Sophia He's Data Blog

Portfolio

Contact: sophiaheli2013@gmail.com

Served as resident data expert in programmatically transforming data and delivering reproducible solutions in Python/R/SQL
Built 40+ automated Tableau dashboards using the jail, courts, and health care data that improved work efficiency and decision-making process for criminal justice partners. Achieved 400+ dashboard views per month.
Implemented statistical and machine learning methods, such as time series analysis, propensity score matching, and logistic regression, on jail population prediction, program evaluation, and pretrial risk assessment.
Contributed to the department receiving $2M MacArthur Criminal Justice Reform Grant to safely reduce the jail population by developing the jail holding status logic and analyzing the racial disparities across the system.

Developed a maintainable and reusable classification model in Python that included data cleaning, feature selection, random forest pipeline, and command line user interface; achieved F1 scores of 0.8, 0.4, and 0.8 for three labels.

Cleaned disorganized survey data provided by Arvato Financial Solutions and transformed features programmatically.
Conducted k-means clustering based on the result from principal components analysis. Identified the characteristics of target customers by interpreting components in each cluster.

The goal of this project is to accurately predict whether an individual makes more than $50,000 using data collected from the 1994 U.S. Census. This sort of task can arise in a non-profit setting, where organizations survive on donations. Understanding an individual's income can help a non-profit better understand how large of a donation to request, or whether or not they should reach out to begin with.
Among three algorithms Random Forest, SVC, and AdaBoost, the final model chosen in is AdaBoost Classifier algorithm, which is short for Adaptive Boosting. AdaBoost combines multiple “weak classifiers” into a single “strong classifier”.
On the testing set (the rest 20%), the model predicted the outcome (Y/N) based on the classification model we generated using the training set. It yields 0.72 F$_{0.5}$ score and 0.86 accuracy score which is the highest among the three algorithms.