Language: English Last updated: 2026-07-14 This page: unsupervised model overview Switch: Chinese
statgpu.unsupervised contains estimators for dimensionality reduction, clustering, density-based grouping, mixture modeling, non-negative matrix factorization, manifold embedding, and approximate nearest neighbor search. The API follows the familiar fit, transform, predict, fit_predict, and score style where those operations make sense for the model.
| Estimator | Main Use | Core Criterion |
|---|---|---|
| PCA | Linear dimensionality reduction | Maximize projected variance / minimize rank-k reconstruction error |
| KMeans | Prototype-based clustering | Minimize squared Euclidean inertia |
| DBSCAN | Density-based clustering with noise | Density reachability and connected components |
| GaussianMixture | Probabilistic soft clustering | Maximize Gaussian mixture log likelihood with EM |
| NMF | Parts-based non-negative factorization | Minimize Frobenius reconstruction error under non-negativity |
| AgglomerativeClustering | Hierarchical clustering | Greedy linkage merges |
| TruncatedSVD | Uncentered low-rank projection | Minimize rank-k dense reconstruction error |
| MiniBatchKMeans | Larger-scale prototype clustering | Approximate inertia minimization with mini-batch updates |
| IncrementalPCA | Batch-wise linear dimensionality reduction | Approximate centered rank-k reconstruction |
| MiniBatchNMF | Larger-scale non-negative factorization | Mini-batch Frobenius reconstruction loss |
| UMAP | Manifold embedding | Fuzzy graph cross-entropy |
| NNDescent | Approximate nearest neighbor search | Iterative neighbor candidate refinement |
| TSNE | Manifold visualization | KL divergence between affinity distributions |
Most unsupervised estimators expose device="auto", "cpu", "cuda", and "torch" following the project-wide device rules. Explicit GPU devices must either run on that backend or raise a clear error; they should not silently fall back to CPU. Some algorithms have narrower support, so check the per-model page before relying on a GPU path.
Dense unsupervised estimators share one backend-aware finite-input check. NaN/Inf is rejected before SVD, eigendecomposition, distance computation, or iterative updates, so users receive a stable public error rather than estimator-specific low-level failures.
Unsupervised estimators do not expose statistical inference fields such as standard errors, p-values, confidence intervals, AIC, or BIC unless the model naturally defines them. For these models, documentation focuses on algorithmic objective, exact versus iterative behavior, device support, and output semantics.
For detailed API behavior and model-specific caveats, continue to the per-model pages linked above.