AUTHOR: Rahul Atram
INSTITUTION: SGGS Institute of Engineering & Technology, Nanded
DATE: 19 April 2026
THE MOMENT OF DISCOVERY
It was a quiet Sunday evening in Nanded, the kind of evening where the air is still and the mind begins to wander. I had just returned from a short trip, unpacking my bags and preparing for the week ahead, when a notification flashed on my screen: the Ex-Machina Hackathon was officially live. The challenge was unique—classify thousands of years of human art history using nothing but digital metadata.
As a third-year Computer Science student, I have always believed that data is not just a collection of numbers; it is a story waiting to be told. However, looking at the spreadsheet of 5,000 messy museum records, it felt more like a riddle than a story. There were missing years, inconsistent measurements, and cryptic notes. But that is exactly where the journey began. What started as a late-night curiosity turned into a high-performance machine learning pipeline that achieved a verified 94.10% accuracy. This is the story of that discovery.
DECODING THE PROBLEM: ART BEYOND THE IMAGE
The competition was a test of "Metadata Intelligence." We were given 4,000 artworks and asked to predict their medium—the physical substance they were made of. This wasn't about looking at photos of paintings; it was about understanding the words used to describe them.
We were dealing with eight beautiful, distinct categories:
- Acrylic: The modern, vibrant medium of the 20th century.
- Ink: The sharp, decisive lines of sketches and calligraphy.
- Oil on Canvas: The heavy, textured gold standard of the masters.
- Oil on Wood and Panel: The rigid, durable ancestors of modern painting.
- Print: The art of reproduction, etching, and woodblocks.
- Tempera: The ancient egg-yolk based paint used in historical icons.
- Watercolor: The translucent, flowing beauty of landscapes.
The challenge wasn't just to get a high score. It was to build a system that truly understood the nuances of art cataloging.
THE "AHA!" MOMENT: LISTENING TO THE DATA
Most people start a machine learning project by immediately writing code. I decided to start by reading. I spent the first hour simply scrolling through the rows of the data, and that is when I found my "Smoking Gun."
I noticed a column called 'Caption'. While other columns were missing or fragmented, the curators at these museums were incredibly consistent in their captions. They would write things like: "A watercolor landscape titled 'The River'..."
This was my breakthrough. I realized that the answer wasn't hidden; it was written in plain English right in front of us. The machine didn't need to guess; it just needed to learn how to read. This "Caption Signal" became the heart of my entire strategy.
THE JOURNEY THROUGH THE EXPERIMENTS
In my quest for accuracy, I followed a path of increasing intelligence. I didn't want to build a "black box"; I wanted to understand the progression of my model's brain.
Step 1: The Keyword Baseline
I started with a technique called TF-IDF. Think of this as a very fast librarian who scans a book for important keywords. If the librarian sees the word "canvas," they guess "Oil." This simple approach got me to a solid 91% accuracy. It was a great start, but art is more than just keywords.
Step 2: The Margin Carver
I then upgraded to a Linear Support Vector Machine. This is a bit like a judge who tries to draw the absolute sharpest line between two different piles of evidence. It brought our accuracy up to 94.20%. We were getting closer to the truth.
Step 3: The Champion Model — SBERT + CatBoost
Then came the real revolution. I introduced a "Transformer" model called Sentence-BERT. Unlike my earlier librarian who only counted words, Sentence-BERT actually "reads" the sentence. It understands context. It knows that "pigment on fabric" is the same as "painting on canvas" even if the words are different.
I combined this "reading brain" with CatBoost—a gradient boosting model that acted as the "historical memory." CatBoost looked at the years (y0/y1) and the size of the artwork (area) and combined them with the text. This hybrid approach allowed us to hit a verified cross-validation peak of 94.10%, with initial probe estimates reaching even higher.
REALITY CHECK: HONESTY IN DATA SCIENCE
As a student, it is tempting to chase a 100% score. But this hackathon taught me a valuable lesson in professional honesty. While my initial probes hit 99% accuracy because of the heavy "Caption Signal," I realized that a truly useful model must be robust.
In my final version, I focused on proper feature integration and handling missing values. I realized that 94.10% is not just a number; it represents a model that is balanced, realistic, and ready for the real world. This intellectual maturity—knowing that data is never perfect—is perhaps the most important thing I learned throughout this experience.
THREE LIFE-CHANGING LESSONS
- Observation is more powerful than Algorithms: The "Caption Signal" was discovered because I spent time looking at the raw data, not just the code. Always look at your data first.
- Baselines are the ground you stand on: Never start with a complex neural network. Start small to understand the "floor" of your performance.
- Art and Science are not enemies: Using AI to understand human creativity was a beautiful experience. Machine learning is simply a new way to appreciate the precision of those who have cataloged human history for centuries.
CONCLUSION: THE FUTURE OF THE MUSEUM
What I built in the Ex-Machina hackathon was more than a classifier. It was a bridge between the historical archives of the past and the intelligent systems of the future. I learned that while a machine doesn't have an "eye" for art, it definitely has an "ear" for the language we use to describe it.
To my fellow students at SGGS and beyond: don't be afraid of the complexity. AI is just a tool, and your curiosity is the power that makes it work. Let’s keep building, keep questioning, and keep telling the stories hidden inside the data.
Let's build the future together.
Journey continues = https://github.com/Rahulatram321/Artistic-Medium-Classification-from-Metadata.git.
Top comments (0)