linear models.
[Report]
Test hypotheses regarding the use of tobacco products among American school children and estimate effects of demographic factors using generalized linear models.
7. Dynamics of a Modified SIR Model
[paper]
In this paper, we proposed a modified SIR (susceptible, infected, recovered) model for modelling the spread of infectious disease. We introduced population birth, death and existence of temporary immunity into the original model.
[paper]
In this paper, we proposed a modified SIR (susceptible, infected, recovered) model for modelling the spread of infectious disease. We introduced population birth, death and existence of temporary immunity into the original model.
