Density-based outlier detection identifies observations in sparsely populated regions of feature space. Unlike methods focused on a point’s absolute distance, it estimates how densely neighboring observations are distributed; points in low-density regions receive higher outlier scores. section 3.2.2
Common density estimates include:
- Multidimensional Gaussian scoring: estimate each feature’s mean and standard deviation, convert values to probabilities, multiply feature probabilities, and flag rows with unusually low resulting probabilities. section 3.2.2
These methods generally require numeric features, numerical encoding for categorical variables, and comparable feature scales through min-max scaling or standardization. Gaussian scoring additionally assumes approximately Gaussian, independent features; multiplying probabilities becomes especially fragile as dimensions increase. section 3.2 section 3.2.2
Robot-navigation sensor data
A two-dimensional KDE of features V6 and V19 can visualize the estimated density, while scikit-learn’s KernelDensity class can score the same observations and highlight the top 30 outliers.
