Feature Engineering in Neural Network
1. The Story: “The Bakery AI Assistant”
Once upon a time in a town, a baker named Maya decided to use a neural network to predict the sales of pastries in her bakery.
She had a small dataset:
- Temperature
- Day of the week
- Holiday or not
- Number of pastries made
- Sales in ₹
But initially, she fed the raw data directly into the neural network. The model gave confusing predictions. Sometimes it thought sales were high on cold Mondays, other times on hot Sundays.
Frustrated, Maya met an AI engineer named Anya. Anya said:
“Your model is like a child learning a language. If you give raw noise, it gets confused. Let’s feed it stories, not just facts.”
She introduced Feature Engineering:
- Converted categorical values like “Day of Week” into numbers using one-hot encoding.
- Normalized numerical values like “Temperature” and “Pastries made”.
- Created interaction features, e.g., Temperature × Holiday to model cold holidays.
- Derived a new feature: is_weekend.
After retraining, the predictions improved significantly! Maya’s bakery boomed!
2. What is Feature Engineering?
Feature Engineering is:
- Creating new relevant inputs from existing data.
- Transforming or encoding data so neural networks can understand them better.
- Sometimes even removing irrelevant features.
3. Why Does It Matter in Neural Networks?
- Neural Networks can learn features, but with limited data, good features act like shortcuts to the truth.
- Feature engineering improves:
- Convergence speed
- Model accuracy
- Interpretability
- It turns complex relationships into learnable patterns.
Feature Engineering in Neural Network – Feature Engineering example with simple python