LOF Anomaly Detection

What is Local Outlier Factor (LOF) and how does it detect anomalies?

Local Outlier Factor (LOF) is a density-based anomaly detection algorithm that identifies outliers by comparing the local density of a point to the densities of its neighbors. Unlike global methods, LOF can detect outliers in datasets with clusters of varying density — a point may be normal in a sparse cluster but anomalous near a dense cluster.

The algorithm computes the reachability distance and local reachability density (LRD) for each point based on its k-nearest neighbors. The LOF score is the ratio of a point's LRD to the average LRD of its neighbors. LOF ≈ 1 means similar density to neighbors (normal); LOF >> 1 means much lower density (anomaly).

The number of neighbors (k) controls the locality: small k focuses on very local structure, large k smooths over broader neighborhoods. The contamination ratio determines the fraction of points classified as anomalies based on the LOF score ranking.

Formula: reach-dist_k(A,B) = max(k-dist(B), dist(A,B)) LRD_k(A) = 1 / (Σ reach-dist_k(A,Bᵢ) / k) LOF_k(A) = (Σ LRD(Bᵢ) / k) / LRD(A) LOF ≈ 1: normal point LOF >> 1: outlier

Example Calculation

With k=5 neighbors: Point P₁ in a dense cluster has LRD=10, its 5 neighbors average LRD=9.5. LOF = 9.5/10 = 0.95 (normal). Point P₂ near the cluster edge has LRD=2, neighbors average LRD=8. LOF = 8/2 = 4.0 (clear outlier). With contamination=0.05, the top 5% highest LOF scores are flagged.

When to Use This Calculator

• When your data has clusters of varying density — LOF detects local anomalies that global methods like Z-score or Isolation Forest would miss
• During multivariate quality inspection when measurements from different test stations form natural clusters with different spreads
• When analyzing equipment vibration or sensor data where normal operating modes have different characteristic patterns and density
• When outlier detection results from Isolation Forest seem to miss anomalies near dense clusters — LOF's density-based approach handles this case
• During process troubleshooting to identify data points that are unusual relative to their local neighborhood, even if they fall within global ranges

Common Mistakes to Avoid

• Setting k (neighbors) too small, causing sensitivity to noise — very small k values make the algorithm unstable; start with k=20 and adjust based on results
• Using LOF on very high-dimensional data — LOF relies on distance metrics that become less meaningful in high dimensions (curse of dimensionality); consider dimensionality reduction first
• Treating all points with LOF > 1 as anomalies — LOF > 1 only means slightly lower density than neighbors; use the contamination threshold or a meaningful LOF cutoff (e.g., LOF > 2)
• Not scaling features before applying LOF — features with larger numerical ranges dominate the distance calculation; standardize all features to zero mean and unit variance

How to Interpret Results

• LOF ≈ 1.0 — the point has similar density to its neighbors; it is a normal member of its local cluster
• LOF 1.5-3.0 — the point is in a less dense area relative to its neighbors; it may be on the boundary of a cluster or a mild outlier
• LOF > 3.0 — the point has much lower density than its neighbors; it is a strong local outlier and should be investigated
• LOF < 1.0 — the point is in a denser area than its neighbors; this is common for points at the center of tight clusters
• If many points have high LOF scores, the data may not have well-defined clusters — consider whether LOF is the right method for this dataset

Related Standards & References

• Breunig, Kriegel, Ng & Sander (2000), "LOF: Identifying Density-Based Local Outliers", ACM SIGMOD — The original paper defining the LOF score computed here
• ISO 16269-4 — Statistical interpretation of data: detection and treatment of outliers, the standard governing how flagged observations should be handled
• ASTM E178 — Standard practice for dealing with outlying observations, the classical complement for univariate outlier decisions
• Aggarwal, Outlier Analysis (Springer) — Comprehensive reference comparing density-based methods like LOF with isolation- and distance-based alternatives

Frequently Asked Questions

How do I choose the number of neighbors (k)?

Start with k=20 as a reasonable default. Smaller k (5-10) detects very local anomalies but is sensitive to noise. Larger k (50+) provides smoother results but may miss localized anomalies. A practical approach: test multiple k values and check if the same points are consistently flagged — persistent outliers across different k values are more likely true anomalies.

When should I use LOF instead of Isolation Forest?

Use LOF when your data has clusters of different densities, as LOF evaluates each point relative to its local neighborhood. A point far from a dense cluster but close to a sparse cluster may be normal — LOF handles this correctly while global methods might flag it. Use Isolation Forest for high-dimensional data or very large datasets where LOF's O(n²) complexity becomes prohibitive.