def cluster(self, df, session, repartition_num=8):
n = df.count()
# index rows
df_index = df.select((row_number().over(
Window.partitionBy(lit(0)).orderBy(self.featureCol)) -
1).alias('id'), "*")
df_features = df_index.select('id', self.featureCol)
# prep for joining
df_features = df_features.repartitionByRange(repartition_num, 'id')
left_df = df_features.select(
df_features['id'].alias('left_id'),
df_features[self.featureCol].alias('left_features'))
right_df = df_features.select(
df_features['id'].alias('right_id'),
df_features[self.featureCol].alias('right_features'))
# join on self where left_id does not equal right_id
joined_df = left_df.join(right_df,
left_df['left_id'] != right_df['right_id'])
# comupte cosine similarity between vectors
joined_df = joined_df.select(
'left_id', 'right_id',
cosine_similarity_udf(
array(joined_df['left_features'],
joined_df['right_features'])).alias('norm'))
ranked = joined_df.select(
'left_id', 'right_id',
rank().over(
Window.partitionBy('left_id').orderBy('norm')).alias('rank'))
knn = ranked.where(ranked['rank'] <= 5)
knn_grouped = knn.groupBy('left_id').agg(
f.collect_list('right_id').alias('nn'))
# generate laplacian
laplacian = knn_grouped.select(
knn_grouped['left_id'].alias('id'),
toVector_udf(
laplacian_vector_udf(knn_grouped['left_id'], knn_grouped['nn'],
lit(n),
lit(self.k_nearest))).alias('lap_vector'))
pca = PCA(k=self.num_eigenvectors,
inputCol='lap_vector',
outputCol='features').fit(laplacian)
eigenvectors = pca.transform(laplacian).select('id', 'features')
model = KMeans(featuresCol='features',
predictionCol=self.predictionCol,
k=self.k).fit(eigenvectors)
predictions = model.transform(eigenvectors).join(df_index, on='id')
return predictions
def pca(dataset, inputCol, k=3):
from pyspark.ml.feature import PCA
model = PCA(k=3, inputCol=inputCol, outputCol=inputCol+'_pca').fit(dataset)
return model.transform(dataset), model
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
def apply_pca(training, testing):
pca = PCA(k=3, inputCol="features", outputCol="pca_features").fit(training)
training = pca.transform(training).cache()
testing = pca.transform(testing).cache()
return training, testing
def preprocessing(df, num_pca=10):
argo_df_og = df
# Cast temp as DoubleType()
argo_df_og = argo_df_og.withColumn("tempTmp", argo_df_og['temp'].cast(DoubleType()))\
.drop("temp")\
.withColumnRenamed("tempTmp", "temp")\
.select("profile_id", "pres", "temp", "lat", "lon", "psal", "date")\
.persist()
argo_filterby = argo_df_og.groupBy("profile_id") \
.agg(min("pres").alias("min_pres"),
max("pres").alias("max_pres"),
count("profile_id").alias("count_profile_id"))
# Now, here are the profile_ids we want to keep, to be inner joined with original argo_df_og
argo_keep_ids = argo_filterby.filter("count_profile_id >= 50 and min_pres <= 25 and max_pres >= 999") \
.select("profile_id")
# Inner join the profile_ids to keep with original argo_df_og to filter and keep only desired IDs
argo_df_keep = argo_keep_ids.join(argo_df_og, "profile_id",
"inner").persist()
#Final filtered df after pressure cleaning
argo_df = argo_df_keep.select("profile_id", "pres", "temp", "lat", "lon", "psal", "date",
month("date").alias("month"), year("date").alias("year")) \
.persist()
#INTERPOLATION
#Create vector mapping correspoding temperatures with pressures
argo_df_listed = argo_df.select('profile_id', 'lat', 'lon', array(argo_df['temp'], argo_df['pres']).alias('temp_pres'))\
.groupBy('profile_id').agg(collect_list('temp_pres').alias('temp_pres_list'),
fn.min(argo_df['lat']).alias('lat'),
fn.min(argo_df['lon']).alias('lon'))
# Ordering by pressure
argo_df_listed = argo_df_listed.select(
'profile_id', 'lat', 'lon',
insane_sort(argo_df_listed['temp_pres_list']).alias('temp_pres_list'))
# Interpolating missing temps at specified grid points
pres = argo_df_listed.select(
'profile_id', 'lat', 'lon',
interp_udf('temp_pres_list').alias('temp_interp'))
# Finding profiles with temps as nans
check_pres = pres.select(
"profile_id", "temp_interp", 'lat', 'lon',
null_udf("temp_interp").alias("temp_interp_hasNA"),
lenarray_udf("temp_interp").alias("temp_interp_len199"))
# Filtering profiles with temps as nans
filtered_pres = check_pres.filter("temp_interp_hasNA == False").select(
"profile_id", "temp_interp", 'lat', 'lon')
# Finding profiles with temps < -5
check_pres = filtered_pres.select(
"profile_id", "temp_interp", 'lat', 'lon',
neg_udf("temp_interp").alias("temp_interp_hasNeg5s"))
# Filtering profiles with temps < -5
argo_df_clean = check_pres.filter("temp_interp_hasNeg5s == False").select(
"profile_id", "temp_interp", 'lat', 'lon')
argo_df_clean = argo_df_clean.select(
'profile_id',
toVector_udf(argo_df_clean['temp_interp']).alias('features'), 'lat',
'lon')
pca = PCA(k=num_pca, inputCol='features',
outputCol='features_pca').fit(argo_df_clean)
argo_df_clean = pca.transform(argo_df_clean)
argo_df_clean = argo_df_clean.select(
'profile_id', argo_df_clean['features_pca'].alias('features'), 'lat',
'lon')
return argo_df_clean
label_mapping = dict(enumerate(labelIndexer.labels()))
reverse_mapping = {}
for key in label_mapping:
reverse_mapping[label_mapping[key]] = key
# ## Dimensionality reduction
#
# Feature selection is not really supported yet in mllib, therefore, we just applied dim reduction using PCA
# In[509]:
pca = PCA(inputCol="features", outputCol="pca", k=15).fit(train_df)
train_df = pca.transform(train_df)
test_df = pca.transform(test_df)
# ## Classification algorithms
# In[ ]:
rf = RandomForestClassifier(labelCol="indexedResult", featuresCol="pca", numTrees=5000)
#rf = RandomForestClassifier(labelCol="indexedResult", featuresCol="features", numTrees=5000)
model = rf.fit(train_df)
# ## Evaluation & results
# In[ ]:
test_df = labelIndexer.transform(test_df)
label_mapping = dict(enumerate(labelIndexer.labels()))
reverse_mapping = {}
for key in label_mapping:
reverse_mapping[label_mapping[key]] = key
# ## Dimensionality reduction
#
# Feature selection is not really supported yet in mllib, therefore, we just applied dim reduction using PCA
# In[509]:
pca = PCA(inputCol="features", outputCol="pca", k=15).fit(train_df)
train_df = pca.transform(train_df)
test_df = pca.transform(test_df)
# ## Classification algorithms
# In[ ]:
rf = RandomForestClassifier(labelCol="indexedResult",
featuresCol="pca",
numTrees=5000)
#rf = RandomForestClassifier(labelCol="indexedResult", featuresCol="features", numTrees=5000)
model = rf.fit(train_df)
# ## Evaluation & results
# In[ ]:
except:
rel['features'] = Vectors.dense(0, 0, 0, 0, 0, 0)
rel['label'] = str(x[14].strip('.'))
return rel
df = spark.sparkContext.textFile('./adult/adult.data').map(lambda line: line.split(',')).\
map(lambda x: Row(**f(x))).toDF()
test = spark.sparkContext.textFile('./adult/adult.test').map(lambda line: line.split(',')).\
map(lambda x: Row(**f(x))).toDF()
# 构建模型
pca = PCA(k=3, inputCol='features', outputCol='pcaFeatures').fit(df)
result = pca.transform(df)
test_data = pca.transform(test)
result.show(truncate=False)
test_data.show(truncate=False)
# 在主成分分析的基础上做逻辑回归
labelIndexer = StringIndexer(inputCol='label', outputCol='indexedLabel').fit(result)
for label in labelIndexer.labels:
print(label)
featureIndexer = VectorIndexer(inputCol='pcaFeatures', outputCol='indexedFeatures').fit(result)
print(featureIndexer.numFeatures)
labelConverter = IndexToString(inputCol='prediction', outputCol='predictedLabel', labels=labelIndexer.labels)
hashingTF = HashingTF(inputCol="filtered", outputCol="rawFeatures", numFeatures=20)
featurizedData = hashingTF.transform(filteredData)
idf = IDF(inputCol="rawFeatures", outputCol="tfidffeatures")
idfModel = idf.fit(featurizedData)
tfidfData = idfModel.transform(featurizedData)
# COMMAND ----------
from pyspark.ml.feature import PCA
pca = PCA(k = 5, inputCol = 'tfidffeatures', outputCol = 'features').fit(tfidfData)
data_pca = pca.transform(tfidfData)
data_pca.select("book_id", "features").show(truncate=False)
# COMMAND ----------
# DBTITLE 1,K-means
from pyspark.ml.clustering import KMeans
from pyspark.ml.feature import VectorAssembler, StringIndexer
from pyspark.ml.evaluation import ClusteringEvaluator
kmeans = KMeans().setK(20).setSeed(1)
model = kmeans.fit(data_pca)
predictions = model.transform(data_pca)
