Course contentsShow
← Machine Learning and Deep Learning
Lesson 141 of 3,5382. Machine Learning Mindset and Core ConceptsFree lesson

Baseline Models: Starting Simple

Establishing simple benchmarks before building complex models to measure real improvement.

Baseline Models: Starting Simple

What you'll learn: Why you should always start with a simple model before building anything complex, and how it helps you measure real progress.

The Power of Starting Simple

Before diving into sophisticated neural networks or ensemble methods, you need to know what "good" actually looks like for your problem. A baseline model is an intentionally simple approach that establishes a minimum performance bar. Think of it like a practice run before the real race—you need to know your starting time to measure improvement.

Why Baselines Matter

Imagine training a complex model for days, achieving 85% accuracy, and celebrating... only to discover that simply predicting the most common class gives you 90% accuracy. Without a baseline, you wasted time and resources on a model worse than doing almost nothing!

Baselines answer critical questions:

  • Is my problem actually hard, or is it trivial?
  • Is my complex model genuinely learning, or am I fooling myself?
  • Where should I invest my effort next?

Common Baseline Approaches

For classification:

  • Predict the majority class (most frequent label)
  • Random guessing based on class frequencies

For regression:

  • Predict the mean of all training targets
  • Predict the median value

These require minimal code and computation but provide invaluable context. If your fancy model only beats the mean by 2%, maybe the problem needs better features, not a bigger model.

The Workflow Impact

Starting with baselines respects the ML development lifecycle you've learned. Before exploring complex architectures, establish your reference point. This keeps you honest, saves time, and guides where to focus your optimization efforts.

Key Takeaway: Always establish a simple baseline model first—it's your measuring stick for whether added complexity actually delivers value, preventing wasted effort on models that don't meaningfully improve over trivial approaches.