kNNi

kNNi is a family of extension methods for the k-Nearest Neighbors classifier and regressor designed to operate on evolving data streams or growing datasets. The "i" generally signals incremental, online, or interactive updates that avoid retraining from scratch as new data arrives.

In a typical kNNi setup, the model maintains a dataset of labeled or unlabeled points along with an index or reservoir to support fast neighbor queries. When a new point arrives, its distance to existing points is computed or approximated, and the nearest neighbors are updated. Depending on the variant, the prediction for a query is formed by aggregating the labels of the k nearest neighbors, with possible weights based on distance, recency, or other relevance signals. Some implementations use data structures such as metric trees, kd-trees, locality-sensitive hashing, or incremental graph structures to speed updates and queries.

Applications include real-time classification, regression on streaming data, anomaly detection, and adaptive user modeling. kNNi can also incorporate drift-detection mechanisms to adjust to changing data distributions, and may use memory limits to keep only a subset of data or compact summaries.

Limitations: like kNN, performance depends on feature scaling and distance metric; high-dimensional data may degrade accuracy; maintaining neighbor sets can be memory-intensive; incremental updates can introduce latency if data arrives rapidly; care is needed to manage concept drift.

kNNi is not a single standard algorithm but a category used in literature and software to describe incremental kNN variants. It is related to online learning and lazy learning paradigms and is often used as a simple baseline in streaming analytics.