Agrupamento Espectral

Spectral clustering starts by representing pairwise similarity as a graph. Observations are nodes, and edges connect similar points with weights that reflect affinity. A graph Laplacian summarizes this connectivity. Eigenvectors associated with selected Laplacian eigenvalues provide a lower-dimensional representation in which graph-connected groups may be easier to separate. A final step assigns cluster labels, often with k-means or another method. The method can handle non-convex structure because it uses graph connectivity rather than relying only on distances to a centroid in the original feature space. For two nested rings, local similarities can encode each ring as a connected group even though the groups are not linearly separable by a simple center-based partition. The eigenvector embedding helps reveal that structure. The number of clusters is usually supplied, so the algorithm does not remove model-selection decisions. The affinity graph is a central modeling choice. It may use an RBF kernel that decreases with squared distance, a nearest-neighbor graph, or a precomputed symmetric similarity matrix. Kernel width or neighbor count controls graph locality. If edges are too sparse, a group can fragment; if too dense, distinct regions become connected. Feature scaling and distance metric affect which pairs are considered similar. A graph with multiple disconnected components can also change the interpretation of normalized-cut clustering. Spectral methods require eigenvalue computations and can be costly for large dense affinity matrices. Sparse graphs and suitable solvers can help, but computational convenience should not determine the affinity alone. A final label-assignment step introduces additional choices and possibly random initialization. Evaluate stability across reasonable graph settings, compare held-out or domain-based usefulness, and inspect sensitivity to initialization. The eigenvectors are a representation of the graph, not direct semantic explanations of groups. Spectral clustering can find useful structure, but it does not guarantee that the chosen clusters correspond to real categories.