LIFTS

LIFTS is a package version of the latent factor imputation workflow developed in this repository. It keeps the original FBI implementation available internally, but exposes a cleaner estimator-style API for simulation experiments and downstream packaging.

The main entry point is LIFTSImputer, which follows the usual pattern:

model = LIFTSImputer(...)
model.fit(X_train, mask_train)
samples = model.impute_samples(X_test, mask_test)

Install

From this folder:

pip install -e .

For a fresh environment, install the listed runtime dependencies first if needed:

pip install -r requirements.txt

Release instructions for PyPI are in RELEASE.md.

Quickstart

from lifts import LIFTSImputer, make_case4_dataset

data = make_case4_dataset(
	num_sample_train=200,
	num_sample_test=50,
	p=20,
	d=5,
	T=48,
	obs_rate=0.5,
	version="v1",
	seed=42,
)

model = LIFTSImputer(
	d=5,
	total_loops=20,
	batch_size=64,
	n_samples=20,
	lambda_selfmask=5.0,
	lambda_enc=1.0,
	enc_warmup=5,
	seed=42,
)

model.fit(data["X_train"], data["mask_train"], verbose=True)
samples = model.impute_samples(data["X_test"], data["mask_test"], n_samples=20)
print(samples.shape)  # (N, n_samples, p, T)

p and T can be passed to the constructor, or inferred from X_train during fit.

Data Format

LIFTS expects numpy arrays with shape (N, p, T):

• X: observed time series values, with missing entries either zero-filled or nan
• mask: binary observation mask with 1 for observed and 0 for missing
• returned imputation samples: (N, n_samples, p, T)

The dataset helpers return both observed and complete arrays:

• X_train, mask_train, X_train_full
• X_test, mask_test, X_test_full

Key Parameters

• lambda_selfmask: weight for the self-mask imputation loss. Default is 5.0.
• lambda_enc: weight for synthetic encoder calibration. Default is 0.0, which preserves the original training behavior.
• enc_warmup: number of training loops before encoder calibration starts. Default is 50.
• enc_syn_batch: number of synthetic trajectories used by encoder calibration. None resolves to min(4000, N) during training.
• enc_syn_mini_batch: chunk size for encoder calibration synthesis. None resolves to min(batch_size, enc_syn_batch).

The training objective is:

L_total = L_forecast + lambda_selfmask * L_selfmask + lambda_enc * L_encoder

L_encoder is only added when lambda_enc > 0 and the warmup has completed.

Included Simulation Cases

• make_case1_dataset(...): linear factor state-space simulation
• make_case3_dataset(...): nonlinear simulation variant
• make_case4_dataset(..., version="v1"): benchmark-compatible Case 4 used by the packaged examples
• make_case4_dataset(..., version="v2"): current local Case 4 generator

Examples

Scripts:

• examples/scripts/case1_minimal.py
• examples/scripts/case3_minimal.py
• examples/scripts/case4_minimal.py

Notebooks:

• examples/notebooks/case1_lifts.ipynb
• examples/notebooks/case4_lifts.ipynb

The notebooks include plot_probabilistic_imputation, exported at the package root:

from lifts import plot_probabilistic_imputation

Backward Compatibility

The lower-level FBIImputer and get_fbi_config APIs remain importable:

from lifts import FBIImputer, get_fbi_config

For new code, prefer LIFTSImputer because it exposes the common training controls directly in the constructor.