Using CONN

CONN evaluates connectivity between prototypes rather than relying only on centroid distances. Its batch and incremental modes therefore require an explicit choice of prototype backend.

Batch mode

The default backend is MiniBatchKMeans and supports batch input only:

import cvi

index = cvi.CONN(
    model_type="MiniBatchKMeans",
    kmeans_k=8,
    kmeans_kwargs={"random_state": 0, "n_init": 10},
)
value = index.get_cvi(samples, labels)

model_type="KMeans" selects ordinary scikit-learn KMeans. For either KMeans backend, kmeans_k may be a positive integer applied to every input label or a dictionary keyed by the original integer labels. Counts larger than a label’s sample count are capped automatically. Do not pass n_clusters in kmeans_kwargs; configure it through kmeans_k.

With the default normalize_batch=True, each feature is min-max normalized over the full batch before prototypes are fitted. Disable this only when the data are already on the intended scale.

Incremental mode

Incremental updates require the FuzzyART backend:

index = cvi.CONN(model_type="Fuzzy")

for sample, label in stream:
    value = index.get_cvi(sample, int(label))

Incremental samples are not normalized by the class because future feature bounds are unknown. They must normally already lie in [0, 1]. The default check_incremental_normalized=True validates that assumption; disabling the check does not normalize the samples.

The FuzzyART parameters rho, alpha, beta, and match_tracking control prototype formation and are passed to the underlying ART model. Streamorder and those parameters can therefore affect the learned prototypes and the resulting criterion trajectory.

Limitations

CONN does not currently implement cvi.CVI.remove() or cvi.CVI.merge(). A KMeans-backed object also rejects incremental samples. Use a new object when changing backend or evaluating another independent partition.