Hands-On Practice Set 2

Let's stop and think about this for a second

This round steps things up from round 1 — instead of practicing skills separately, you'll be combining them all at once. That includes reading a problem's requirements and choosing an algorithm, diagnosing overfitting/underfitting from a train/test accuracy gap, and scoping and planning an ML project. Give each task about 10 minutes.

Let's connect it to a real scenario

Task 1: For a loan approval project where 'interpretability (being able to explain why a decision was made) is a business requirement' — decide whether to choose Decision Tree, Random Forest, or Neural Network, and explain why. Task 2: Given a result of 95% training accuracy and 60% test accuracy, decide what's going on (overfitting) and write down at least two ways to fix it. Task 3: For an e-commerce company's project on 'which product recommendations a customer would like', decide whether to approach it with Supervised or Unsupervised Learning and explain your reasoning. Task 4: Write down three questions you'd want to ask stakeholders before starting an ML project (about data availability, success metrics, and deployment plans).

Can decide on algorithm selection based on business requirements (like interpretability)
Can write up an overfitting diagnosis and fix approach
Have a set of ML project scoping questions written down

Let's look at it together

text

# Task 1 - interpretability-first choice
Requirement: must explain WHY a loan was denied
-> Decision Tree (single tree, human-readable path)
-> NOT Neural Network (black box, hard to explain to
   a regulator or a rejected applicant)
-> Random Forest is a middle ground (feature_importances_
   helps, but individual decisions are harder to trace
   than a single tree)

# Task 2 - overfitting fixes
1. Add regularization (Ridge/Lasso, or reduce max_depth
   for trees)
2. Get more training data, or simplify the feature set
   (remove noisy/irrelevant features)

You should see

You'll come away with an algorithm selection decision, an overfitting diagnosis/fix, and a set of ML project scoping questions.

Try it in 5 minutes

Connect your Task 4 stakeholder question list to the Project 2 (Customer Churn) scenario, and answer them yourself as a mock interview (self Q&A).

One thing to watch out for

There's no universal 'best algorithm' — remember that every ML project involves trade-offs depending on dataset characteristics, business requirements, and deployment constraints.