The competitions taught me things no course ever did — and a few things I had to unlearn.
Six months, eleven competitions, one bronze medal, and an opinion about Kaggle that is more complicated than I expected it to be. I joined the platform in my second year at BITS thinking it would be the fastest path to ML expertise. I was partly right and partly wrong in ways that took me a while to understand.
The quality of the discussion forums on Kaggle is extraordinary. During the Tabular Playground Series competition I participated in last year, someone in the discussion section posted a 4,000-word explanation of why gradient boosting outperforms neural networks on structured data with less than 100k rows. This was better than any lecture I had sat through. The culture of sharing, especially in the getting-started and intermediate competitions, is real.
The other thing that is great: feedback loops. In college projects you submit something, get a grade two weeks later, and the feedback is usually vague. On Kaggle you submit a prediction, get a leaderboard score in 10 minutes, and can immediately test whether your hypothesis was right. I learned more about what actually moves model performance in two months of this than in a full semester of ML coursework.
Kaggle teaches you to win competitions, which is not the same thing as building useful ML systems. After six months I was reasonably good at feature engineering, ensemble stacking, and cross-validation strategies. I knew almost nothing about data pipelines, model serving, monitoring for data drift, or how to have a conversation with a non-technical stakeholder about what a model can and cannot do. These gaps showed up badly in my internship interview at a data science firm, where the first interview question was about how I would handle a situation where model performance degraded after deployment.
There is also a leaderboard addiction that is subtle and real. I spent a weekend optimizing a model from 0.847 to 0.849 AUC on a competition that had no practical stakes and no real-world equivalent. That was time I could have spent building something, writing something, or resting. The dopamine loop of the leaderboard is very well-designed and it does not always serve your learning.
Do Kaggle, but do it with a plan. Pick three or four competitions in domains you actually care about. Go deep on those rather than wide across everything. Read the top solutions after competitions end — this is where the real learning is. And alternate it with building actual things: projects where the data is messy, the requirements are unclear, and the definition of done is yours to figure out. The combination is stronger than either alone.
My bronze medal is in my Kaggle profile and I am genuinely proud of it. But the thing I am more proud of is the recommendation system I built for my BITS project afterward, which handled real users and broke in real ways and taught me things no competition ever could. Both were necessary. Neither was sufficient.
Join Sparrow — written by college students, for college students
Read unlimited articles, spark the ones you love, and share your own voice.
Written by
Sneha GuptaCS at BITS Pilani. Kaggle nerd. Failed startup co-founder. I write about ML, building things that flop, and Pilani winters.
135 followers
Responses
Sign in to join the conversation
Sign in