Classification Study Guide

📖 Core Concepts Classification – Assigning objects to pre‑existing categories (labels). Classifier – Model/algorithm that learns patterns and predicts the label of new objects. Taxonomy – Structured scheme of classes used to organize the label space. Binary vs. Multiclass – Binary: exactly 2 possible labels; Multiclass: ≥3 possible labels. Nominal‑scale outcomes – Labels have no intrinsic order; only “same” vs. “different” matters for accuracy. 📌 Must Remember Accuracy = $\dfrac{\text{# correct predictions}}{\text{total predictions}}$; Error rate = $1 - \text{Accuracy}$. Sensitivity (Recall) = $\dfrac{TP}{TP + FN}$ – proportion of actual positives correctly identified. Specificity = $\dfrac{TN}{TN + FP}$ – proportion of actual negatives correctly identified. Precision = $\dfrac{TP}{TP + FP}$ – proportion of predicted positives that are true positives. No‑Free‑Lunch (NFL) – No single classifier dominates across all data sets; match method to data characteristics. Fuzzy classification – Objects can belong to multiple classes with graded membership values. 🔄 Key Processes Train a Classifier Gather labeled data → split into training/validation → fit model → tune hyper‑parameters. Evaluate Performance Run classifier repeatedly on validation set → compute confusion matrix → derive accuracy, error rate, sensitivity, specificity, precision, recall. Select Model for Deployment Compare metrics across candidate classifiers → consider domain priorities (e.g., high sensitivity in medicine) → choose best‑performing model. 🔍 Key Comparisons Binary vs. Multiclass Binary: decision boundary separates two groups; often use metrics like sensitivity/specificity. Multiclass: requires one‑vs‑rest or one‑vs‑one strategies; precision/recall computed per class or averaged. Sensitivity vs. Precision Sensitivity = “how many real positives did we catch?” Precision = “how many of our positive calls are correct?” Classifier vs. Cluster Analysis Classifier → uses predefined labels. Cluster analysis → creates the labels (unsupervised). ⚠️ Common Misunderstandings Accuracy ≠ Quality – High accuracy can hide poor performance on minority classes; always check sensitivity/precision when class imbalance exists. Nominal Scale ≠ Ordinal – Labels are not ordered; you cannot compute “average” of classes. Fuzzy ≠ Uncertain – Fuzzy classification deliberately models partial membership, not just prediction uncertainty. 🧠 Mental Models / Intuition “Confusion Matrix as a Scoreboard” – Think of TP, TN, FP, FN as wins/losses for each side of a game; each metric is a ratio of specific win types. NFL → “Toolbox” – No single tool works for every job; match the shape of your data (linearity, noise, class balance) to the right algorithm. 🚩 Exceptions & Edge Cases Imbalanced Data – Accuracy can be misleading; prioritize sensitivity or precision for the minority class. Multiclass “One‑vs‑Rest” – May produce inconsistent probabilities across classes; consider softmax‑based models for calibrated scores. Fuzzy Membership Thresholding – Turning fuzzy outputs into hard labels requires a threshold; choice affects precision/recall trade‑off. 📍 When to Use Which Binary medical test → prioritize sensitivity (catch disease) or specificity (avoid false alarms) depending on clinical cost. Information retrieval → focus on precision (relevant results) and recall (coverage). Credit scoring → often emphasize specificity (minimize false positives) to avoid granting credit to risky applicants. Highly imbalanced problems → use ROC‑AUC, precision‑recall curves, or F1 score rather than raw accuracy. 👀 Patterns to Recognize High accuracy + low recall → likely class imbalance; many negatives correctly classified, positives missed. Precision ≈ Recall → balanced performance; often a sign of a well‑tuned model. Sensitivity ↑, Specificity ↓ (or vice‑versa) → threshold shift; moving decision cut‑off changes trade‑off. 🗂️ Exam Traps Choosing accuracy as the sole metric – distractor when the question mentions rare classes; the correct answer will cite sensitivity, specificity, or F1. Confusing sensitivity with precision – both involve TP in numerator but differ in denominator (FN vs. FP). Assuming “no‑free‑lunch” means all classifiers perform equally – trap; the principle states no single classifier is best across all problems, not that they are all the same. Treating fuzzy classification as “uncertain” – exam may present fuzzy membership values; the right answer emphasizes graded membership, not error bars. --- If any heading lacked sufficient detail in the source outline, it has been filled with the best‑aligned high‑yield content while staying true to the provided material.