DBSCAN クラスタリング

DBSCAN means Density-Based Spatial Clustering of Applications with Noise. It defines local neighborhoods using a radius epsilon and a minimum number of points min_samples. A core point has enough observations in its neighborhood to meet the threshold. A cluster grows by connecting density-reachable core points. Points near a core point but with too few neighbors to be core can be border points. Observations not assigned to a cluster are treated as noise or outliers for this run. Unlike k-means, DBSCAN does not require the number of clusters as an input and can identify curved or irregularly shaped dense regions. Its notion of density depends on the distance metric, feature scaling and parameters. A too-small epsilon may label many points as noise; a too-large epsilon may merge nearby groups. Increasing min_samples generally demands denser support for core status. Parameter choice should reflect meaningful neighborhood scale and be inspected with domain knowledge, not chosen solely to obtain an attractive number of clusters. A single global density threshold can struggle when one genuine cluster is much less dense than another. High-dimensional distance concentration can also weaken neighborhood intuition. The outcome may vary with distance metric and feature representation. In scikit-learn, label -1 denotes noise, and border points associated with multiple clusters can lead to implementation-dependent assignment details. DBSCAN's noise label does not mean a point is erroneous, dangerous or permanently outside every cluster; it means the point was not assigned under this metric and parameterization. Evaluate cluster stability across reasonable settings, inspect how many points are noise and whether clusters make sense for the task. If every point must be assigned or cluster densities vary substantially, compare with other methods. Unlike centroid-based methods, DBSCAN does not naturally provide a prediction rule for assigning arbitrary new points without additional design. Document scaling, metric, epsilon and min_samples so results can be reproduced.