Outliers can often be detected using clustering methods because they either form small, distinct groups or are isolated from major clusters without strict statistical assumptions.
| Method | Key Assumption | Strengths | Weaknesses | Typical Use Case |
|---|---|---|---|---|
| DBSCAN | Clusters are areas of high density separated by low-density regions | - No need to specify number of clusters- Can find arbitrarily shaped clusters- Explicitly identifies noise (anomalies) | - Struggles with varying densities- Sensitive to parameter choice (epsilon, minPoints) | Spatial data clustering and density-based anomaly detection |
| Isolation Forest | Anomalies are easier to isolate via random splits | - Efficient on large, high-dimensional datasets- Requires fewer assumptions- Scales well with data size | - Not suited for small datasets- Less interpretable than density-based methods | High-dimensional tabular data anomaly detection |
| Local Outlier Factor (LOF) | Anomalies have significantly lower local density compared to neighbors | - Good for local anomaly detection- Adapts to density variations | - Sensitive to choice of k (number of neighbors)- Poor performance on high-dimensional data | Detecting subtle anomalies in medium-sized tabular datasets |
DBSCAN
Purpose:
- Detects anomalies based on density rather than explicit statistical models.
Key Idea:
- Points in low-density regions are identified as anomalies.
Steps:
- Cluster data points based on neighborhood density.
- Identify points that do not belong to any dense cluster (noise) as anomalies.
Local Outlier Factor (LOF)
Purpose:
- A density-based method that identifies anomalies by comparing a point’s local density to its neighbors.
Key Idea:
- Points with substantially lower local density than their neighbors are considered outliers.
Steps:
- For each point, compute local density based on its k-nearest neighbors.
- Calculate the LOF score: a value higher than 1 suggests potential anomaly.
- Points with high LOF scores are flagged as anomalies.