SBSM Model Parameters

This tutorial describes the parameters of sbsm and suggests appropriate tuning

Constructor

SBSM(
    C: float = 64.0,
    alpha: float = 1.0,
    match_score: int = 2,
    mismatch_score: int = -1,
    gap_score: int = -2,
    k_neighbors: int = 15,
    lambda_: float = 0.9,
    nu1: float = 0.001,
    nu2: float = 0.001,
    process_num: int = 30
)

Parameter Tuning Guide for SBSM

The performance of the SBSM model depends on the configuration of its hyperparameters.y:

💡 We strongly recommend tuning key parameters before using the model in practical applications.

The table below ranks the SBSM hyperparameters by how strongly they influence model performance (higher = more sensitive).

Parameter	Type	Default	Valid Range	Suggested Range	Impact	Description
alpha	float	1.0	0 ~ ∞	0.1 ~ 5.0	🔥🔥🔥🔥🔥	Kernel scaling factor before training. Affects decision boundary sharpness.
C	float	64.0	0 ~ ∞	0.1 ~ 100.0	🔥🔥🔥🔥🔥	Controls trade-off between margin and error.
k_neighbors	int	15	0 ~ len(X_train)	5 ~ min(200, len(X_train)-1)	🔥🔥🔥🔥	Number of neighbors used in kernel weighting. Dataset-size dependent.
lambda_	float	0.9	0 ~ 1	0.1 ~ 0.99	🔥🔥🔥	Decay ratio in global kernel alignment.
nu1	float	0.001	0 ~ ∞	1e-4 ~ 1e-1	🔥🔥	Regularization weight for Laplacian graph term.
nu2	float	0.001	0 ~ ∞	1e-4 ~ 1e-1	🔥🔥	Regularization weight for L2 norm term.
match_score	int	2	0 ~ ∞	1 ~ 5	🔥	Score for matching characters in alignment. Biological prior-based.
mismatch_score	int	-1	-∞ ~ 0	-5 ~ -1	🔥	Penalty for mismatched characters. Should be negative.
gap_score	int	-2	-∞ ~ 0	-10 ~ -1	🔥	Penalty for insertions/deletions in alignment. Should be negative.
process_num	int	30	1 ~ ∞	1 ~ CPU cores	🧊	Number of parallel processes for kernel computation. Impacts speed only.

By following this tuning strategy, you can make SBSM adapt effectively to your dataset and maximize classification performance. For a comprehensive explanation of the parameters, please refer to the SBSM paper.