avbirt

Importance-Weighted Adversarial Variational Bayes for Item Factor Analysis

avbirt is the reference implementation accompanying:

Luo N, Ji F. Generative Adversarial Networks for High-Dimensional Item Factor Analysis: A Deep Adversarial Learning Algorithm. Psychometrika. 2025 Nov 11; 90(5):1–24. doi: 10.1017/psy.2025.10059. PMID: 41216666; PMCID: PMC12805202.

It fits a confirmatory or exploratory Item Factor Analysis (IFA) model to ordinal response data using the Graded Response Model (GRM), combining:

• Importance-Weighted ELBO (IW-ELBO) — a tighter bound on the marginal log-likelihood than the standard ELBO, approaching MML as the number of importance weights R → ∞.
• Adversarial Variational Bayes (AVB) — the KL divergence between the variational posterior and the prior is estimated by a discriminator network rather than computed in closed form.
• Adaptive Contrast (AC) — normalises posterior samples before passing them to the discriminator, making the adversarial task tractable in high dimensions.
• Doubly Reparameterized Gradients (DReG) — reduces encoder gradient variance.

Installation

pip install avbirt                 # core estimator + presets
pip install avbirt[tune]           # + automatic tuner (optuna)
pip install avbirt[examples]       # + notebook deps (deepirtools, matplotlib, ...)

From source (development):

git clone https://github.com/Feng-Ji-Lab/AVBIRT.git
cd AVBIRT
pip install -e ".[dev,tune,examples]"

Requirements: Python ≥ 3.9, PyTorch ≥ 2.0, Pyro ≥ 1.8. A CUDA-capable GPU is strongly recommended for realistic model sizes.

Quick Start

import torch
from avbirt import IWAVB_estimator

# Response matrix: N respondents × M items, integer responses in {0, ..., C-1}
dataset = torch.randint(0, 5, (1000, 45))

# Confirmatory IFA: define a binary Q-matrix (M items × P factors)
Q = torch.zeros(45, 3)
Q[:15, 0] = 1
Q[15:30, 1] = 1
Q[30:, 2] = 1

estimator = IWAVB_estimator(
    representation_size=3,    # P = 3 latent factors
    num_questions=45,         # M = 45 items
    num_categories=5,         # C = 5 response categories
    num_students=1000,        # N = 1000 respondents
    Q=Q,
)

estimator.fit(dataset, batch_size=32, iw_num=50)

scores   = estimator.predict(dataset)         # (N, P) NumPy array of factor scores
loadings = estimator.get_loadings()           # (M, P) loading matrix
ll       = estimator.log_likelihood(dataset)  # approximate marginal log-likelihood

For exploratory IFA (no Q-matrix), simply omit the Q argument.

Presets — verified configs in one call

Four hyperparameter presets ship with the package, each a 50-rep-verified winner from the tuning study. Pick one from your data shape and fit in one call:

from avbirt import IWAVB_estimator, recommend_preset, list_presets

print(list_presets())                              # catalogue + verdicts
name = recommend_preset(N=5000, M=70, C=5, P=7)    # -> "normal"
est = IWAVB_estimator.from_preset(name, num_students=5000, num_questions=70,
                                  num_categories=5, representation_size=7, Q=Q)
est.fit(data)                                       # preset fit kwargs applied

Ratios are AVBIRT/IWAVE MSE (50-rep): < 1 means AVBIRT wins; a PASS verdict means every metric is within 2× of IWAVE (not necessarily better).

Automatic tuner

For data that doesn't match a preset, tune on your own data (pip install avbirt[tune]):

from avbirt import tune

result = tune(data, representation_size=5, num_categories=5, Q=Q, n_trials=50)
est = result.best_estimator        # refit on all data with the winning config

By default the tuner maximizes the held-out IW-ELBO (no ground truth needed). In a simulation study, pass true_params={"disc": ..., "intercept": ...} to tune for parameter recovery instead. See docs/tuner_and_presets.md.

Examples

The examples/ directory contains executed simulation notebooks that reproduce the paper's studies, plus an S1–S4 comparison notebook (shipped as an unexecuted template):

• simulation_normal.ipynb — multivariate normal latent ability distribution (N=5000, M=70, C=5, P=7).
• simulation_multimodal.ipynb — 3-mode multimodal latent ability distribution (N=5000, M=45, C=3, P=3).
• comparison_S1_S4.ipynb — IWAVE-vs-IWAVB across all four scenarios (S1–S4): a live single-rep fit plus the cached 50-rep results.

Both simulation notebooks compare avbirt against the IWAVE baseline from deepirtools over 3 independent repetitions and report MSE for discrimination, intercept, ability, and factor covariance estimates. Per-rep training logs are saved under examples/logs/.

To regenerate the notebooks from the source script:

python examples/make_notebooks.py

Package Structure

avbirt/
├── __init__.py            ← IWAVB_estimator, tune, presets helpers
├── presets.py             ← canonical S1-S4 presets + recommend_preset/list_presets
├── scenarios.py           ← S1-S4 simulation generators
├── metrics.py             ← parameter-recovery MSE + alignment
├── defaults.py            ← profile registry
├── models/                ← constraints, distributions, grm, networks
├── core/                  ← iwavb (forward), estimator (+ from_preset), init
└── tuner/                 ← tune(), search space, objectives, Optuna driver

Running Tests

pytest                              # full suite
pytest -m "not slow"                # skip end-to-end integration tests
pytest tests/test_constraints.py    # single module

Implementation Status

Beta (0.2.0). The public API is IWAVB_estimator (with .from_preset), tune, and the preset helpers (PRESETS, recommend_preset, list_presets). Internal modules may still change. See CHANGELOG.md.

Citation

If you use this package in your research, please cite:

@article{luo2025avbirt,
  author  = {Luo, Nanyu and Ji, Feng},
  title   = {Generative Adversarial Networks for High-Dimensional Item Factor
             Analysis: A Deep Adversarial Learning Algorithm},
  journal = {Psychometrika},
  year    = {2025},
  volume  = {90},
  number  = {5},
  pages   = {1--24},
  doi     = {10.1017/psy.2025.10059},
  pmid    = {41216666},
  pmcid   = {PMC12805202},
}

License

Released under the MIT License.