SpikeSlabHeredityClassification¶
Classify data instances using sparse Bayesian logistic regression with explicit main effects and interaction terms under a heredity-structured spike-and-slab prior.
The node reuses the existing coordinate-ascent variational Bayes core from the main-effects-only Spike-Slab VB classifier while expanding the design with pairwise or three-way interaction columns and recomputing the interaction prior probabilities from the current main-effect inclusion probabilities at each sweep. The default configuration uses weak heredity, centered pairwise products, and lasso screening before interaction expansion, which is the recommended practical setting for EEG-like data where interactions may be meaningful but should remain sparser than main effects. Interaction posterior inclusion probabilities are exposed alongside the main-effect probabilities in the fitted weight packet. More Info... Version 0.7.0
Ports/Properties¶
data¶
Data to process.
cond_field¶
The name of the instance data field that contains the conditions to be discriminated. This parameter is ignored if the packet has previously been processed by a DescribeStatisticalDesign node.
feature_scaling¶
Feature scaling to use. The spike-and-slab prior is scale-sensitive, so scaling the features is recommended for proper shrinkage behavior. The default of 'std' (standardization to zero mean and unit variance) is a good choice for most data. The 'robust' option uses median and interquartile range, which is more resilient to outliers.
sample_size_factor¶
Sample size correction factor or effective sample size. When observations are not independent (e.g., multiple trials per participant), the effective information content may be considerably lower than the nominal sample count, causing the posterior to be overconfident. This parameter tempers the data-dependent terms of the variational bound by the factor n_eff/n, reducing the data's influence on the posterior to match the effective information content. Values between 0 and 1 (inclusive) are interpreted as a multiplicative factor applied to the actual number of observations; for instance, if you have 2000 trials from 30 participants, set this to 30/2000 = 0.015. Values > 1 must be integers and are interpreted as an absolute effective sample size; for instance, 30 means that the effective sample size is 30 regardless of the actual number of observations.
expected_relevant_features¶
Expected number or fraction of relevant features. Controls the prior inclusion probability pi = p0/p, where p0 is the expected relevant count. Values between 0 and 1 are interpreted as a fraction of the total feature count; values greater than or equal to 1 must be integers and are interpreted as an absolute count. The default of 0.1 works well for moderate-dimensional problems (p up to ~2000). For very high-dimensional datasets (p > 5000), values of 0.02-0.05 are recommended to prevent over-selection of noise features.
slab_scale¶
Scale of the slab (relevant-feature) component. Controls the prior width of the non-zero coefficients on standardized features: larger values allow relevant features to take on larger coefficient values, while smaller values constrain them to stay closer to zero. The default of 1.0 works robustly across a wide range of problems; empirically, values from 0.5 to 2.0 produce nearly identical results on most datasets.
include_bias¶
Include a bias (intercept) term. If disabled, input features should be centered or augmented with a constant feature.
learn_inclusion_probability¶
Learn the global prior inclusion probability from the data during fitting. When enabled, the method adapts how many features it expects to be relevant based on what the data supports. When disabled (the default), the inclusion probability is fixed at the value implied by the expected relevant features parameter.
learn_slab_scale¶
Learn the slab scale from the data during fitting. When enabled, the method adapts the expected magnitude of relevant coefficients based on the data. When disabled (the default), the slab scale is fixed at its configured value throughout fitting.
support_threshold¶
Threshold on the estimated probability that a feature is relevant. Features whose inclusion probability meets or exceeds this threshold are reported as active in the support mask. The default of 0.5 selects features that are more likely relevant than not (the median probability model).
max_iterations¶
Maximum number of coordinate-ascent (CAVI) sweeps. Convergence typically occurs within 100-200 iterations for most problems; the default of 500 is a generous safety margin.
convergence_tolerance¶
Convergence tolerance on the maximum change in posterior inclusion probabilities or posterior mean coefficients over a full sweep. The default of 1e-5 yields tight convergence; values up to 1e-4 are acceptable for faster fitting with minimal accuracy loss.
laplace_solver¶
Solver for the inner CAVI coordinate sweep. 'scipy' uses scipy.optimize.minimize with L-BFGS-B for the 2D Laplace subproblem (the reference implementation). 'newton' replaces that with a hand-rolled 2D damped Newton iteration using the closed-form Hessian, which is numerically equivalent but avoids scipy's per-call overhead. 'numba' additionally JIT-compiles the entire per-feature sweep (x_appm updates, coef_lin dot product, Newton iteration, alpha update) as a single fused kernel, which is the fastest strictly-IEEE option for moderate to large feature counts. 'numba_fastmath' is the same kernel built with LLVM fast-math (reassoc, reciprocal, finite-only); it is typically 20-50% faster again but permits small reorderings at the ~1e-14 level. All numba variants pay a one-time JIT compilation cost on first use in a process (cached to disk).
random_seed¶
Random seed. Different values may give slightly different outcomes.
initialize_once¶
Calibrate the model only once. If set to False, then this node recalibrates whenever a non-streaming data chunk is received that has both training labels and associated training instances.
dont_reset_model¶
Do not reset the model when the preceding graph is changed.
probabilistic¶
Use probabilistic outputs. If enabled, the node will output for each class the predicted probability that a given trial belongs to that class.
verbosity¶
Verbosity level. Higher values print more fitting diagnostics.
heredity¶
The heredity structure coupling interaction inclusion to its parent main effects. 'strong' requires all parent mains to be active before an interaction can enter. 'weak' allows an interaction when at least one parent is active. 'none' removes heredity coupling and uses a plain spike-and-slab prior on the expanded design. 'custom' uses the probabilities supplied in custom_heredity_probabilities and currently supports pairwise interactions only.
custom_heredity_probabilities¶
Custom pairwise heredity probabilities (a_0, a_1, a_2, a_3) used when heredity is set to 'custom'. These correspond to the probabilities of including an interaction when neither parent is active, only the first parent is active, only the second parent is active, or both parents are active.
interaction_form¶
How to construct interaction feature columns. 'centered_product' multiplies training-centered parent features, reducing overlap between main and interaction effects. 'orthogonalised_product' additionally removes the projection onto the parent mains and the intercept so the interaction column captures only non-additive structure.
interaction_order¶
Maximum interaction order. A value of 1 disables interaction expansion and reduces the node to the main-effects-only spike-and-slab VB classifier. A value of 2 adds pairwise interactions, and 3 additionally adds three-way interactions.
include_self_interactions¶
Include squared terms alongside the main effects. These act as simple quadratic terms whose inclusion depends on the corresponding main effect.
main_effect_screening¶
Optional screening applied to the main effects before interaction expansion. The default 'lasso' reuses the package-like lasso warm start to keep the expanded design tractable in moderate-to-high-dimensional problems.
main_effect_screening_k¶
Optional cap on the number of main effects retained for interaction expansion. When omitted with lasso screening, all non-zero lasso-selected mains are kept.
vb_prescreen¶
Run an additional main-effects-only spike-and-slab VB pass after screening and keep only features whose posterior inclusion probability is at least 0.5 . This is more principled than lasso-only screening but costs an extra fit.
interaction_inclusion_scale¶
Multiplier mapping the main-effect prior inclusion probability to the baseline interaction inclusion probability. Interactions are therefore encouraged to be sparser than main effects by default.
interaction_slab_scale¶
Multiplier mapping the main-effect slab scale to the slab scale used for interaction terms.
set_breakpoint¶
Set a breakpoint on this node. If this is enabled, your debugger (if one is attached) will trigger a breakpoint.
metadata¶
User-definable meta-data associated with the node. Usually reserved for technical purposes.