Sklearn Workflow (Cheatsheet)

Context: FIT1043_MOC · Weeks 5–7 modelling in ONE skeleton · code canon in the LAB notebooks (30_Projects/FIT1043_Labs/Week7-Classification-sklearn.ipynb, Week7-Clustering-KMeans.ipynb) · applied depth in Scikit-learn Classification and Clustering.

Unit-specific: FIT1043's linear regression is done with scipy linregress, not sklearn (Week 5 lab) — see Linear Regression in Python (scipy). sklearn here is for classification and clustering (Week 7). The sklearn regression rows below are the standard API for reference/other units.

Quick Revision

• 🎯 Objective: every sklearn model, same four verbs ➔ Estimator() → fit(X_train, y_train) → predict(X_test) → evaluate.
• ⚡ Key Constraint: shapes — $X\in {\mathbb{R}}^{N\times D}$ (2-D, even for one feature via df.loc[:, ['weight']]), $y\in {\mathbb{R}}^N$ (1-D); most beginner errors are a 1-D $X$.

🧩 Workflow Anatomy (execution order)

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
X = df.iloc[:, [0, 1]].values       # (N, D) — always 2-D (or df.loc[:, ['a','b']])
y = df.iloc[:, 2].values            # (N,)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=0)
 
sc = StandardScaler()               # scale AFTER splitting
X_train = sc.fit_transform(X_train) # fit on TRAIN only
X_test  = sc.transform(X_test)      # transform (never fit) TEST
 
model = SomeEstimator()             # swap this line per task ↓
model.fit(X_train, y_train)         # learn parameters from TRAIN only
y_pred = model.predict(X_test)      # apply to UNSEEN data

Why split ➔ evaluating on training data rewards memorisation — the overfitting trap; test error is the honest estimate.

🎛 Estimator Swap Table

Task	Estimator line	Evaluate with	Concept note
linear regression (FIT1043)	linregress(x, y) (scipy.stats) → slope/intercept/r/p/std_err	$r$, plot the line	Linear Regression in Python (scipy)
classification (tree)	DecisionTreeClassifier(criterion='entropy', random_state=0) (sklearn.tree)	confusion matrix + metrics	Decision Trees and Regression Trees
classification (forest)	RandomForestClassifier(n_estimators=20, criterion='entropy') (sklearn.ensemble)	confusion matrix	Random Forest
clustering (no labels!)	KMeans(n_clusters=k, init='random') (sklearn.cluster)	labels_, cluster_centers_; no y in fit(X)	k-means Clustering
(other units)	LinearRegression() / PolynomialFeatures(degree=k) (sklearn.linear_model)	MSE, $R^2$	Linear and Polynomial Regression

📏 Evaluation Calls

Task	Micro-syntax	Reads as
regression error	mean_squared_error(y_test, y_pred)	mean of squared residuals — the loss regression minimises
fit quality	r2_score(y_test, y_pred) or model.score(X_test, y_test)	proportion of variance explained
confusion matrix	confusion_matrix(y_test, y_pred)	rows=actual, cols=predicted ➔ TP/FP/FN/TN
accuracy	accuracy_score(y_test, y_pred)	correct ÷ total — misleading on imbalance
full report	classification_report(y_test, y_pred)	precision, recall, F1 per class
k-means labels	model.labels_ · centroids model.cluster_centers_	assignment after convergence

• Metric choice is the exam question ➔ spam filter wants precision, disease screen wants recall — derive from the confusion matrix, don’t memorise (Classification Evaluation (Confusion Matrix and Metrics)).
• Read coefficients ➔ model.coef_, model.intercept_ — same quantities as R’s fit$coefficients (R Toolkit (Cheatsheet)).

✍️ Practice

df has hours, attendance → passed (0/1). Build an 80/20 split, fit a decision tree, print the confusion matrix and accuracy, and state which metric you'd report if failing students are rare and missing one is costly.

Reference solution

from sklearn.model_selection import train_test_split
from sklearn.tree import DecisionTreeClassifier
from sklearn.metrics import confusion_matrix, accuracy_score
X, y = df[['hours', 'attendance']], df['passed']
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.2, random_state=42)
model = DecisionTreeClassifier().fit(X_tr, y_tr)
y_pred = model.predict(X_te)
print(confusion_matrix(y_te, y_pred), accuracy_score(y_te, y_pred))
# rare + costly misses ⇒ report RECALL on the 'fail' class, not accuracy

• Key moves: 2-D X; fit on train only; metric justified by cost structure.

⚠️ Common Mistakes

• 💡 1-D X crash ➔ a single-bracket df['w'] is $(N,)$; estimators need 2-D $(N,1)$ — use df.loc[:, ['w']] or df.iloc[:, [0]].
• 💡 Fitting on everything ➔ split FIRST; any statistic learned from test rows (even StandardScaler) is leakage — fit_transform train, transform test.
• 💡 Accuracy on imbalance ➔ 95% accuracy on a 95/5 class split is the majority-class baseline, not skill.
• 💡 Set random_state ➔ without it, splits/forests/KMeans seeds change every run and results aren’t reproducible.