◆ Data Science · Case Study 03

Poisonous or edible? A classification where being wrong has a body count

8,124 real mushrooms, 22 traits each, one question: edible or poisonous. Watch the whole pipeline, then meet the lesson: a model can be 99.9% accurate and still unsafe. Only the confusion matrix, not the accuracy score, will tell you so.

← Back to the case study ⬇ Download raw data (CSV)

🍄

Read this first

This is a clean, curated field-guide dataset, and it is nearly separable. Never eat a wild mushroom based on a model, or on this page. The value here is the method: how to handle a classification problem where one kind of error is far more expensive than the other.

Loading the pipeline…

STEP 1

The raw data

The source is the UCI Mushroom dataset, drawn from a 1981 Audubon Society field guide. Every column is categorical, a single letter in the raw file, decoded here for reading. The target is the class: edible or poisonous. Note the flagged stalk-root cells, marked missing in the source.

edible / poisonous = the label ✎ gold = missing value (stalk-root) 22 traits total; a readable subset is shown

STEP 2

Cleaning

All categorical, so no scaling or outliers, but two judgment calls. veil-type has the same value for every mushroom, so it carries zero information and is dropped. stalk-root is missing for 2,480 rows; rather than throw away a third of the data, the missing value is kept as its own category. Then every trait is one-hot encoded, turning 21 columns into 116.

raw columns →

→

one-hot features

Rows (all kept)

Constant column dropped · veil-type

stalk-root missing · kept as its own category

Features after one-hot encoding

STEP 3

The tell: odor

Before any model, one feature carries most of the story. Grouped by odor, the classes separate almost perfectly: every foul, fishy, spicy, pungent, creosote, or musty mushroom is poisonous; almond and anise are edible. Only "no odor" mixes, and even then it leans edible.

Odor vs class

Green edible, red poisonous · stacked counts

Spore print color

green spore = poison

Gill size

narrow leans poison

Bruises

no bruises leans poison

STEP 4

Two models, and why accuracy lies

A logistic regression baseline and a shallow decision tree, both trained on 70% of the data and scored on the 30% held out. Both clear 99.8% on every metric. That is exactly where a careful analyst gets suspicious, not satisfied.

The confusion matrix splits the mistakes. A false positive throws away a good mushroom. A false negative feeds someone poison. Here every error is the second kind.

STEP 5

ROC, and tuning the threshold to be safe

The ROC curve confirms the models rank poison above edible almost perfectly (AUC near 1.0). But the 0.50 cutoff is a choice, not a law. Since a missed poison is the costly error, drag the threshold down and watch the false negatives disappear, at some cost elsewhere.

ROC curve

true-positive vs false-positive rate

Decision threshold 0.50

correct edible

correct poison

FP · good tossed

FN · poison missed

STEP 6

Live model: run the decision tree

This is the real trained tree, running in your browser. Set the traits and it walks the same rules it learned, showing the exact decision path to its verdict. These eight traits are the ones the tree actually uses.

Verdict

EDIBLE

confidence

Decision path

Data: UCI Machine Learning Repository · Mushroom (Audubon Society field guide) ·
Analysis, models, and visuals by Carlos Abel Vivanco / AbleV Labs · Read the full notebook → · Back to the case study →