Examples¶

The repository's scripts are complete programs rather than fragments. Run them from the repository root so relative project configuration and the local package checkout are available.

uvpip

Bash

uv sync --extra eval
uv run python examples/models/online_iforest.py

Bash

python -m pip install -e ".[eval]"
python examples/models/online_iforest.py

Most model scripts compute average precision and ROC AUC and therefore require the eval extra. Dataset-backed scripts can download and cache a registered NPZ artifact on first use.

General numeric streams¶

Family	Example	Additional requirement
Isolation forest	`online_iforest.py`	Registered SHUTTLE dataset
Isolation forest	`mondrian_isolation_forest.py`	Registered SHUTTLE dataset
Isolation forest	`asd_iforest.py`	Registered SHUTTLE dataset
Isolation forest	`streamRHF.py`	Registered SHUTTLE dataset
Isolation forest	`xstream.py`	Registered SHUTTLE dataset
Random cut forest	`random_cut_forest.py`	Registered SHUTTLE dataset
Half-space trees	`half_space_trees.py`	Registered SHUTTLE dataset
Local density	`local_outlier_factor.py`	Registered SHUTTLE dataset
Cell neighborhood	`cell_neighborhood.py`	Registered SHUTTLE dataset
Observer distance	`sdostream.py`	Registered SHUTTLE dataset
Stationary region	`stationary_region_neighbor.py`	Registered SHUTTLE dataset
Multi-aspect sketch	`mstream.py`	Registered SHUTTLE dataset
Projection histogram	`streaming_loda.py`	Registered SHUTTLE dataset
Randomized hashing	`streaming_rshash.py`	Registered SHUTTLE dataset

Graph and time-series streams¶

Model	Example	Data source
AnoEdge-L	`anoedge.py`	Maps two SHUTTLE features to integer-like edge identifiers for demonstration
ISCONNA	`isconna.py`	Seeded synthetic edge stream
MIDAS-R	`midas.py`	Seeded synthetic edge stream
Signed graph sketch	`signed_graph_sketch.py`	Seeded synthetic multi-graph stream
X-Lag DAMP	`xlag_damp.py`	Seeded synthetic periodic series with an injected discord; no `eval` extra needed

The AnoEdge example is an API demonstration, not a claim that tabular SHUTTLE rows are a scientifically meaningful graph benchmark.

Experimental, reconstruction, and pipeline examples¶

Example	Purpose	Additional requirement
`adaptive_svm.py`	Budgeted adaptive-kernel SVM heuristic	Registered SHUTTLE dataset
`graph_gated_svm.py`	Graph-gated linear SVM heuristic	Registered FRAUD dataset
`online_autoencoder_ensemble.py`	NumPy online autoencoder ensemble	Registered SHUTTLE dataset
`autoencoder.py`	User-supplied PyTorch architecture and optimizer	`aberrant[dl,eval]` and registered SHUTTLE dataset
`knn.py`	FAISS-backed KNN distance	`aberrant[faiss,eval]` and registered SHUTTLE dataset
`pipeline.py`	Scaler/KNN versus scaler/PCA/KNN	`aberrant[faiss,eval]` and registered SHUTTLE dataset

Read each script's warm-up and learning policy before comparing its metrics. Several demonstrations intentionally warm up only on labeled-normal rows; that is a curated-normal protocol and is not equivalent to fully unsupervised test-then-train learning.