Cimentar Futuro Foundation

Learning Objectives

Interpret accuracy, precision, and recall for a classification model
Explain why accuracy alone can be misleading on imbalanced data
Recognize common sources of bias in training data

A model's accuracy score alone can be dangerously misleading. If only 1% of transactions in a dataset are fraudulent, a model that predicts "not fraud" every single time will be 99% accurate while catching zero fraud, which is why analysts also use precision (of the cases flagged positive, how many actually were) and recall (of all actual positive cases, how many were caught). This module also covers bias: if a hiring model is trained on ten years of a company's past hiring decisions, and those decisions reflected human bias against a particular group, the model will learn and often amplify that same bias, even without being told anyone's demographic group directly, because other correlated variables can act as proxies.

Module Quiz

1. Recall measures:

How fast a model runsThe total size of the datasetOf all actual positive cases, the proportion the model correctly identifiedHow many features the model uses

2. On a dataset where only 1% of cases are truly positive, why can a 99% accuracy score be misleading?

A model could reach 99% accuracy by simply predicting 'negative' every time and catch zero real positivesThe dataset would need to be deleted99% accuracy is always a sign of a broken modelAccuracy cannot be calculated on imbalanced data

3. A hiring model trained on biased historical decisions is likely to:

Only reflect bias if demographic data is explicitly includedAutomatically correct the bias on its ownLearn and potentially amplify the same bias found in its training dataHave no relationship to the training data at all

Module 9: Evaluating Models & Avoiding Bias