def _convert_vector(df, dtype):
from pyspark.ml.linalg import VectorUDT
from pyspark.mllib.linalg import VectorUDT as OldVectorUDT
for field in df.schema:
col_name = field.name
if isinstance(field.dataType, VectorUDT) or \
isinstance(field.dataType, OldVectorUDT):
df = df.withColumn(col_name, vector_to_array(df[col_name], dtype))
return df
def cast_spark_df_with_vector_to_array(input_spark_df):
"""
Finds columns of vector type in a spark dataframe and
casts them to array<double> type.
:param input_spark_df:
:return: a spark dataframe with vector columns transformed to array<double> type
"""
vector_type_columns = [
_field.name for _field in input_spark_df.schema
if isinstance(_field.dataType, VectorUDT)
]
return reduce(
lambda df, vector_col: df.withColumn(vector_col,
vector_to_array(vector_col)),
vector_type_columns,
input_spark_df,
)
def _convert_vector(df, precision):
from pyspark.ml.linalg import Vector
from pyspark.mllib.linalg import Vector as OldVector
types_set = {struct_field.dataType for struct_field in df.schema}
found_vectors = not types_set.isdisjoint({Vector, OldVector})
if not found_vectors:
return df
import pyspark
if LooseVersion(pyspark.__version__) >= LooseVersion('3.0'):
raise ValueError("Vector columns are not supported for pyspark<3.0.0.")
# pylint: disable=import-error
from pyspark.ml.functions import vector_to_array
# pylint: enable=import-error
for struct_field in df.schema:
col_name = struct_field.name
if struct_field.dataType in {Vector, OldVector}:
df = df.withColumn(col_name,
vector_to_array(df[col_name], precision))
return df
def _validate_and_convert_feature_col_as_array_col(dataset, features_col_name):
features_col_datatype = dataset.schema[features_col_name].dataType
features_col = col(features_col_name)
if isinstance(features_col_datatype, ArrayType):
if not isinstance(
features_col_datatype.elementType,
(DoubleType, FloatType, LongType, IntegerType, ShortType),
):
raise ValueError(
"If feature column is array type, its elements must be number type."
)
features_array_col = features_col.cast(ArrayType(
FloatType())).alias("values")
elif isinstance(features_col_datatype, VectorUDT):
features_array_col = vector_to_array(features_col,
dtype="float32").alias("values")
else:
raise ValueError(
"feature column must be array type or `pyspark.ml.linalg.Vector` type, "
"if you want to use multiple numetric columns as features, please use "
"`pyspark.ml.transform.VectorAssembler` to assemble them into a vector "
"type column first.")
return features_array_col
def process_bus_data(bus_df):
""" Method to process raw business data from Yelp."""
def select_elibigble_bus(row):
""" Select businesses which fall into selected categores."""
global categories
try:
# Return true if business falls into category list, else false.
row_cats = row.split(',')
for cat in row_cats:
if cat.strip() in categories:
return True
return False
except (TypeError, AttributeError):
# Returns false if business has no defined categories.
return False
def unpack_bus_attributes(row):
""" Unpacks Business attributes and assigns them an index value."""
# List to store business attributes.
unpacked = list()
# Unpack all attributes except PriceRange and Parking
temp = [row[s] for s in bus_attributes]
# Process PriceRange
try:
priceRange = int(row["RestaurantsPriceRange2"])
except (TypeError, ValueError):
# If no price range specified - default=2
priceRange = 2
#Process Parking
try:
parking = 1 if (row["BusinessParking"].find("True")) != -1 else -1
except AttributeError:
parking = 0
# Process WiFi
if row["WiFi"] == 'no' or row["WiFi"] == "u'no'":
wifi = -1
elif row["WiFi"] == None:
wifi = 0
else:
wifi = 1
# Tokenize all Boolean attributes.
for i in temp:
if i == "True":
unpacked.append(1)
elif i == "False":
unpacked.append(-1)
else:
unpacked.append(0)
# Append the Parking and PriceRange attributes
unpacked.append(wifi)
unpacked.append(parking)
unpacked.append(priceRange)
# Print any arrays that are not of desired length (=30).
if len(unpacked) != 30:
print(unpacked)
return _convert_to_vector(
csc_matrix(np.asarray(unpacked).astype(float)).T)
def unpack_bus_categories(row):
"""Unpacks all business cattegories."""
# List to store business categories.
unpacked = list()
# Unpack all attributes except PriceRange and Parking
for cat in row.split(','):
unpacked.append(cat.strip())
return unpacked
def unpack_price_range(row):
""" Returns price range."""
return int(row[-1])
# Package the functions above into Spark SQL user-defined functions
udf_select_eligible_bus = udf(select_elibigble_bus, BooleanType())
udf_unpack_bus_attributes = udf(unpack_bus_attributes, VectorUDT())
udf_unpack_bus_categories = udf(unpack_bus_categories,
ArrayType(StringType()))
udf_unpack_price_range = udf(unpack_price_range, IntegerType())
# Find businesses to include.
eligible_bus = bus_df.withColumn("include", udf_select_eligible_bus(col("categories"))) \
.filter(col("include") == True)
# Process business attributes feature.
all_bus_attributes = set(
bus_df.select("attributes").take(1)[0].attributes.asDict().keys())
bus_attributes_to_exclude = {
'AcceptsInsurance', 'AgesAllowed', 'ByAppointmentOnly', 'Caters',
'Corkage', 'DietaryRestrictions', 'HairSpecializesIn', 'Open24Hours',
'RestaurantsAttire', 'RestaurantsPriceRange2', 'BusinessParking',
'WiFi'
}
bus_attributes = list(all_bus_attributes - bus_attributes_to_exclude)
bus_attributes.sort()
eligible_attr = eligible_bus.withColumn(
"unpackedAttr", udf_unpack_bus_attributes(col("attributes")))
# Process business categories feature.
eligible_cats = eligible_attr.withColumn(
"unpackedCats", udf_unpack_bus_categories(col("categories")))
cv = CountVectorizer(inputCol="unpackedCats", outputCol="vectorizedCats")
vectorized_cats = cv.fit(eligible_cats).transform(eligible_cats)
# Un-bundle price range from all other attributes.
unpacked_pr = vectorized_cats.withColumn(
"priceRange", udf_unpack_price_range(col("unpackedAttr")))
unpacked_pr.take(1)
# Reduce dimensions of attributes and categories features, respectively.
pca_attr = PCA(k=3, inputCol="unpackedAttr",
outputCol="pcaAttr").fit(unpacked_pr)
temp = pca_attr.transform(unpacked_pr)
temp.show()
pca_cats = PCA(k=1, inputCol="vectorizedCats",
outputCol="pcaCats").fit(temp)
temp2 = pca_cats.transform(temp)
temp2.show()
# Assemble into final feature vector.
va = VectorAssembler(
inputCols=["stars", "priceRange", "pcaAttr", "pcaCats"],
outputCol="featureVec")
features = va.transform(temp2).select("business_id", "stars", "categories",
"featureVec")
features.take(1)
# Unpack
n_features = len(features.select("featureVec").take(1)[0].featureVec)
final = features.withColumn("f", vector_to_array(col("featureVec"))) \
.select(["business_id", "stars", "categories"] + [col("f")[i] for i in range(n_features)])
return final, n_features
# url=db_url,
# table="text_data_train_nm_tfidf",
# mode="overwrite",
# properties=db_properties,
# )
# test_finished.write.jdbc(
# url=db_url,
# table="text_data_test_nm_tfidf",
# mode="overwrite",
# properties=db_properties,
# )
# SVM
train_pred = train_pred.withColumn("rawPrediction",
vector_to_array("rawPrediction"))
test_pred = test_pred.withColumn("rawPrediction",
vector_to_array("rawPrediction"))
train_pred.createOrReplaceTempView("train_pred")
test_pred.createOrReplaceTempView("test_pred")
train_finished = spark.sql("""
SELECT review_id,
ROUND(rawPrediction[1], 3) AS svm_pred
FROM train_pred
""")
test_finished = spark.sql("""
SELECT review_id,
ROUND(rawPrediction[1], 3) AS svm_pred
)
if scaler_model is None:
scaler = StandardScaler(
inputCol=temp_vector_col,
outputCol=temp_normalized_vector_col,
withStd=parse_parameter(bool, scale, "scale", True),
withMean=parse_parameter(bool, center, "center", False),
)
scaler_model = fit_and_save_model(trained_parameters, "scaler_model", scaler, assembled_wo_nans)
output_df = transform_using_trained_model(scaler_model, assembled, scaler_model_loaded)
# convert the resulting vector back to numeric
temp_flattened_vector_col = temp_col_name(output_df)
output_df = output_df.withColumn(temp_flattened_vector_col, vector_to_array(temp_normalized_vector_col))
# keep only the final scaled column.
output_column = input_column if output_column is None or not output_column else output_column
output_column_value = sf.col(temp_flattened_vector_col)[0].alias(output_column)
output_df = output_df.withColumn(output_column, output_column_value)
final_columns = list(dict.fromkeys((list(df.columns) + [output_column])))
output_df = output_df.select(final_columns)
return default_spark_with_trained_parameters(output_df, trained_parameters)
def process_numeric_robust_scaler(
df,
input_column=None,
lower_quantile=None,