Research Interest:
Network Data & Graphical Models
Multivariate Analysis & Machine Learning
Time Series
Bayesian Theory
Computational Social Science
Analyzing HR Data using Machine Learning
Park YJ, Choi BY, Woo YJ, Yoo SY, and Kim SR
Abstract
Since the initiation of Google's PiLab (People & Innovation Lab) in 2006, the widespread adoption of data-driven HR (Human Resources) and organizational management systems has aimed to bolster corporate profitability through the utilization of diverse candidate data.
In this study, we collected HR data from 14,999 employees on Kaggle and applied various machine learning algorithms, including Logistic Regression, Decision Trees, Support Vector Machines, Random Forests, and Artificial Neural Networks. Our objective was to identify the key factors influencing turnover. To gauge model performance from a company's perspective, we introduced a novel metric named "Total Gain," calculated as the difference between assigned "Profit" and "Loss." The additional metrics, including PPV, FNR, TPR, etc., provide a more nuanced understanding of the model's performance. Our results revealed that the Random Forest model consistently outperformed other models, highlighting "employee satisfaction" as the most influential variable affecting turnover.
Our final presentation
Throughout the Multivariate & Big Data Analysis Practice (Capstone Design) course, we regularly presented our findings, garnering positive evaluations from Dr. Kim (my recommender). Notably, our work was incorporated into the curriculum for the subsequent semester, emphasizing the significance and applicability of our research in the academic domain..
Codebook
Multivariate Analysis of Diabetes in 2015
Park YJ
Abstract
This study delves into the potential correlation between the wealth gap, education level, and diabetes beyond lifestyle habits and health conditions. Survey data, consisting of 22 distinct categories, was collected from 253,680 Americans through the BRFSS (Behavioral Risk Factor Surveillance System). Using R, I employed Logistic Regression, Decision Tree, and Artificial Neural Networks to establish predictive models for diabetes.
In the Decision Tree analysis, an initial model considering all variables resulted in misclassifying all respondents as not having diabetes. However, a revised model incorporating only the top five important variables demonstrated significant improvement, as confirmed by a high AUC (Area Under the Curve) value.
Conversely, the Artificial Neural Network analysis, examining randomly sampled reduced data with both all variables and important ones, did not produce accurate classification, prompting the exclusion of this modeling approach.
Throughout all machine learning models, the variable "self-assessed health score" consistently emerged as a crucial factor. Notably, a discernable relationship was observed, wherein higher income among respondents correlated with elevated self-assessed health scores.
I concluded that the onset of diabetes reflects a complex interplay of physical factors, such as blood pressure, cholesterol, and socioeconomic conditions. This finding underscores the importance of considering both health and socio-economic determinants in understanding and addressing diabetes.
Codebook
Predicting the Index Trend of 'KOSPI' through Time Series Analysis
Kim SY and Park YJ
Abstract
In 2020, the South Korean domestic stock market underwent pronounced fluctuations driven by the COVID-19 pandemic, resulting in significant amplitude in KOSPI (Korean Composite Stock Price Index). Existing studies suggest that the search volume for specific keywords is instrumental in reflecting public attention and participation, particularly in the economic field, showcasing a correlation between increased search volume and heightened interest and engagement in the stock market. This research aims to forecast future patterns of public attention.
Daily data collection from NAVER, South Korea's premier search engine, spanning from 2016 to 2021, formed the basis of our study. Utilizing ARIMA (Autoregressive Integrated Moving Average) models, we identified suitable models and estimated parameters, ultimately selecting the ARIMA (2,1,1) model as the most fitting. The model predicted a slight decline in KOSPI search volume over the next 10 weeks, suggesting a trend of increased interest in stock investments in Korea since 2020, followed by a gradual decrease in 2021 due to economic recovery. We observed that our predictive model aligned with the real-world data after 10 weeks.