IRT（Item Response Theory）のサンプル

py_irt_sample.py

import pandas as pd
import torch
from py_irt.models.two_param_logistic import TwoParamLog

class SafeTwoParamLog(TwoParamLog):
    def fit_MCMC(self, models, items, responses, num_epochs=1000):
        from pyro.infer import MCMC, NUTS

        nuts_kernel = NUTS(self.model_vague, adapt_step_size=True)
        hmc_posterior = MCMC(
            nuts_kernel, num_samples=1000, warmup_steps=100
        )
        hmc_posterior.run(models, items, responses)  # ← run() は None を返すので、分ける！
        self.trace = hmc_posterior
        return hmc_posterior  # ✅ これで正しい TracePosterior を返す

# データ準備
data = pd.DataFrame({
    'user_id': ['A', 'A', 'A', 'B', 'B', 'B', 'C', 'C', 'C'],
    'item_id': ['Q1', 'Q2', 'Q3', 'Q1', 'Q2', 'Q3', 'Q1', 'Q2', 'Q3'],
    'response': [1, 1, 0, 1, 0, 0, 0, 1, 1]
})
user2idx = {u:i for i,u in enumerate(data['user_id'].unique())}
item2idx = {q:i for i,q in enumerate(data['item_id'].unique())}

# Tensor に変換
subjects = torch.tensor(data['user_id'].map(user2idx).to_numpy(), dtype=torch.long)
items = torch.tensor(data['item_id'].map(item2idx).to_numpy(), dtype=torch.long)
responses = torch.tensor(data['response'].to_numpy(), dtype=torch.float)

# モデル初期化
model = SafeTwoParamLog(
    priors='vague',
    num_items=len(item2idx),
    num_subjects=len(user2idx),
    verbose=True,
    device='cpu'
)

# # 🔄 MCMC推定
mcmc = model.fit_MCMC(subjects, items, responses, num_epochs=1000)
# 各パラメータのサンプル取得
samples = mcmc.get_samples()

import pandas as pd

# 各パラメータの統計量を DataFrame に
summary_df = pd.DataFrame({
    "theta_mean": samples["theta"].mean(0).numpy(),
    "theta_std": samples["theta"].std(0).numpy(),
    "b_mean": samples["b"].mean(0).numpy(),
    "b_std": samples["b"].std(0).numpy(),
    "a_mean": samples["a"].mean(0).numpy(),
    "a_std": samples["a"].std(0).numpy(),
})

print(summary_df)