ML Task Families

Naming the task picks the math

The first decision in any ML project is naming the task family. The family fixes the target shape, the metric, the legal split strategy, and the baseline. It is not taxonomy trivia. It is the contract that decides whether your evaluation is honest.

The seven families you will see most

Regression predicts a number on a continuous scale (price, demand, latency).
Binary classification picks one of two labels, usually with a probability (churn, fraud, click).
Multi-class classification picks one of K labels (digit, intent, category).
Multi-label classification assigns any subset of K labels (tags on an article).
Ranking orders items for a user (search results, recommendations).
Forecasting predicts future values from past values (sales, capacity).
Anomaly detection flags rare or unusual patterns (fraud, machine failure).

The trap of the wrong family

Calling everything classification because the final UI shows yes or no is the most common framing mistake. Often the real problem is ranking (top-10 leads to call), forecasting (capacity next week), or scoring (risk between 0 and 1). The wrong family quietly poisons the metric.

Code

A task-family map for one team's backlog·python

task_map = {
    "house_price_estimate": "regression",
    "will_user_churn": "binary_classification",
    "support_ticket_topic": "multi_class_classification",
    "article_tags": "multi_label_classification",
    "search_results": "ranking",
    "weekly_demand": "forecasting",
    "machine_failure_alert": "anomaly_detection",
}

Same data, two families, two metrics·python

# Treat as classification: did the lead convert?
from sklearn.metrics import roc_auc_score
auc = roc_auc_score(y_converted, p_convert)

# Treat as ranking: are the top-K leads the right ones?
top_k_idx = p_convert.argsort()[::-1][:50]
precision_at_50 = y_converted.iloc[top_k_idx].mean()

Naming the task picks the math

The seven families you will see most

The trap of the wrong family

Code

External links

Exercise

Progress

Comments 0