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:
- What decision or prediction is required?
- What are the inputs available at decision time?
- What is the target, if the task is supervised?
- How is success measured?
- 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
| Pattern | What the model does | Example |
|---|---|---|
| Regression / forecasting | Predicts a number | Estimate electricity demand or delivery time |
| Classification | Assigns a class or probability | Detect spam or identify a defective product |
| Ranking / recommendation | Orders candidates for a user or system | Rank search results or recommend courses |
| Clustering / discovery | Finds groups without supplied labels | Segment customers or group documents |
| Anomaly detection | Flags unusual observations | Investigate unusual sensor readings |
| Representation / perception | Converts raw signals into useful structure | Recognize objects in images or transcribe speech |
| Decision support | Produces a prediction for a human or policy | Prioritize 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
| Domain | Possible task | Data examples | Risk to examine |
|---|---|---|---|
| Healthcare | Classify risk or assist image interpretation | Records, images, signals | False negatives, privacy, distribution shift |
| Finance | Detect unusual transactions or estimate risk | Transaction histories, applications | Imbalance, changing behavior, fairness |
| Manufacturing | Predict maintenance or defects | Sensor streams, inspection images | Sensor failure, delayed labels, safety |
| Natural language | Classify, search, translate, summarize | Text and speech | Ambiguous language, bias, changing usage |
| Computer vision | Detect or classify objects/defects | Images and video | Lighting, camera, and population shift |
| Transport | Forecast demand or estimate travel time | GPS, schedules, weather | Safety, feedback loops, changing routes |
| Agriculture | Estimate yield or detect crop stress | Weather, satellite, field sensors | Regional conditions, missing measurements |
| Education | Recommend practice or identify support needs | Activity and assessment records | Privacy, 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.
- Inputs: message text, sender information, and metadata available when the message arrives.
- Target: a label from historical moderation or user feedback.
- Model output: probability that the message is spam.
- Decision rule: choose a threshold; messages above it may be filtered, while uncertain messages may be reviewed.
- Metrics: precision matters because incorrectly filtering legitimate mail is costly; recall matters because missed spam is also undesirable.
- Changing environment: spammers adapt, so the data distribution and labels must be monitored.
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:
- Predict the repair time for a machine from sensor data.
- Recommend elective subjects to a student.
- Group customer support messages when no categories exist.
- Identify suspicious card transactions.
One reasonable answer
- 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:
- automation is not automatically machine learning;
- a prediction is not a causal explanation;
- a high offline score is not proof of safe deployment;
- the model output is not always the final business decision.
Rapid revision
- Can I turn an application idea into
inputs → model → output → action? - Can I identify whether the task is regression, classification, ranking, or clustering?
- Can I choose a metric based on the cost of mistakes?
- Can I name one deployment risk for any application?