Contact

k-Nearest Neighbors (kNN)

A k-Nearest Neighbors (kNN) model predicts a property by finding the most structurally similar molecules already present in its training data — in PAULY’s models, similarity is measured via a Jaccard metric over Atom-Centered Fragment (ACF) counts — and averaging their known values.

Because it makes no assumption about the underlying mathematical relationship between structure and property, a kNN model is robust and easy to validate (PAULY’s kNN models are evaluated via leave-one-out cross-validation). Its accuracy, however, depends entirely on the training set actually containing molecules similar enough to the one being predicted — for a genuinely novel structure, a kNN model’s prediction is only as good as its nearest available neighbour.

Models using this approach