class SVMTextEncoder(BaseEstimator, TransformerMixin):
# number of jobs to execute in parallel
NUM_JOBS = 3
# number of folds to apply to svm fit
NUM_FOLDS = 3
# !! add tuning
def __init__(self, metric, random_seed):
super().__init__()
self._vect = TfidfVectorizer(ngram_range=[1, 2], max_features=30000)
self._random_seed = random_seed
if metric in classification_metrics:
self._model = LinearSVC(class_weight="balanced",
random_state=random_seed)
self.mode = "classification"
elif metric in regression_metrics:
self._model = LinearSVR(random_state=random_seed)
self.mode = "regression"
else:
raise AttributeError(
"metric not in classification or regression metrics")
def fit(self, X, y):
raise NotImplemented
def transform(self, X):
X = pd.Series(X.squeeze()).fillna(MISSING_VALUE_INDICATOR).values
Xv = self._vect.transform(X)
if self.mode == "classification":
out = self._model.decision_function(Xv)
else:
out = self._model.predict(Xv)
if len(out.shape) == 1:
out = out.reshape(-1, 1)
return out
def fit_transform(self, X, y=None, **kwargs):
assert y is not None, "SVMTextEncoder.fit_transform requires y"
X = pd.Series(X.squeeze()).fillna(MISSING_VALUE_INDICATOR).values
Xv = self._vect.fit_transform(X)
self._model = self._model.fit(Xv, y)
if self.mode == "classification":
# Aim for NUM_FOLDS and stratified k-fold. If that doesn't work, fallback to uniform sampling.
num_folds = min(self.NUM_FOLDS, y.value_counts().min())
if num_folds < 2:
cv = KFold(n_splits=self.NUM_FOLDS,
random_state=self._random_seed)
out = cross_val_predict(
self._model,
Xv,
y,
method="decision_function",
n_jobs=self.NUM_JOBS,
cv=cv,
)
else:
out = cross_val_predict(
self._model,
Xv,
y,
method="decision_function",
n_jobs=self.NUM_JOBS,
cv=num_folds,
)
else:
out = cross_val_predict(self._model,
Xv,
y,
n_jobs=self.NUM_JOBS,
cv=self.NUM_FOLDS)
if len(out.shape) == 1:
out = out.reshape(-1, 1)
return out
