Projects

Besides the projects in the Publications section, here are some more projects that I have conducted.

1. Automatic Differentiation Python Library

We built a software performing automatic differentiation (AD) for the user. The AD software, SwiftDiff, sequentially evaluates elementary functions, and avoids the complexity of symbolic differentiation and precision issues of numerical differentiation. The system implements multiple methods of AD that compute the derivatives of a function in a single flow with machine precision and accuracy.

2. Mortality Prediction and Interpretation

We used National Health And Nutrition Examination Survey I Epidemiologic Follow-up Study (NHEFS) dataset to predict a person’s risk of dying. PCA was applied to deal with multicollinearity and reduce runtime and complexity for the later training. We performed KNN, Logistic Regression, Random Forest, AdaBoost, and Neural Network on data, and determined Random Forest to be the best model based on F1 score (0.803) and AUC (0.861). Feature importance was interpreted using SHAP value.

3. Does money buy happiness? What affects world happiness?

We performed several ordinary least squares models(OLS) and then mixed effect models to explore the relationship between happiness score and and other economic factors, social factors and etc collected by World happiness Report. We discovered that higher GDP values are positively correlated with higher happiness. Many other cultural, and social factors also have significant influence on happiness, such as percentage of people who have social support and the extent of satisfaction with the freedom to choose the lifestyle. Developed countries usually have positive relation between happiness score and health life expectancy, but some derdeveloping countries have negative relation between the two variables.

4. Ethical Analysis of Boston 311 Service: Observational Study

Like many cities in the U.S., Boston offers a 311 line to connect citizens with non-emergency city services. This study combined the 311 service request datasets with demographic data, such as income, education level, poverty rate, race distribution, poverty rate, to answer the question: Does Boston provide 311 services fairly and equally to all citizens? As a result, while we found it unlikely that there is any explicit bias in Boston’s 311 service provision (e.g. intentional prioritization of well-off neighborhoods or witholding of services from needy ones), we found discrepancies along demographic lines in the way the service is used, making it likely that neighborhoods do not receive equal treatment from city government through 311. Additionally, our analysis suggested that whiter and more well-off neighborhoods have a stronger preference for using the Citizen Connect App compared to their less well-off counterparts, which increased the likelihood that their cases are solved quickly and successfully. These factors, taken together, provided strong evidence for the existence of conditions leading to unequal outcomes along demographic lines, and a logical explanation for how it might occur.