distancei

Distancei is a context-aware dissimilarity measure used in data analysis and machine learning to quantify how different two points are, with the measure adapted to the specific context or task. Unlike fixed distance metrics, distancei allows the importance of features to vary by context, enabling more flexible similarity evaluation across domains such as clustering, retrieval, and anomaly detection.

Formally, for a context i, distancei between x and y in R^n is defined as D_i(x,y) = sqrt( (x - y)^T A_i (x - y) ), where A_i is a positive semidefinite matrix that encodes feature relevance under context i. Equivalently, distancei can be written as sum_k w_k,i |x_k - y_k|^p with p ≥ 1, where the weights w_k,i are derived from the eigenstructure or a learned metric model.

Properties: If A_i is positive semidefinite, D_i is a metric, satisfying non-negativity, identity of indiscernibles, symmetry, and the triangle inequality. When A_i varies with i, distancei becomes a family of contextual distances; if A_i is not PSD, some metric properties may fail.

Learning and usage: Contextual matrices A_i or weights w_k,i are typically learned from labeled data using metric learning methods, cross-validation, or task-specific optimization. Once learned, distancei supports nearest-neighbor search, clustering with context-sensitive distance, and similarity-based recommendations.

Examples and limitations: In a dataset with two relevant features, A_i could emphasize the first feature in one context and the second in another. Limitations include the need for sufficient data to learn A_i and potential overfitting, plus added computational cost.

See also: distance, metric, Mahalanobis distance, metric learning, contextual similarity.

---