§ 1.4Module I

Applications of Machine Learning

1.4 — Applications of Machine Learning

Recall first. Choose one everyday system that appears to “make a decision.” Is it definitely using machine learning, or could a fixed rule/database be enough? Explain your guess.

Start with the task, not the buzzword

An application is an ML application when a model learns a useful mapping or structure from data and uses it to support a real task. Describe it with five questions:

  1. What decision or prediction is required?
  2. What are the inputs available at decision time?
  3. What is the target, if the task is supervised?
  4. How is success measured?
  5. What is the cost of a wrong or delayed decision?

This prevents a domain list (“ML is used in healthcare, finance, and education”) from becoming an explanation. The same algorithm can serve different tasks, and the same task can require different models or metrics.12

Common application patterns

PatternWhat the model doesExample
Regression / forecastingPredicts a numberEstimate electricity demand or delivery time
ClassificationAssigns a class or probabilityDetect spam or identify a defective product
Ranking / recommendationOrders candidates for a user or systemRank search results or recommend courses
Clustering / discoveryFinds groups without supplied labelsSegment customers or group documents
Anomaly detectionFlags unusual observationsInvestigate unusual sensor readings
Representation / perceptionConverts raw signals into useful structureRecognize objects in images or transcribe speech
Decision supportProduces a prediction for a human or policyPrioritize inspections or support triage

The output is not automatically the final decision. A human, rule, threshold, or downstream system may act on it.

Domain examples and their risks

DomainPossible taskData examplesRisk to examine
HealthcareClassify risk or assist image interpretationRecords, images, signalsFalse negatives, privacy, distribution shift
FinanceDetect unusual transactions or estimate riskTransaction histories, applicationsImbalance, changing behavior, fairness
ManufacturingPredict maintenance or defectsSensor streams, inspection imagesSensor failure, delayed labels, safety
Natural languageClassify, search, translate, summarizeText and speechAmbiguous language, bias, changing usage
Computer visionDetect or classify objects/defectsImages and videoLighting, camera, and population shift
TransportForecast demand or estimate travel timeGPS, schedules, weatherSafety, feedback loops, changing routes
AgricultureEstimate yield or detect crop stressWeather, satellite, field sensorsRegional conditions, missing measurements
EducationRecommend practice or identify support needsActivity and assessment recordsPrivacy, label quality, unequal access

These are candidate applications, not claims that an ML model is automatically appropriate. A high-stakes application needs a clear failure policy, monitoring, and human or organizational accountability.

Case study — spam filtering

Task: classify an incoming message as spam or legitimate.

A fixed rule such as “reject every message containing one word” may be enough for a narrow situation. ML becomes useful when patterns are numerous, changing, and difficult to encode manually—but it still needs labelled feedback and monitoring.

Prediction is not explanation

If an application predicts hospital readmission from a set of correlated variables, a high score does not prove that changing one variable will cause readmission to fall. Prediction asks “what tends to improve accuracy?” Causal analysis asks “what would happen if we intervened?” Confusing the two can produce harmful decisions.2

Exercise

For each proposed application, fill in task, target, metric, and one deployment risk:

  1. Predict the repair time for a machine from sensor data.
  2. Recommend elective subjects to a student.
  3. Group customer support messages when no categories exist.
  4. Identify suspicious card transactions.
One reasonable answer
  1. Regression; target repair time; MAE; risk: sensor or operating-condition shift. 2. Ranking/recommendation; target may be clicks, completion, or satisfaction; metric must match the educational goal; risk: reinforcing narrow choices. 3. Unsupervised clustering; no supplied target; metric must be judged for useful, stable groups; risk: arbitrary clusters being treated as real categories. 4. Classification or anomaly detection; target fraud/not fraud if labels exist; precision/recall or a cost-sensitive metric; risk: class imbalance and false declines.

Exam lens

Application answer template: state the domain, task type, inputs, target, output, metric, and risk. Do not merely name an industry.

Common traps:

Rapid revision

Sources

Footnotes

  1. Mitchell, T. M. (1997). Machine Learning. McGraw-Hill — examples of learning tasks and performance measures.

  2. Alpaydın, E. (2020). Introduction to Machine Learning, 4th Ed. MIT Press — application framing, prediction, and learning paradigms. 2