Ensembles and multimodal analysis

Ensemble search

mll.Ensemble controls candidate construction and aggregation.


ENSEMBLE = mll.Ensemble(
    max_sizes=(3,),
    selection_strategies=(
        "top_k",
        "best_per_resolution",
        "best_per_learner_type",
        "caruana",
        "super_learner",
    ),
    aggregation_strategies=(
        "mean_proba",
        "weighted_mean_proba",
        "median_proba",
    ),
    optimize_metric="log_loss",
)

Selection is based on inner or training-side out-of-fold evidence. Outer-test performance does not choose MPMA-E.

Multimodal data

For multimodal analyses, define one sample table and one or more modalities.


SAMPLES = mll.Samples(
    path="samples.tsv",
    sample_id_col="sample_id",
    target_col="label",
    group_col="participant_id",
)
 
MODALITIES = (
    mll.Modality(
        name="microbiome",
        path="microbiome.tsv",
        format="table",
        primary=True,
        sample_id_col="sample_id",
        feature_prefixes=("g__", "s__"),
    ),
    mll.Modality(
        name="clinical",
        path="clinical.tsv",
        format="table",
        sample_id_col="sample_id",
        feature_cols=("age", "bmi"),
    ),
)

Feature columns should be declared explicitly for general tabular modalities unless the table has been verified to contain only features.

Integration strategies

mll.Integration supports several stages of multimodal integration:

Stage	Examples
Unimodal	evaluate one modality independently
Early	concatenate modality features
Intermediate	modality-specific PCA or joint PCA before modelling
Late	combine modality predictions, including weighted or Super Learner strategies

Example:


INTEGRATIONS = (
    mll.Integration("unimodal"),
    mll.Integration("early_concat"),
    mll.Integration("intermediate_modality_pca", n_components=(5, 10)),
    mll.Integration("late_weighted_mean_proba"),
)

Integration models that learn a representation are fitted within the corresponding training partition.