SVM Models API¶

Both public classes are experimental heuristics. They do not claim equivalence to a published incremental One-Class SVM optimizer.

IncrementalOneClassSVMAdaptiveKernel ¶

Python

IncrementalOneClassSVMAdaptiveKernel(nu: float = 0.1, initial_gamma: float = 1.0, gamma_bounds: tuple[float, float] = (0.001, 100.0), adaptation_rate: float = 0.1, buffer_size: int = 200, sv_budget: int = 100, tolerance: float = 1e-06, seed: int | None = None)

Bases: BaseModel

Experimental incremental one-class kernel model with adaptive gamma.

This is a custom budgeted support-vector heuristic, not an implementation of a published incremental One-Class SVM optimizer. Support vectors and the recent-data buffer are stored in raw feature coordinates. Kernel values are computed after applying the current running standardization to both operands, so all comparisons use one consistent coordinate system.

Parameters:

Name	Type	Description	Default
`nu`	`float`	Heuristic support-vector weight scale in `(0, 1]`. It controls the initial weight and caps weights added for negative margins; it does not carry a solver-backed One-Class SVM `nu` guarantee.	`0.1`
`initial_gamma`	`float`	Initial positive RBF-kernel coefficient.	`1.0`
`gamma_bounds`	`tuple[float, float]`	Positive inclusive lower and upper bounds for adaptive gamma. The interval must contain `initial_gamma`.	`(0.001, 100.0)`
`adaptation_rate`	`float`	Interpolation fraction in `(0, 1]` toward the median-distance gamma estimate at each adaptation step.	`0.1`
`buffer_size`	`int`	Maximum recent raw samples used to estimate gamma.	`200`
`sv_budget`	`int`	Maximum retained support vectors.	`100`
`tolerance`	`float`	Non-negative margin tolerance used for support-vector insertion and `predict_one`.	`1e-06`
`seed`	`int \| None`	Seed for the model-local subsampling generator used by gamma estimation.	`None`

learn_one ¶

Python

learn_one(x: dict[str, float]) -> None

Incrementally learn from one sample.

predict_one ¶

Python

predict_one(x: dict[str, float]) -> int

Predict normal (1) or anomalous (-1).

score_one ¶

Python

score_one(x: dict[str, float]) -> float

Compute anomaly score (higher means more anomalous).

get_model_info ¶

Python

get_model_info() -> dict[str, object]

Return current model diagnostics.

GraphGatedOneClassSVM ¶

Python

GraphGatedOneClassSVM(graph: dict[int, list[int]] | None = None, threshold: float = 0.0, learning_rate: float = 0.01, nu: float = 0.5, lambda_reg: float = 0.01)

Bases: BaseModel

Graph-gated ensemble of incremental linear one-class SVM heuristics.

The graph determines which node models are updated and scored. This custom anomaly detector is unrelated to the published GADGET distributed averaging and optimization algorithm.

Traversal starts at nodes with no incoming edge. A node's outgoing edges are traversed only when its local score exceeds threshold. Each visited node owns an incremental linear hinge-style heuristic.

Parameters:

Name	Type	Description	Default
`graph`	`dict[int, list[int]] \| None`	Directed adjacency mapping from integer node identifiers to child-node lists. `None` uses the chain `0 -> 1 -> 2`. The mapping and child lists are copied.	`None`
`threshold`	`float`	Local score gate for traversing outgoing edges.	`0.0`
`learning_rate`	`float`	Step size used by every node model.	`0.01`
`nu`	`float`	Coefficient in every node model's bias update. This custom update does not provide the guarantees of a solved One-Class SVM.	`0.5`
`lambda_reg`	`float`	Weight-decay coefficient used by every node model.	`0.01`