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)
df_train = pd.read_csv(os.path.join(DATA_PATH, 'q_train.csv'))
df_test = pd.read_csv(os.path.join(DATA_PATH, 'q_test.csv'))
df_test = df_test.drop(
df_test.head(100).index
) # Not to use the validation data used in 5.1 for model selection
dict_latent_traits = pickle.load(
open(os.path.join(DATA_PATH, 'known_latent_traits.p'), "rb"))
# define latent traits calibrator (known latent traits)
latent_traits_calibrator = KnownParametersCalibrator(dict_latent_traits)
file = open("outputs/5_3_read.txt", 'w')
# pipeline difficulty
pipe_b = FeatureEngAndRegressionPipeline(
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,
df_train = pd.read_csv(os.path.join(DATA_PATH, 'q_train.csv'))
df_test = pd.read_csv(os.path.join(DATA_PATH, 'q_test.csv'))
df_test = df_test.drop(
df_test.head(100).index
) # Not to use the validation data used in 5.1 for model selection
dict_latent_traits = pickle.load(
open(os.path.join(DATA_PATH, 'known_latent_traits.p'), "rb"))
# define latent traits calibrator (known latent traits)
latent_traits_calibrator = KnownParametersCalibrator(dict_latent_traits)
file = open("outputs/5_3_ling.txt", 'w')
# pipeline difficulty
pipe_b = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([LinguisticFeaturesComponent()]),
RegressionModule([
SklearnRegressionComponent(RandomForestRegressor(random_state=SEED),
latent_trait_range=B_RANGE)
]))
# pipeline discrimination
pipe_a = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([LinguisticFeaturesComponent()]),
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,
for min_df in np.arange(0.00, 0.11, 0.02):
for max_df in np.arange(0.90, 1.01, 0.02):
file = open(
"outputs/5_3_ir_mindf_%.2f_maxdf_%.2f.txt" % (min_df, max_df), 'w')
file.write("MIN_DF = %.2f - MAX DF = %.2f" % (min_df, max_df))
# pipeline difficulty
vec_b = TfidfVectorizer(stop_words='english',
preprocessor=preproc,
min_df=min_df,
max_df=max_df)
pipe_b = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([
IRFeaturesComponent(vec_b,
concatenate_correct=True,
concatenate_wrong=True)
]),
RegressionModule([
SklearnRegressionComponent(
RandomForestRegressor(random_state=SEED),
latent_trait_range=B_RANGE)
]))
# pipeline discrimination
vec_a = TfidfVectorizer(stop_words='english',
preprocessor=preproc,
min_df=min_df,
max_df=max_df)
pipe_a = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([
IRFeaturesComponent(vec_a,
# define latent traits calibrator (known latent traits)
latent_traits_calibrator = KnownParametersCalibrator(dict_latent_traits)
for min_df in np.arange(0.00, 0.11, 0.02):
for max_df in np.arange(0.90, 1.01, 0.02):
file = open("outputs/5_1_model_selection_RF_mindf_%.2f_maxdf_%.2f.txt" % (min_df, max_df), 'w')
file.write("MIN_DF = %.2f - MAX DF = %.2f" % (min_df, max_df))
# pipeline difficulty
vec_b = TfidfVectorizer(stop_words='english', preprocessor=preproc, min_df=min_df, max_df=max_df)
pipe_b = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([
IRFeaturesComponent(vec_b, concatenate_correct=True, concatenate_wrong=True),
LinguisticFeaturesComponent(),
ReadabilityFeaturesComponent(),
]),
RegressionModule([
SklearnRegressionComponent(RandomForestRegressor(random_state=SEED), latent_trait_range=B_RANGE)
])
)
# pipeline discrimination
vec_a = TfidfVectorizer(stop_words='english', preprocessor=preproc, min_df=min_df, max_df=max_df)
pipe_a = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([
IRFeaturesComponent(vec_a, concatenate_correct=True, concatenate_wrong=True),
LinguisticFeaturesComponent(),
ReadabilityFeaturesComponent(),
]),
RegressionModule([
df_gte = pd.read_csv(os.path.join(DATA_PATH, 'a_gte.csv'))
df_sap = pd.read_csv(os.path.join(DATA_PATH, 'a_sap.csv'))
df_train = pd.read_csv(os.path.join(DATA_PATH, 'q_train.csv'))
df_test = pd.read_csv(os.path.join(DATA_PATH, 'q_test.csv'))
dict_latent_traits = pickle.load(
open(os.path.join(DATA_PATH, 'known_latent_traits.p'), "rb"))
# This first section is to train the model - I assume that the parameters used here are the ones of the best performing
# model, as obtained in the scripts 5_1_*
pipeline_difficulty = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([
IRFeaturesComponent(TfidfVectorizer(
stop_words='english',
preprocessor=vectorizer_text_preprocessor,
min_df=0.02,
max_df=0.92),
concatenate_correct=True,
concatenate_wrong=True),
LinguisticFeaturesComponent(),
ReadabilityFeaturesComponent(),
]),
RegressionModule([
SklearnRegressionComponent(RFRegressor(n_estimators=100,
max_depth=20,
random_state=SEED),
latent_trait_range=DIFFICULTY_RANGE)
]))
pipeline_discrimination = FeatureEngAndRegressionPipeline(
FeatureEngineeringModule([
IRFeaturesComponent(TfidfVectorizer(
stop_words='english',