Junseok Yang

17-18 EPL Team Attack Pattern Analysis

STAT 430 - Practice of Applied Statistics

(Report) (Code)

• • Constructed attacking sequences from spatio-temporal football data, utilizing event types and spatial coordinates

  • Applied unsupervised learning using Frechet and Gower distances with K-medoids clustering to define latent tactical patterns and verify team-level associations

  • Discovered that Big 6 teams sharing structurally similar attacking clusters with tactical variations while Leicester City uniquely relying on fast, efficient duel-based sequences, achieving the highest goal output outside the Big 6

Political Polarization

LING 413 - Corpus Linguistics

(Report)

• • Worked on a project analyzing user data provided by a non-profit organization ‘Braver Angels’ supervised by professor Tori Ellison and Jonathan Dunn to find potential key terms of political polarization in the U.S.

  • Implemented Topic Modeling (LDA) by preprocessing raw text and tuning hyperparameters to extract unique and distinguishing terms across diverse topics

  • Performed document clustering to identify user behavioral patterns by correlating topic interests with demographic features like profession and joining date

Fall 2023

STAT 426

Statistical Modeling II (R)

---

STAT 430

Time Series ML (Python)

Analysis & Forecast of Chinese Stock Market

(Report)

• • Analyzed 500 Chinese stocks price data with diverse time-series perspectives (Classic ARIMA / Clustering / Deep Learning GRU)

  • Worked on cluster analysis part by applying algorithms like DBSCAN to group stocks with similar price trends and found real estate / Food / Medicine showing notable growth and stabled fluctuation

---

STAT 542

Statistical Learning (Python)

Spring 2024

STAT 431

Applied Bayesian Analysis (R)

STAT 510

Mathematical Statistics

---

STAT 530

Bioinformatics (R)

Partial Dopaminergic & Tyrosine Hydroxylase of Dopamine Synthesis

(Report)

• • Used MERFISH dataset to figure out the existence of partial dopaminergic neurons in preoptic area (POA) and spatial evidence of tyrosine hydroxylase (TH) and L-DOPA exocytosis for dopamine synthesis

  • Implemented dimensionality reduction algorithms such as UMAP and found out a spatial structure and molecular features of partial dopaminergic cells within POA that are potentially interacted with dopamine synthesis

  • Conducted spatial autocorrelation to test evidence of spatial location difference between TH/Ddc (dopadecarbosylase) groups and detected a notable pattern of L-DOPA exocytosis from concentrated to dispersed

Fall 2024

STAT 430

Practice of Applied Statistics (R)

STAT 571

Multivariate Analysis (R)

Spring 2025

LING 413

Corpus Linguistics (Python)

PHIL 460

Philosophy of Statistics

Fall 2020

STAT 107

Data Science Discovery (Python)

STAT 400

Statistics & Probability I

MATH 415

Applied Linear Algebra

Spring 2021

STAT 212

Biostatistics (R)

STAT 410

Statistics & Probability II

STAT 420

Methods of Applied Statistics (R)

Fall 2021

STAT 207

Data Science Exploration (Python)

STAT 425

Statistical Modeling I (R)

---

STAT 430

Unsupervised Learning (Python)

Market Segmentation of Credit Card Customers

(Report & Code) (Presentation)

• • Explored a bank data from Kaggle to discover notable insights of customers' leaving status based on their demographics and banking information using Python

  • Applied diverse algorithms (Hierarchical Agglomerative, K-Prototype Clustering) from Scikit-learn to visualize a clustering structure and set an assumption of customers with lower and unstable income to be less economically active and likely to leave referring to the structure and exploratory data analysis

  • Computed supervised evaluation scores (Complete/Homogeneity/V-Score) of the algorithms' predicted clustering label created based on the assumption to compare with the pre-assigned label and identified poor match between the predicted and original labels (weak or no association between customers' info and their leaving status)

Spring 2022

STAT 385

Statistical Programming Methods (R)

European Football League Player Stats (Revamping)

(Streamlit) (Shiny App / Old Version)

• • Developing an interactive Streamlit web app using Python to analyze and visualize major European football league (Big 5 & Eredivisie/Primeira Liga/Jupiler) players' performance data from Fbref, building on a prior R Shiny version

  • Integrating data scraping (BeautifulSoup), radar chart visualization (Plotly), and player comparison features using cosine similarity to allow dynamic selection and comparison of similar players based on diverse in-game performance stats

---

STAT 443

Professional Statistics (R)

Consulting Project: Air Pollutant in Gucheng, China

(Report) (Presentation)

• • Consulted with a mock client about a project using an air pollutant dataset from Gucheng, China using R

  • Visualized diverse time-series-related plots with ggplot and observed relatively higher air pollutant (PM2.5) level in winter, weekend, and nighttime

  • Tested various algorithms (Linear Regression / Multinomial / K-Nearest Neighbors) to find the best model which predicts a next day's air pollutant (PM2.5) level based on its previous day's meteorological variables values

---

CS 307

Modeling & Learning in Data Science (Python)

Denoise Autoencoder

(Report)

• • Built autoencoders with 'FashionMNIST' data in PyTorch to observe a relationship between noisy data and the autoencoders' performances

  • Trained the autoencoders using manipulated data with three different noise type (No noise / Gaussian / RandomErasing)

  • Evaluated the autoencoders' performances by comparing their average train & test loss and visualizing t-SNE plots of latent spaces

Fall 2022

STAT 430

Fundamental of Deep Learning (Python)

Lyrics Generator

(Report)(Presentation) (Github)

• • Built a neural network structure (LSTM) with PyTorch to generate lyrics with a dataset containing 6 million songs from Kaggle

  • Trained and tested the model with different hyperparameters (Sequence length / Batch size) by genres and discovered longer sequence length not necessarily guarantee better model performances

  • Generated sample lyrics and observed Rap to be relatively short and negative whereas Country and R&B to be long and romantic

---

STAT 432

Basics of Statistical Learning (R)

Beijing Housing Prices

(Report) (Presentation)

• • Conducted both supervised and unsupervised analysis with a Beijing housing price dataset from Kaggle to understand price trends in Beijing, China

  • Applied diverse regression algorithms such as LASSO and XGBoost to predict 'price' and figured out Number of rooms / Age & Condition / Distance from an epicenter to be the most significant predictors

  • Clustered the data using K-Means, grouped the observations as 'over'/'under' priced within their clusters and discovered the trend of 'overpriced' properties to be relatively old and close to the epicenter

---

STAT 448

Advanced Data Analysis (SAS)

LING 402

Tools & Tech Speech & Language Process (Python, Linux (Bash))

Spring 2023

STAT 430

Baseball Analytics (R)

Pitching Similarity

(Shiny App)

• • Created an R Shiny app with MLB Statcast data

  • Designed pitching location heatmaps with ggplot to visualize and compare selected pitchers' pitching location and diverse stats such as pitch type and speed

---

STAT 440

Statistical Data Management (R, SQL)

CS 277

Algorithms & Data Structures for Data Science (Python)