def __init__(self, random_state: int = None, known_latent_traits: Dict[str, Dict[str, float]] = None):
if known_latent_traits is not None:
latent_traits_calibrator = KnownParametersCalibrator(latent_traits=known_latent_traits)
if set(known_latent_traits.keys()) != {DIFFICULTY, DISCRIMINATION}:
raise ValueError("wrong keys in known_latent_traits dictionary")
else:
latent_traits_calibrator = IRTCalibrator(DIFFICULTY_RANGE, DISCRIMINATION_RANGE)
vec_diff = TfidfVectorizer(stop_words='english', preprocessor=vectorizer_text_preprocessor, max_features=1000)
feat_eng_regression_pipeline_difficulty = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([IRFeaturesComponent(vec_diff, concatenate_correct=True, concatenate_wrong=True)]),
RegressionModule([
SklearnRegressionComponent(
RandomForestRegressor(n_estimators=250, max_depth=50, random_state=random_state),
latent_trait_range=DIFFICULTY_RANGE
)
])
)
vec_disc = TfidfVectorizer(stop_words='english', preprocessor=vectorizer_text_preprocessor, max_features=800)
feat_eng_regression_pipeline_discrimination = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([IRFeaturesComponent(vec_disc, concatenate_correct=True, concatenate_wrong=True)]),
RegressionModule([
SklearnRegressionComponent(
RandomForestRegressor(n_estimators=100, max_depth=25, random_state=random_state),
latent_trait_range=DISCRIMINATION_RANGE
)
])
)
estimator_from_text = FeatureEngAndRegressionEstimatorFromText(
{
DIFFICULTY: feat_eng_regression_pipeline_difficulty,
DISCRIMINATION: feat_eng_regression_pipeline_discrimination
}
)
super().__init__(latent_traits_calibrator, estimator_from_text)
FeatureEngineeringModule([ReadabilityFeaturesComponent()]),
RegressionModule([
SklearnRegressionComponent(RandomForestRegressor(random_state=SEED),
latent_trait_range=B_RANGE)
]))
# pipeline discrimination
pipe_a = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([ReadabilityFeaturesComponent()]),
RegressionModule([
SklearnRegressionComponent(RandomForestRegressor(random_state=SEED),
latent_trait_range=A_RANGE)
]))
# create estimator from text form the previous pipelines
estimator_from_text = FeatureEngAndRegressionEstimatorFromText({
DIFFICULTY:
pipe_b,
DISCRIMINATION:
pipe_a
})
model = Text2PropsModel(latent_traits_calibrator, estimator_from_text)
model.calibrate_latent_traits(None)
# define parameters for randomized CV
dict_params = {
DIFFICULTY: [{
'regressor__n_estimators': randint(20, 200),
'regressor__max_depth': randint(2, 50)
}],
DISCRIMINATION: [{
'regressor__n_estimators': randint(20, 200),
'regressor__max_depth': randint(2, 50)
}],
concatenate_correct=True,
concatenate_wrong=True),
LinguisticFeaturesComponent(),
ReadabilityFeaturesComponent(),
]),
RegressionModule([
SklearnRegressionComponent(RFRegressor(n_estimators=100,
max_depth=20,
random_state=SEED),
latent_trait_range=DISCRIMINATION_RANGE)
]))
model = Text2PropsModel(
KnownParametersCalibrator(dict_latent_traits),
FeatureEngAndRegressionEstimatorFromText({
DIFFICULTY:
pipeline_difficulty,
DISCRIMINATION:
pipeline_discrimination
}))
model.train(df_train)
print('[INFO] model trained')
# Here I estimate the latent traits for the test set
dict_predictions_test_set = model.predict(df_test)
# I have to convert the dictionary of the prediction in the right format as model.predict returns a dict of lists
# (one list for each latent trait)
dict_predicted_latent_traits = dict()
dict_predicted_latent_traits[DIFFICULTY], dict_predicted_latent_traits[
DISCRIMINATION] = dict(), dict()
for idx, q_id in enumerate(df_test[Q_ID].values):
dict_predicted_latent_traits[DIFFICULTY][q_id] = dict_predictions_test_set[
DIFFICULTY][idx]