PSOKMeans

My undergraduate thesis at UNIFEI: using Particle Swarm Optimization as an initialization mechanism for K-Means clustering. Fancy words for “can we make clustering less random and more reliable?”

The Problem

K-Means clustering has a dirty secret: results vary wildly depending on where you randomly place the initial centroids. Run it twice, get different clusters. Not ideal when you’re trying to do science.

The Solution

Instead of random initialization, use PSO to search for good starting positions. Each “particle” in the swarm encodes a set of centroids. The swarm explores the solution space, particles share information about promising regions, and after a few iterations you get better starting points than random chance.

The Results

It worked. PSO initialization consistently outperformed random initialization on the standard datasets (Iris, Glass, Seeds, Yeast). The thesis has the numbers and statistical analysis.

The Tech

• Python with NumPy, SciPy, PySwarm
• Rand Index for evaluation (measures clustering quality against ground truth)
• 100 iterations per test, 4 rounds, statistical aggregation
• Matplotlib 3D visualizations of cluster results and particle swarm positions

The Experience

Wrote 40+ pages overnight to hit the deadline. Busy semester - I was preparing to leave for University of Maryland and didn’t want loose ends. Poor time management or clutch performance? Both.

My advisor was Isabella Drummond, who taught most of our AI courses: neural networks, data mining, the works. She gave me pointers and freedom to work independently. Good mentorship style.

Presenting to the “banca” (thesis committee) was fun. Got some of my favorite professors there, exchanged a lot of ideas. My friends from Utyl Business Platforms and my girlfriend (now wife) came to watch.

Reflection

Sometimes I wonder if I should have pushed the academic path further before chasing industry money. The research mindset - systematic experimentation, statistical rigor, building on existing work - still shapes how I approach engineering problems. But the pull of actually building things people use was too strong.