sheaf

Hierarchical structure: community detection (Leiden, Louvain), reconciliation across forecast hierarchies, and conformal prediction that respects hierarchy constraints.

Problem

Given a set of points or a graph, you want to find groups: k-means for embeddings, Leiden for networks, hierarchical clustering for dendrograms. When predictions have hierarchical structure (e.g., forecasts at country/region/city levels), they should be coherent -- the parts should sum to the whole. Conformal prediction can provide calibrated intervals that respect this structure.

This library provides the algorithms. It is domain-agnostic.

Examples

See examples/README.md for the full runnable example map.

Embeddings to communities. Build a kNN graph from 2D points and detect clusters via Leiden:

cargo run --example embedding_clustering --features knn-graph

Forecast reconciliation. Given a tree of predictions, reconcile them so they are structurally coherent, then produce calibrated prediction intervals:

cargo run --example forecast_reconciliation

The clustering-evaluation metrics check whether two point sets share the same cluster structure: for example, whether a generated set of earthquake locations forms the same geographic clusters as the USGS catalog.

What it provides

• Clustering: k-means, DBSCAN, hierarchical clustering.
• Community detection: kNN graph construction (feature-gated), Leiden/Louvain/label propagation.
• Hierarchy + conformal: hierarchical reconciliation, split conformal prediction with coherence guarantees.
• Metrics: clustering evaluation helpers.

Custom distance metrics

K-means and DBSCAN accept a pluggable distance metric via with_metric. Built-in metrics re-exported from clump: Euclidean, SquaredEuclidean, CosineDistance, InnerProductDistance. Implement DistanceMetric for your own.

use sheaf::{Kmeans, CosineDistance};

let km = Kmeans::with_metric(8, CosineDistance)
    .with_seed(42)
    .with_seeding_alpha(2.0); // oversampling factor for k-means++
let result = km.fit(&data)?;

use sheaf::cluster::{Dbscan, CosineDistance};

// epsilon is compared against cosine distance (range [0, 2])
let db = Dbscan::with_metric(0.3, 5, CosineDistance);
let labels = db.fit(&data);

Usage

[dependencies]
sheaf = "0.1"

use sheaf::{HierarchicalConformal, HierarchyTree, ReconciliationMethod};

// Build hierarchy, get summing matrix
let h_tree = HierarchyTree::from_raptor(&tree);
let s = h_tree.summing_matrix();

// Calibrate on held-out data
let mut cp = HierarchicalConformal::new(s, ReconciliationMethod::Ols);
cp.calibrate(&y_calib, &y_hat_calib, 0.1)?; // 90% coverage

// Coherent prediction intervals
let (lower, upper) = cp.predict_intervals(&y_hat_test)?;

References

• Principato et al. (2024). "Conformal Prediction for Hierarchical Data."
• Qiu & Li (2015). "IT-Dendrogram: A new representation for hierarchical clustering."
• Sarthi et al. (2024). "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval."

License

MIT OR Apache-2.0