🐧 Penguin Clustering Analysis

Illustration source: @allison_horst

Welcome to my penguin clustering project! In this analysis, I explored real biological data collected in Antarctica to uncover natural groupings among penguins — potentially even hinting at the discovery of a new species! 🧬

---

📦 Dataset Overview

The dataset (penguins.csv) was collected by Dr. Kristen Gorman and the Palmer Station, Antarctica LTER. It contains biometric measurements of penguins from three known species: Adelie, Chinstrap, and Gentoo — although species labels were not included.

These species are known to inhabit the region, and the goal of this project is to uncover natural groupings based on their physical measurements.

📊 Features

The key biometric features recorded in the dataset are listed below:

The dataset includes several biometric features measured from each penguin. The image above shows how culmen length and culmen depth are defined — terms used in the dataset to describe bill dimensions.

Column	Description
culmen_length_mm	Length of the penguin's bill (culmen) in mm
culmen_depth_mm	Depth of the bill (culmen) in mm
flipper_length_mm	Flipper length in mm
body_mass_g	Body mass in grams
sex	Biological sex (MALE / FEMALE)

---

🧹 Data Cleaning

To prepare the dataset for analysis:

• I confirmed that there were no missing values in the dataset.
• I created a shorter alias 'df' for the original DataFrame 'penguins_df'.
• I converted the 'sex' column into binary dummy variables (sex_MALE, sex_FEMALE) using pd.get_dummies.

This ensured that all features used for clustering were numerical and complete.

---

⚖️ Standardisation

I standardised all numerical features using StandardScaler from scikit-learn. This transformed the data so that each feature had a mean of 0 and a standard deviation of 1, ensuring that features like body mass or flipper length didn’t dominate the clustering due to differences in scale.

---

📈 Elbow Analysis

To determine the optimal number of clusters (k), I applied the Elbow Method. I plotted the model’s inertia (within-cluster sum of squares) across values of k from 1 to 10.

There was a clear drop in inertia up to k = 5 or k = 6, suggesting meaningful separation in those cases. A subtle inflection between k = 7 and k = 8 hinted at potential deeper structure in the data.

---

🧪 Clustering with k = 5

I proceeded by clustering the standardised data using KMeans with k = 5. After fitting the model, I assigned the resulting clusters to each penguin in the dataset.

Here’s a random sample of penguins from the clustered dataset:

Index	culmen_length_mm	culmen_depth_mm	flipper_length_mm	body_mass_g	sex_MALE	label
159	45.9	17.1	190.0	3575.0	0.0	2
67	35.5	17.5	190.0	3700.0	0.0	2
75	36.7	18.8	187.0	3800.0	0.0	2
166	48.5	17.5	191.0	3400.0	1.0	4
270	43.2	14.5	208.0	4450.0	0.0	3

---

🔍 Visualising Cluster Distributions

To better understand how each feature contributed to the clustering, I created box plots showing the distribution of each variable across the five clusters.

These visualisations helped highlight how flipper length, culmen size, and body mass varied between clusters — possibly corresponding to different species or sex-based differences.

---

🐧 Plot Twist!

When I plotted the sex_MALE variable alongside the cluster labels, something unexpected showed up — in Cluster 4, both males ♂️ and females ♀️ appeared! 🤯

This overlap might be due to small males from one species clustering with larger females from another. It raised a compelling question: Had I completely missed the third species — or maybe even stumbled upon a fourth? 🔍🐧

Curious, I tried clustering again with k = 6, hoping the mix would disappear. It didn’t. Same result at k = 7. So I pushed it further — what about k = 8?

At k = 8, the sex-based overlap seemed to vanish. Could this mean there are actually four species, not just three? 🤔 To dig deeper, I used t-SNE to project the clusters into two dimensions and get a better sense of how the groups separate visually. 📈

In the t-SNE map with five clusters, we can clearly see five distinct groups. One of them shows a subtle internal separation, which might reflect the similarity between two subgroups the model wasn’t able to split. There's also some evidence of the algorithm stretching the borders of another cluster and misplacing a few points. 📏📊

Now take a look at the t-SNE map for k = 8...

Here, the map looks messy. Several clusters are overlapping or spread in ways that don't align with clear group boundaries. What initially looked like a new species turns out to be just noise or natural variation. 📉

So, are there four species? Most likely not. What we saw was probably just intra-species variation, especially between males and females that share similar features. 🙃

Still — a fun detour, and a reminder that sometimes, the data keeps its secrets. 🧬🐧💥

---

📌 Conclusion

This project demonstrates how unsupervised learning can uncover natural groupings within real-world biological data. Using only physical measurements — and no species labels — I was able to:

• Identify meaningful clusters of penguins.
• Reveal potential overlaps and latent structure.
• Open the door to future biological questions (and possibly discoveries 🧪).

---

📁 Repository Contents

• notebook.ipynb – Main Jupyter notebook with the full clustering analysis
• README.md – Project overview and explanation
• /data – Contains the penguin dataset (penguins.csv)
• /plots – Visualisations generated during the analysis
• /figures – Illustrative images used in the README and notebook

---

• 🙌 Acknowledgements

• Dataset: Palmer Station LTER
• Illustration: Allison Horst – penguins art
• 🧠 Powered by Python, pandas, scikit-learn & matplotlib.