Generative vs Discriminative ML

Product teams mix both: a classifier routes tickets; a generator drafts replies. Knowing which paradigm you need saves months of misfit tooling.

When to use discriminative models

• Fixed labels with abundant labeled data (fraud, churn, intent ID)
• Need calibrated probabilities for thresholds
• Smaller, cheaper models on tabular or vision features

See AI track supervised learning lessons for baselines.

When to use generative models

• Open-ended language, code, or creative variation
• Few-shot instructions instead of thousands of labels
• Assistants that must synthesize prose from many sources (with RAG)

Hybrid pattern

# Pseudocode: route then generate
intent = classifier("user message")  # discriminative
if intent == "billing":
    draft = llm.generate(policy_prompt + user_message)  # generative
else:
    draft = template_fallback(intent)

Important interview questions and answers

1. Q: Can one product use both?
   A: Yes—routing, safety classifiers, and RAG rerankers are often discriminative around a generative core.

Self-check

1. When would you refuse an LLM for a churn model?
2. What does the hybrid pseudocode optimize?

Tip: Sketch a flowchart: route (discriminative) → draft (generative) before buying GPU hours.