Local Outlier Factor

LOF is a density-based anomaly score. For each observation, it finds a neighborhood of k nearest points and estimates local reachability density, which reflects how closely packed the point is relative to its neighbors. LOF compares the average local reachability density of neighbors with the point's own density. A score near one indicates comparable local density; a substantially larger score indicates that the point is less dense than nearby observations. It is a relative local comparison rather than a global distance-from-center measure. This locality can surface anomalies that global methods miss. A point can be unusual in a dense neighborhood while still lying in a region where the overall dataset is broad. Conversely, a point far from the main cloud may belong to a distinct dense subgroup and receive an ordinary local score. Whether that subgroup should count as anomalous depends on the task. The neighbor count controls the scale of comparison. Small neighborhoods are sensitive to fine structure and noise; large ones may compare across separate groups. Distance metric and feature units are important because LOF uses nearest neighbors. Scale numerical variables appropriately and encode categories without inventing misleading numeric order. Choose settings based on expected group sizes and evaluate score behavior, ideally with reviewed cases. Scikit-learn distinguishes outlier detection, where LOF identifies unusual observations in the fitted dataset, from novelty detection, where a model is fitted on reference data and used on new points. The latter requires novelty=True, and the documentation warns against using the novelty scoring methods on the training data in that mode. LOF scores are not calibrated probabilities and may use inverted sign conventions in APIs. Inspect whether higher or lower values indicate greater abnormality. A neighborhood anomaly is evidence relative to a selected reference set and metric, not a definitive judgment about data quality or risk.