Issues in Machine Learning

1.3 — Issues in Machine Learning

Recall first. A model reports 99% accuracy for detecting a rare disease. Before celebrating, write three questions you would ask about the data, evaluation, and deployment context.

The master diagnostic

A machine-learning project can fail even when the algorithm is implemented correctly. Diagnose it through four questions:

1. Data: Are the examples representative, valid, labelled consistently, and available at prediction time?
2. Model: Can the chosen representation and model express the useful relationship?
3. Evaluation: Does the metric and split measure the real objective without leakage?
4. Deployment: Will the environment, users, costs, and risks remain similar after release?

The “issues in ML” topic is therefore broader than overfitting. Overfitting is one failure mode inside a larger data–model–evaluation–deployment system.¹²

Data issues

Insufficient or unrepresentative data

A large dataset can still be poor if it misses important groups, conditions, seasons, or failure cases. A model trained only on sunny-day images may not generalize to night-time images; more copies of the same sunny examples do not fix coverage.

Ask:

• Who or what generated the examples?
• Which cases are missing?
• Are the deployment cases drawn from a similar population?
• Does time change the distribution?

Missing, noisy, or inconsistent values

Missing measurements may be random, systematic, or informative. Sensor errors and outliers can distort a model. Labels can disagree because different annotators use different definitions. Cleaning is not automatically harmless: deleting unusual records may delete the very failures the system must detect.

Class imbalance

If 99 of 100 records are negative, a model that always predicts negative has 99% accuracy but zero recall for the positive class. The metric must reflect the cost of the mistakes, and evaluation should report more than one number when the classes are imbalanced.³

Data leakage

Leakage occurs when information unavailable at the intended prediction time enters training or evaluation. Examples:

• using a final exam score to predict whether a student will pass;
• calculating a feature from the outcome being predicted;
• fitting a scaler, imputer, or feature selector on the entire dataset before splitting;
• allowing near-duplicate records from one person to appear in both training and test sets.

Leakage can produce an impressive score that collapses in deployment. It is an evaluation failure, not evidence that the model has learned a useful causal signal.

Model and representation issues

• Irrelevant features: add noise and computation without useful signal.
• Too few or weak features: the target may not be predictable from the available inputs.
• Wrong model capacity: a model may be too simple or unnecessarily complex.
• Objective mismatch: optimizing average error may be wrong when rare severe errors matter more.
• Interpretability and actionability: a high score does not tell a decision-maker what action to take or whether the relationship is causal.
• Computation and reproducibility: training may be too slow, expensive, or difficult to reproduce with the same data and settings.

These issues interact. A more expressive model cannot recover information that the features never contain.

Evaluation and deployment issues

Distribution shift and concept drift

The training distribution may differ from the future distribution. User behavior, policies, sensors, prices, or language can change. A static test score is evidence about the test distribution, not a lifetime guarantee.

Metric choice

A metric encodes what counts as a good prediction. MAE treats errors linearly; MSE penalizes larger errors more. Accuracy treats all classification mistakes equally, which is often inappropriate for rare or high-cost events. Choose the metric before tuning toward a score, or the project can optimize the wrong thing.

Privacy, fairness, and security

Data may contain sensitive information or proxies for protected characteristics. Historical decisions may reflect unequal treatment. Attackers may manipulate inputs or extract information from a model. These are system-design and governance issues, not problems solved merely by increasing accuracy.

Worked diagnosis — a loan-default model

A model reports 98% accuracy. Investigation finds:

1. 95% of loans did not default.
2. A feature records whether the account entered collections after the loan was issued.
3. The data is split randomly, although the economic conditions changed over ten years.
4. Income is missing more often for one demographic group.

Diagnosis:

• Accuracy is weak evidence because the classes are imbalanced.
• The collections feature is leakage: it is not available at decision time.
• A time-aware split may be needed to estimate future performance.
• Missingness may be a coverage, fairness, or measurement issue requiring investigation rather than silent deletion.

The algorithm might be perfectly coded and the project would still be untrustworthy.

Exercise

A factory-failure dataset contains many normal operating records, very few failures, and multiple readings from the same machine. A random row-level split gives excellent test accuracy. List at least four issues and one remedy for each.

Exam lens

Framework: answer “issues in ML” under data, model, evaluation, deployment, and ethics. For every issue, state the failure mechanism, its consequence, and a mitigation.

High-value distinctions:

• data quality vs. model complexity;
• overfitting vs. leakage;
• class imbalance vs. low overall accuracy;
• correlation/prediction vs. causation;
• test performance vs. guaranteed future performance.

Rapid revision

• Can I explain why 99% accuracy may be useless?
• Can I identify a feature that leaks future information?
• Can I explain why a random split can be wrong for grouped or time-dependent data?
• Can I connect each issue to a concrete remedy?

Sources