SDSS17 Stellar Classification - MATH 4323 Project

Directory Structure

Stellar-Classificaton/
├── data/                          <- All data files go here
│   ├── star_classification.csv    <- Raw dataset from Kaggle (do not modify)
│   ├── df_clean.rds               <- Cleaned, unscaled, with class labels
│   ├── df_supervised_train.rds    <- Scaled, with class, 80% train split
│   ├── df_supervised_test.rds     <- Scaled, with class, 20% test split
│   ├── df_unsupervised.rds        <- Scaled, no class labels
│   └── class_labels.rds           <- Class labels for post-hoc cluster comparison
├── report/                        <- Final report files (docx, images, etc.)
├── 00_eda.R
├── 01_preprocessing.R
├── 02_pca.R
├── 03_knn.R
├── 04_svm.R
├── 05_kmeans.R
├── 06_hierarchical.R
├── 07_comparison.R
└── README.md

Scripts (run in order)

Script	Owner	Purpose
00_eda.R	Diego	Exploratory data analysis on raw data
01_preprocessing.R	Diego	Cleans data, scales, splits, saves all .rds files
02_pca.R	Anthony	PCA for dimensionality reduction and visualization
03_knn.R	Asibong	KNN tuning and evaluation
04_svm.R	Diego	SVM kernel selection, tuning, and evaluation
05_kmeans.R	Anthony	K-Means clustering and silhouette analysis
06_hierarchical.R	Asibong	Hierarchical clustering, linkage selection, dendrogram
07_comparison.R	Shared	Compares test errors (supervised) and silhouette scores (unsupervised)

Data Files

You only need to run 01_preprocessing.R once. It generates all the .rds files below.

File	Used by	Description
df_clean.rds	Anyone	99,999 rows, 7 cols (u, g, r, i, z, redshift, class). Unscaled. Good for reference or re-scaling.
df_supervised_train.rds	KNN, SVM	79,999 rows, scaled features + class column.
df_supervised_test.rds	KNN, SVM	20,000 rows, scaled features + class column.
df_unsupervised.rds	PCA, K-Means, Hierarchical	99,999 rows, 6 scaled features, no class labels.
class_labels.rds	K-Means, Hierarchical	Vector of 99,999 class labels for checking if clusters match known categories.

How to Load Data

At the top of your script, set the data directory and load what you need:

DATA_DIR <- file.path("C:/Users/diego/Downloads/Stellar-Classificaton", "data")

# Supervised scripts (KNN, SVM)
df_train <- readRDS(file.path(DATA_DIR, "df_supervised_train.rds"))
df_test  <- readRDS(file.path(DATA_DIR, "df_supervised_test.rds"))

# Unsupervised scripts (PCA, K-Means, Hierarchical)
df_unsupervised <- readRDS(file.path(DATA_DIR, "df_unsupervised.rds"))
class_labels    <- readRDS(file.path(DATA_DIR, "class_labels.rds"))

Important Notes

• Always use set.seed(42) before any randomized operation so results are reproducible and KNN/SVM use the same train/test split.
• Do not re-scale the .rds data. It is already standardized (mean=0, sd=1). Scaling again will distort values.
• Do not re-run preprocessing unless the pipeline changes. The .rds files are the single source of truth.
• Dropped columns from raw data: obj_ID, run_ID, rerun_ID, cam_col, field_ID, spec_obj_ID, plate, MJD, fiber_ID (metadata), alpha, delta (sky coordinates with no predictive value).
• One sentinel row with -9999 values in u, g, z was removed.

Required Libraries

install.packages(c(
  "tidyverse",    # data manipulation and plotting
  "e1071",        # SVM (tune, svm)
  "class",        # KNN (knn)
  "caret",        # confusion matrices, model utilities
  "factoextra",   # eclust(), fviz_cluster(), fviz_dend()
  "cluster",      # silhouette()
  "corrplot",     # correlation heatmaps
  "GGally",       # ggpairs scatterplot matrix
  "gridExtra"     # arranging multiple plots
))