Master of Science in Data Science — Unsupervised Learning Methods
A project-based course exploring traditional and modern unsupervised learning techniques for dimensionality reduction, clustering, anomaly detection, and multivariate data visualization.
This course introduces a broad range of unsupervised learning methods applied to real-world datasets through extensive programming assignments and projects. Unlike supervised learning, unsupervised methods uncover hidden structure in data without labeled responses — making them essential tools for exploratory analysis, pattern recognition, and anomaly detection.
- Principal Component Analysis (PCA) — Represent relationships among many continuous variables by reducing to orthogonal components
- Factor Analysis — Model latent structure underlying observed variables
- Cluster Analysis — Identify natural groupings of individuals within data
- Block Clustering — Simultaneously cluster rows and columns (biclustering) to identify groups of variables
- Log-Linear Models — Explore relationships and dependencies among categorical variables
- Association Rules — Mine frequent patterns and co-occurrence rules (e.g., Apriori, FP-Growth)
- Lattice Displays — Visualize conditional relationships across multiple variables
- Multidimensional Scaling (MDS) — Embed high-dimensional data into lower-dimensional space preserving distances
- t-SNE — t-Distributed Stochastic Neighbor Embedding for non-linear dimensionality reduction and visualization
- Autoencoders — Detect anomalies by measuring reconstruction error from neural network encoders
- Probabilistic Deep Learning — Model uncertainty and detect outliers using probabilistic approaches
📦 msds411-unsupervised-learning/
├── 📁 assignments/
│ ├── assignment_01_pca_political/
│ ├── assignment_02_clustering_housing/
│ ├── assignment_03_unsupervised_pretraining_credit/
│ └── assignment_04_anomaly_detection_fraud/
├── 📁 data/
│ ├── raw/
├── requirements.txt
└── README.md
| Category | Tools |
|---|---|
| Language | R 4.3+ |
| Core Libraries | tidyverse, data.table, matrixStats |
| Machine Learning | caret, tidymodels, cluster, factoextra |
| Association Rules | arules, arulesViz |
| Deep Learning | keras, tensorflow |
| Visualization | ggplot2, lattice, plotly, ggfortify |
| Dimensionality Reduction | stats, Rtsne, umap, factoextra |
| Notebooks | R Markdown, Quarto |
| Environment Management | renv |
- R 4.3+
- RStudio (recommended)
- Git
# Clone the repository
git clone https://github.com/jep9731/academic-MSDS411-coursework.git
cd academic-MSDS411-coursework# Create and activate environment
install.packages("renv")
renv::restore()install.packages(c(
"tidyverse",
"data.table",
"cluster",
"factoextra",
"caret",
"tidymodels",
"arules",
"arulesViz",
"Rtsne",
"umap",
"keras",
"tensorflow",
"plotly",
"lattice"
))Methods: PCA, Feature Engineering
Context: The conventional view of political discourse treats voters and politicians as falling on a single left-to-right, liberal-to-conservative dimension. This assignment challenges that assumption, exploring whether social and political attitudes are better understood as multidimensional. Data are sourced from the Pew Research Center, requiring up-front data preparation before analysis.
Management Problem: You are working as a pollster or political consultant serving a hypothetical client — a lobbyist, politician, or news organization. Your task is to gather and tag documents related to an upcoming election or piece of legislation. Beyond simple source-level labels (e.g., "Republican" vs. "Democrat"), you develop a more nuanced, data-driven categorization of political leanings using unsupervised methods.
Reference: Hatley (2018) illustrates R methods for working with Pew Research Center data.
Methods: K-Means Clustering, Hierarchical Clustering
Dataset: Melbourne Housing Market (January 2016)
Context: Cluster analysis is applied to identify natural groupings among residential properties in the Melbourne housing market.
Management Problem: You own a real estate firm and want to identify distinct housing types to support marketing, territory assignment, and property appraisal. As a real estate investor, you are also interested in identifying undervalued properties by detecting anomalies or outliers relative to comparable homes.
Methods: Unsupervised Pretraining, Logistic Regression, Dimensionality Reduction
Dataset: German Credit (OpenML #31)
Context: This assignment evaluates whether unsupervised learning can improve the performance of a downstream supervised classifier. Traditional logistic regression has been used to classify credit card applicants as good or bad risks.
Management Problem: You work for a credit card company aiming to grow its customer base without taking on excessive risk. Prior analysis established that misclassifying a bad customer costs five times as much as declining a good one. The question is whether unsupervised pretraining — learning latent structure in the data before supervised training — can reduce misclassification costs and improve model accuracy.
Methods: DBSCAN, Local Outlier Factor (LOF), Isolation Forest
Context: Fraud detection is framed as an anomaly detection problem applied to sales transaction data. The company suspects some sales representatives may be misreporting product sales to inflate commissions.
Management Problem: You are asked to recommend a fraud detection method by comparing two approaches — DBSCAN/LOF and Isolation Forest — across 133,731 unlabeled training transactions spanning 798 products. Evaluation is conducted on 15,732 labeled test transactions (14,462 normal, 1,270 fraudulent). Given the class imbalance, the primary evaluation metric is F1 score, combining precision and recall. Analysis includes EDA and supervised evaluation using binary labels (normal vs. fraud).
By the end of this course, students are able to:
- Apply dimensionality reduction techniques to continuous multivariate data
- Identify and interpret natural groupings using a variety of clustering methods
- Mine association rules and model dependencies in categorical data
- Visualize high-dimensional data using scaling and embedding techniques
- Build and deploy autoencoder-based anomaly detection systems
- Communicate unsupervised learning findings clearly to technical and non-technical audiences
For questions related to coursework or collaboration, reach out via GitHub Issues or the contact information below.
| Student | Joshua Pasaye |
joshuapasaye2027@u.northwestern.edu |
|
| Program | Master of Science in Data Science |
| Course | MSDS 411 — Unsupervised Learning Methods |
This repository contains coursework completed as part of the MSDS 411-DL: Unsupervised Learning Methods course in the Master of Science in Data Science program.