Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/LauraSilRu/exoplanet-profiler/llms.txt

Use this file to discover all available pages before exploring further.

05_clustering_and_evaluation.ipynb takes the 4 PCA components from notebook 04 and applies K-Means clustering to discover natural groupings among the 731 exoplanets. After evaluating K values from 2 to 10, K=2 is selected based on a Silhouette score of 0.4950. Stability is confirmed across multiple random seeds and cross-validated with DBSCAN and hierarchical clustering.

​
Input Contract

Before running this notebook, verify the following:
  • File: data/processed/pca/pca_scores.csv must exist (produced by notebook 04)
  • Shape: 731 rows × 5 columns in the file (pl_name, PC1, PC2, PC3, PC4); the clustering matrix X is 731 × 4 (PC1–PC4 only)
  • Variance: 4 components collectively retain 90.35% of original variance
  • Identifier: pl_name is unique and non-null; it is separated from the feature matrix before clustering and re-joined to labels at export
pl_name is never passed to the K-Means algorithm.

​
K Evaluation

K-Means is evaluated for every value of K from 2 to 10. The Silhouette Coefficient quantifies how well each point fits its own cluster relative to the nearest alternative cluster (1.0 is perfect separation; 0.0 is ambiguous; negative is misassigned). The elbow method plots within-cluster sum of squares (inertia/WCSS) to identify the point of diminishing returns.
The evaluation starts at K=2 because the Silhouette Coefficient requires at least two clusters to compare intra-cluster cohesion against inter-cluster separation. The metric is mathematically undefined for K=1.
KInertia (WCSS)Silhouette Score
24023.360.4950 ⭐
32820.810.3223
42392.980.3238
5+—< 0.28
K=2 is selected. It achieves the highest Silhouette score by a clear margin and corresponds to the primary elbow in the inertia curve. Higher K values produce marginally lower inertia at the cost of substantially weaker cluster separation.

​
Stability Analysis

The K=2 model is retrained with five different random seeds to verify that the solution is not an artifact of centroid initialization.
SeedInertiaSilhouetteAgreement with K=2 Baseline
42 (baseline)4023.360.4950—
104023.360.4950100.00%
20264023.360.4950100.00%
99994023.360.4950100.00%
1004025.210.479699.04%
Three seeds reproduce the baseline solution exactly. Seed 100 produces a marginally higher inertia and assigns 7 of 731 planets differently (99.04% agreement), which represents a negligible boundary effect. The K=2 solution is highly stable, reproducible, and free from sensitivity to centroid initialization.

​
Cross-Validation with Alternative Algorithms

Two alternative clustering algorithms are applied to independently verify that the K=2 structure reflects a genuine pattern in the data rather than an artifact of the K-Means objective function. Hierarchical (Agglomerative) Clustering. The dendrogram shows two clearly dominant primary branches, confirming that the 731 exoplanets naturally split into two large groups when built bottom-up. Label agreement with the K-Means K=2 assignment is 97.13%. The 2.87% disagreement occurs at the boundary between the two populations. DBSCAN (Density-Based). DBSCAN does not require a pre-specified K. It identifies one large, dense majority cluster and approximately 60 isolated points classified as noise or outliers (label = -1). This is consistent with the K-Means Cluster 1 structure — a small population of extreme objects with unusual orbital and physical properties, separated from the mainstream. The convergence of three independent algorithms on the same two-population structure provides strong evidence that K=2 reflects a real, stable partition of the exoplanet data.

​
Cluster Profiles

Cluster 0 — Standard Population

667 exoplanets (91.24%)The mainstream of TESS discoveries. Captures rocky planets, Super-Earths, and Sub-Neptunes with typical orbital periods, moderate radii, and stellar hosts spanning a wide range of effective temperatures. Represents the statistical norm in the sample.

Cluster 1 — Exotic Objects

64 exoplanets (8.76%)Massive gas giants and Hot Jupiters characterized by extreme orbital parameters, high planetary mass, and large planetary radii. These are the outliers confirmed by DBSCAN — physically real objects that sit in the tail of every feature distribution.

​
Label Export

After the final K=2 model is validated, two label columns are added to the dataset and exported:
  • Cluster_K2 — integer cluster label (0 or 1)
  • Familia_Planeta — human-readable descriptive label:
    • 'Familia 0 (Población Estándar)' for Cluster 0
    • 'Familia 1 (Población Exótica / Atípica)' for Cluster 1
Labels are joined to the PCA scores by pl_name — never by row position — to guarantee correct alignment even if row order changes.

​
Output Artifacts

FileDescription
data/processed/clustered_exoplanets.csv731 rows with pl_name, PC1–PC4, Cluster_K2, Familia_Planeta
This file closes the ExoProfiler ML pipeline. It is the primary input for any downstream profiling, visualization, or product layer (e.g., the Streamlit application).

​
Running the Notebook

jupyter notebook notebooks/05_clustering_and_evaluation.ipynb
Prerequisite: data/processed/pca/pca_scores.csv must exist and contain 731 rows. Run notebook 04 first if this file is missing or outdated.

​
Astrophysical Context

The two ExoProfiler clusters align closely with NASA’s recognized planetary size categories:
  • Terrestrials & Super-Earths — rocky planets with radius < 2 R⊕
  • Sub-Neptunes — intermediate gaseous planets, 2–4 R⊕
  • Gas Giants / Hot Jupiters — massive planets > 4 R⊕ with high mass and often short, hot orbits
Cluster 0 (Standard Population) captures the continuum of Terrestrials, Super-Earths, and Sub-Neptunes — the dominant population in TESS discoveries and the statistical backbone of the dataset.Cluster 1 (Exotic Objects) isolates Gas Giants and Hot Jupiters at the extreme end of the mass and radius distributions, characterized by anomalous orbital parameters and high insolation flux.The clusters are data-driven, not manually defined. Their correspondence with NASA’s categories provides interpretive grounding but should not be treated as an exact classification system — the boundary between populations is a continuous gradient, not a sharp physical threshold.

Build docs developers (and LLMs) love

Get started for freeTalk to us