An evaluation metric is a quantitative measure used to assess the performance of a machine learning model, software system, or business process. In AI and data science, evaluation metrics provide objective criteria for comparing models, tuning hyperparameters, and determining whether a solution meets its intended goals. Choosing the right metric matters because it directly shapes how a model is optimized and what trade-offs are prioritized.
For classification tasks, common evaluation metrics include accuracy, precision, recall, F1 score, and area under the ROC curve (AUC-ROC). Regression problems typically rely on mean absolute error (MAE), mean squared error (MSE), and R-squared. In ranking and recommendation systems, metrics like normalized discounted cumulative gain (NDCG) and mean average precision (MAP) measure how well results are ordered. Each metric highlights different aspects of model performance, and no single metric tells the whole story.
Beyond machine learning, evaluation metrics are used throughout software engineering and product development. Teams track metrics such as latency (p50, p95, p99), error rates, throughput, and uptime to gauge system health, while product teams monitor conversion rates, user retention, and engagement scores to evaluate feature success. The key is selecting metrics that align with the specific outcomes your stakeholders care about.
Using evaluation metrics well means understanding their limitations, avoiding vanity metrics that look impressive but don't lead to action, and establishing baselines for meaningful comparison. A disciplined approach to metrics helps teams iterate faster and make decisions based on evidence rather than gut feeling.