def compute_category_similarity():
ds = DataSet.open(dataset_dir)
TrainSearchStream = ds.get_table("TrainSearchStream")
AdsInfo = ds.get_table("AdsInfo")
Category = ds.get_table("Category")
SearchInfo = ds.get_table("SearchInfo")
TrainSearchStream.add_join_column("AdCategoryID", [TrainSearchStream.get_column("AdID_ref"),
AdsInfo.get_column("CategoryID_ref")])
TrainSearchStream.add_join_column("AdCategoryLevel", [TrainSearchStream.get_column("AdID_ref"),
AdsInfo.get_column("CategoryID_ref"),
Category.get_column("Level")])
TrainSearchStream.add_join_column("AdCategoryParentID", [TrainSearchStream.get_column("AdID_ref"),
AdsInfo.get_column("CategoryID_ref"),
Category.get_column("ParentCategoryID")])
TrainSearchStream.add_join_column("SearchCategoryID", [TrainSearchStream.get_column("SearchID_ref"),
SearchInfo.get_column("CategoryID_ref")])
TrainSearchStream.add_join_column("SearchCategoryLevel", [TrainSearchStream.get_column("SearchID_ref"),
SearchInfo.get_column("CategoryID_ref"),
Category.get_column("Level")])
TrainSearchStream.add_join_column("SearchCategoryParentID", [TrainSearchStream.get_column("SearchID_ref"),
SearchInfo.get_column("CategoryID_ref"),
Category.get_column("ParentCategoryID")])
print(TrainSearchStream)
def test_sum():
import time
ds = DataSet.open(dataset_dir)
tss = ds.get_table("TrainSearchStream")
p = tss.get_column("Position")
t = time.time()
print(p.sum())
print(time.time() - t)
def check_raw_dataset():
# Category(CategoryID: int32, Level: int8, ParentCategoryID: int8): 68 row(s) - compressed: 0.06 MB - comp. ratio: 0.01
# TrainSearchStream(SearchID: int32, AdID: int32, Position: int8, ObjectType: int8, HistCTR: float32, IsClick: float32): 190,157,735 row(s) - compressed: 1479.89 MB - comp. ratio: 2.21
# AdsInfo(AdID: int32, LocationID: float32, CategoryID: float32, Price: float32, IsContext: float32): 36,893,298 row(s) - compressed: 280.61 MB - comp. ratio: 2.51
# SearchInfo(SearchID: int32, IPID: int32, UserID: float32, IsUserLoggedOn: float32, LocationID: float32, CategoryID: float32): 91,019,228 row(s) - compressed: 1043.73 MB - comp. ratio: 2.00
# userInfo(UserID: int32, UserAgentID: int32, UserAgentOSID: int32, UserDeviceID: int32, UserAgentFamilyID: int32): 4,284,823 row(s) - compressed: 20.32 MB - comp. ratio: 4.02
# Location(LocationID: int32, RegionID: float32, CityID: float32): 4,080 row(s) - compressed: 0.38 MB - comp. ratio: 0.12
# PhoneRequestsStream(UserID: int32, IPID: int32, AdID: int32, PhoneRequestDate: bytes168): 13,717,580 row(s) - compressed: 139.27 MB - comp. ratio: 3.10
# VisitsStream(UserID: int32, IPID: int32, AdID: int32, ViewDate: bytes168): 286,821,375 row(s) - compressed: 2548.20 MB - comp. ratio: 3.54
ds = DataSet.open(raw_dir)
for table in ds.tables:
print(table.short_descr())
def load_dataset():
"""'Raw'-dataset is the result of loading the CSV sources data into dazzle tables, only filtering out data
that we don't want to further process.
The method is programmed in a non-destructive way so as to be able to launch it several times
before getting the job done.
"""
import os
from dazzle.core.dataset import DataSet
if DataSet.exists(raw_dir):
ds = DataSet.open(raw_dir)
else:
ds = DataSet(raw_dir, force_create=True)
# Notes:
# - many of the following attributes should be unsigned int instead of signed int, but numexpr works only on
# signed data.
# - Simlarly to pandas, we use the types required to contain the existing data, not the types we desire to use
if ds.get_table("Category") is None:
t = Table.from_csv("Category", ds, os.path.join(csv_dir, "Category.tsv"), delimiter='\t', chunksize=10**7,
usecols=['CategoryID', 'ParentCategoryID', 'Level'],
dtype={'CategoryID': 'i4', 'ParentCategoryID': 'i1', 'Level': 'i1'})
t = None
# Notice the filter attribute that does not exist in pandas.read_csv(). It makes it possible to skip some rows
# based on a numexpr expression. IsClick == IsClick is true iff IsClick is not na
if ds.get_table("TrainSearchStream") is None:
t = Table.from_csv("TrainSearchStream", ds, os.path.join(csv_dir, "trainSearchStream.tsv"), delimiter='\t', chunksize=10**7,
usecols=['SearchID', 'AdID', 'Position', 'ObjectType', 'HistCTR', 'IsClick'],
dtype={'SearchID':'i4', 'AdID':'i4', 'Position':'i1', 'ObjectType':'i1', 'HistCTR':'f4', 'IsClick':'f1'},
filter='(ObjectType == 3) & (IsClick == IsClick)')
t = None
# We avoid to load the string fields. We will see this problem later with Don
if ds.get_table("AdsInfo") is None:
t = Table.from_csv("AdsInfo", ds, os.path.join(csv_dir, "AdsInfo.tsv"), delimiter='\t', chunksize=10**7,
usecols=['AdID', 'LocationID', 'CategoryID', 'Price', 'IsContext'],
dtype={'AdID':'i4', 'LocationID':'f4', 'CategoryID':'f4', 'Price': 'f4', 'IsContext': 'f1'})
t = None
# We avoid to load the string fields. We will see this problem later with Don
if ds.get_table("SearchInfo") is None:
t = Table.from_csv("SearchInfo", ds, os.path.join(csv_dir, "SearchInfo.tsv"), delimiter='\t', chunksize=10**7,
usecols=['SearchID', 'IPID', 'UserID', 'IsUserLoggedOn', 'LocationID', 'CategoryID'],
dtype={'SearchID':'i4', 'IPID':'i4', 'UserID':'f4', 'IsUserLoggedOn':'f1',
'LocationID':'f4', 'CategoryID':'f4'})
t = None
if ds.get_table("userInfo") is None:
t = Table.from_csv("userInfo", ds, os.path.join(csv_dir, "userInfo.tsv"), delimiter='\t', chunksize=10**7,
usecols=['UserID', 'UserAgentID', 'UserAgentOSID','UserDeviceID', 'UserAgentFamilyID'],
dtype={'UserID':'i4', 'UserAgentID':'i4', 'UserAgentOSID':'i4',
'UserDeviceID':'i4', 'UserAgentFamilyID':'i4'})
t = None
if ds.get_table("Location") is None:
t = Table.from_csv("Location", ds, os.path.join(csv_dir, "Location.tsv"), delimiter='\t', chunksize=10**7,
usecols=['LocationID', 'CityID', 'RegionID'],
dtype={'LocationID': 'i4', 'CityID':'f4', 'RegionID': 'f4'})
t = None
if ds.get_table("PhoneRequestsStream") is None:
t = Table.from_csv("PhoneRequestsStream", ds, os.path.join(csv_dir, "PhoneRequestsStream.tsv"), delimiter='\t', chunksize=10**7,
usecols=['UserID', 'IPID', 'AdID', 'PhoneRequestDate'],
dtype={'UserID':'i4', 'IPID':'i4', 'AdID':'i4', 'PhoneRequestDate': 'object'})
t = None
if ds.get_table("VisitsStream") is None:
t = Table.from_csv("VisitsStream", ds, os.path.join(csv_dir, "VisitsStream.tsv"), delimiter='\t', chunksize=10**7,
usecols=['UserID', 'IPID', 'AdID', 'ViewDate'],
dtype={'UserID':'i4', 'IPID':'i4', 'AdID':'i4', 'ViewDate': 'object'})
t = None
return ds
def preprocess_dataset():
# This step takes around 3 mins (6 mins if dataset must be copied)
#
# Table.add_reference_column(), which uses pandas is partly responsible for this. In addition,
# it consumes a lot of RAM.
# 1. Make a copy of the raw dataset, if this has not already be done: we don't want to reload the whole CSV stuff
# if something wrong happens
if not DataSet.exists(dataset_dir):
print("Copying dataset ...")
raw_ds = DataSet.open(raw_dir)
ds = raw_ds.copy(dataset_dir) # almost 3 mins !
else:
ds = DataSet.open(dataset_dir)
# 2. Rebuild each table. This means:
#
# 2.1 inserting a nan row at the head of each table. This is necessary because we use index=0 in each RefColumn
# for indicating a null reference
#
# 2.2 assigning the desired dtype of each column
#
# 2.3 Setting data in each column using the setting dtype
#
# 2.4 Replace Numpy NA values by those of the corresponding column class (Ref/Literal) and dtype
#
Category = ds.get_table("Category")
Location = ds.get_table("Location")
userInfo = ds.get_table("userInfo")
AdsInfo = ds.get_table("AdsInfo")
SearchInfo = ds.get_table("SearchInfo")
TrainSearchStream = ds.get_table("TrainSearchStream")
print("Re-building tables with given dtypes ...")
Category.rebuild({"CategoryID": np.int32, "Level": np.int8, "ParentCategoryID": np.int32})
Location.rebuild({"LocationID": np.int16, "RegionID": np.int8, "CityID": np.int16})
userInfo.rebuild({"UserID": np.int32, "UserAgentID": np.int32, "UserAgentOSID": np.int8,
"UserDeviceID": np.int16, "UserAgentFamilyID": np.int8})
AdsInfo.rebuild({"AdID": np.int32, "LocationID": np.int16, "CategoryID": np.int32, "Price": np.float32,
"IsContext": np.int8})
SearchInfo.rebuild({"SearchID": np.int32, "IPID": np.int32, "UserID": np.int32, "IsUserLoggedOn": np.int8,
"LocationID": np.int16, "CategoryID": np.int32})
TrainSearchStream.rebuild({"SearchID": np.int32, "AdID": np.int32, "Position": np.int8,
"ObjectType": np.int8, "HistCTR": np.float32, "IsClick": np.int8})
# 3. Add references between columns: foreign keys (like LocationID in AdsInfo) are kept
# but an additional column (xxx_ref) is added with the index of the row containing the referenced value
#
print("Building references from AdsInfo ...")
AdsInfo.add_reference_column(AdsInfo.get_column("LocationID"), Location.get_column("LocationID"))
AdsInfo.add_reference_column(AdsInfo.get_column("CategoryID"), Category.get_column("CategoryID"))
print(AdsInfo)
print("Building references from SearchInfo ...")
SearchInfo.add_reference_column(SearchInfo.get_column("UserID"), userInfo.get_column("UserID"))
SearchInfo.add_reference_column(SearchInfo.get_column("LocationID"), Location.get_column("LocationID"))
SearchInfo.add_reference_column(SearchInfo.get_column("CategoryID"), Category.get_column("CategoryID"))
print("Building references from TrainSearchStream ...")
TrainSearchStream.add_reference_column(TrainSearchStream.get_column("SearchID"), SearchInfo.get_column("SearchID"))
TrainSearchStream.add_reference_column(TrainSearchStream.get_column("AdID"), AdsInfo.get_column("AdID"))
print(TrainSearchStream)
print("Done")
