metriset

Metriset is a concept in information science describing a formal collection of evaluation metrics intended to assess system performance in a consistent way. It formalizes which metrics to report, how to compute them, and how to interpret results.

A metriset typically includes core metrics for a given task (e.g., classification: accuracy, precision, recall, F1; regression: MSE, MAE; ranking: NDCG, MAP), along with ancillary measures such as calibration, robustness to distribution shift, fairness across groups, latency, memory usage, and energy consumption. It prescribes data handling practices, such as using a standard test split, cross-validation, or bootstrapping for confidence intervals, and a defined reporting format.

The aim is to enable comparability across studies, graphs, and experiments, by providing a shared metric set and recommended aggregation methods. Researchers can extend a metriset with task-specific metrics for specialized domains (e.g., medical imaging, recommender systems). Implementation typically involves a metrics registry or library.

Proponents argue it improves transparency and reproducibility; critics caution that a fixed metric set can bias evaluation toward certain objectives or overlook context. Metriset is distinct from benchmarks, which provide tasks and datasets, though it often interacts with them by specifying metrics.

Related concepts include evaluation benchmarks, metric registries, and standardized reporting practices that accompany a metriset in research and industry contexts.