Core concepts

HUG-IML process

HUG-IML extracts High Utility Gain patterns from labelled tabular data, transforms every row into a binary pattern-presence matrix, and trains an interpretable downstream classifier on that matrix.

The practical modeling sequence is:

Detect or provide feature types: integer, float, and categorical.
Discretize numerical columns into bins.
Build transactions from observed feature-bin/category items.
Mine top High Utility Gain patterns with information-gain filtering.
Transform samples into a sparse binary matrix indicating which patterns fire.
Fit the downstream estimator, logistic regression by default.
Expose pattern labels, support, utility, information gain, coefficients, and local active-pattern explanations.

Pattern anatomy

A learned pattern is readable because it is expressed in the vocabulary of the input data:

glucose=[157.1,177.3) AND bmi=[31.8,39.1)   coef=+1.41 support=0.067
checking_status=no_checking                 coef=+1.12 support=0.390

Useful fields include:

Field	Meaning
`pattern`	Human-readable feature interval/category combination.
`support`	Fraction of training samples where the pattern is active.
`utility`	Pattern utility score used during mining.
`information_gain`	Supervised filtering signal for target relevance.
`coefficient`	Downstream classifier contribution, when available.

Key hyperparameters

Parameter	Role	Typical starting point
`B`	Numerical bin count, or upper bound when adaptive binning is enabled.	`5` to `15`
`L`	Maximum pattern length. `L=1` mines singleton patterns; larger values allow interactions.	`1` for larger datasets, `2` for interaction discovery
`G`	Minimum information gain threshold.	`1e-3` to `1e-5`
`topK`	Maximum number of retained patterns. In the native 1.1.x path this budget is applied inside mining, so it is both an interpretability cap and a performance control. `-1` computes an automatic budget from the mined item universe.	`50` to `300` for compact audits; use `-1` only when the candidate space is known to be manageable
`adaptive_binning`	Selects per-feature bin resolution with supervised information gain.	`True` when tuning `B` is undesirable

Missing values

Numerical NaN and infinite values are treated as not observed. The corresponding item is absent from the transaction, so patterns requiring that feature do not fire for that row. This avoids fabricating values through mean or median imputation.

When missingness is itself meaningful, add explicit binary indicators before fitting:

X_aug = X.copy()
for col in X.columns:
    if X[col].isna().mean() > 0.05:
        X_aug[f"{col}__MISSING"] = X[col].isna().astype(int)

Adaptive binning

Adaptive binning selects a bin count per numerical feature by evaluating candidate B values using supervised information gain and stopping when marginal gains become small.

from hugiml import HUGIMLClassifier

clf = HUGIMLClassifier(
    adaptive_binning=True,
    b_candidates=[2, 3, 5, 7, 10, 15],
    min_marginal_gain_ratio=0.02,
    L=2,
    G=1e-4,
)

Performance and mining behavior

The 1.1.x native path keeps the modeling interface stable but changes where work is bounded internally. The effective topK value is passed into the native mining stage before candidate retention, rather than being treated only as a post-processing cap. This is closer to the original HUGIML Java implementation and is important for larger datasets because fewer non-final candidates need to be materialized.

For L=1 fits, the native hot path fuses transaction preparation, singleton pattern mining, information-gain filtering, top-K retention, and sparse matrix construction. This hot path supports adaptive binning without first materializing a separate binned matrix, which is the recommended route for large adaptive singleton workflows. Set use_hotpath=False only when comparing against the older three-stage path for debugging or benchmarking.

Transaction construction is performed in row-stripe chunks on the non-fused path, and materialized native transactions now store compact item ids with shared item-level utility lookup. The resulting model is intended to match the previous transaction semantics while reducing repeated utility storage and making memory use less bursty. This is most useful for wide data, large batches, and cross-validation loops.

For interaction mining, structured constraints are applied exactly. EUCS pruning is disabled by default because aggressive pair-level pruning can remove valid higher-order patterns in edge cases. Prefer controlling complexity with L, G, and topK first.

Operational stability controls

Use n_jobs=-1 to allow the native backend to use all available OpenMP threads. The adaptive bin-selection and bin-code application stages can use this parallelism before the main mining step, so it benefits adaptive workflows as well as the fused L=1 path.

max_fit_seconds is a wall-clock budget for the native mining stage. If the budget or memory pressure prevents the full configuration from completing, HUGIML attempts safer fallback configurations, records the degraded outcome in fit_metadata_.degraded, and raises a clear HUGIMLTimeoutError or HUGIMLMemoryError only when it cannot recover. Inspect fit_metadata_ after fitting to review pattern counts, stage timings, memory estimates, OpenMP thread count, and whether a fallback was used.

Constant and zero-utility columns

A column that is constant after preprocessing cannot contribute useful pattern utility. The estimator may emit HUGIMLConvergenceWarning for these columns during prepareXy or fit. This warning is expected: the column is ignored for pattern mining, while the rest of the dataset continues to train normally. In production pipelines, drop known constant columns before fitting if you want a quieter log.