Machine learning study · 2025
Traffic, thirty minutes ahead
A machine-learning project from the Uber × upGrad MentorMind program that predicts traffic congestion thirty minutes ahead using historical data.
01 · The problem
Why this was
worth building.
Urban traffic changes across peak hours, weekdays, weekends, seasons, and junctions. Predicting that congestion in advance can support better route planning and smarter transport systems.
02 · What I wanted to achieve
Understand traffic patterns before choosing a model
Build useful time-based features from historical data
Compare models and evaluate thirty-minute forecasts
03 · System architecture
How the pieces
work together.
04 · Technical decisions
Explore before modelling
I used exploratory analysis and visualisations to compare junctions, identify peak-hour behaviour, and understand how traffic changed over time.
Engineer time-based features
The forecasting data was prepared with features that represent recurring traffic behaviour across hours, days, and different junctions.
Compare prediction models
I trained and evaluated multiple machine-learning approaches rather than relying on the first result, then compared their forecasting performance.
05 · Project targets
These are design targets and project measures, not claims about production usage.
The hardest part
The most important challenge was finding meaningful patterns in raw historical data. Traffic varied by both time and junction, so preprocessing and exploratory analysis shaped the modelling choices that followed.
The choice I made
More complex models are not automatically more useful. I compared approaches using evaluation results and interpretability rather than choosing an algorithm only because it was more advanced.
What I learned
This project strengthened my understanding of the complete machine-learning pipeline. More importantly, it taught me that good machine learning begins with understanding the data before choosing an algorithm.