instancecentric

Instancecentric is a term used in machine learning and data analysis to describe approaches, perspectives, or methodologies that emphasize individual data points, rather than focusing primarily on global aggregates, classes, or distributions. The term is not part of a single standardized framework, but is often used to contrast with class-centric or category-centric viewpoints that prioritize group-level patterns or aggregated statistics.

In practice, instance-centric approaches may involve per-instance decisions, losses, or explanations. Examples include instance-based learning methods, such as k-nearest neighbors, where predictions are derived from a subset of training examples closest to a query instance. Other instance-centric techniques include per-instance loss weighting or calibration, where each data point contributes to optimization or confidence estimates according to its own characteristics or difficulty. In explainability, instance-specific explanations aim to justify a prediction for a single input rather than providing global model insights.

Applications of instance-centric thinking appear in anomaly detection, where individual outliers are evaluated on their own merits; active learning, which selects specific data points for labeling; and calibration tasks that produce per-instance confidence scores. It can also influence dataset curation and sampling strategies that target informative or representative instances rather than broad class-level coverage.

Because the term is used variably across literature, its exact meaning is context-dependent. It is often described as a complementary perspective to class-centric methods, highlighting the value of treating data points as unique entities within learning and evaluation processes.