dna-casestudy / utils.py
davidna22's picture
Upload folder using huggingface_hub
dad00c5 verified
raw
history blame
40.3 kB
import pandas as pd
import polars as pl
import numpy as np
import json
import gc
import folium
import html
from matplotlib import pyplot as plt
import seaborn as sns
import xgboost as xgb
from xgboost import plot_importance
from bs4 import BeautifulSoup
import plotly.express as px
import plotly.graph_objects as go
import plotly.figure_factory as ff
from plotly.subplots import make_subplots
import plotly.io as pio
from statsmodels.graphics.tsaplots import plot_pacf, plot_acf
from statsmodels.tsa.stattools import kpss, adfuller
from bertopic import BERTopic
from collections import defaultdict
color_pal = sns.color_palette("tab10")
impute_cols = [
'MeanTemp', 'MinTemp', 'MaxTemp', 'DewPoint',
'Percipitation', 'WindSpeed', 'MaxSustainedWind',
'Gust', 'Rain', 'SnowDepth', 'SnowIce',
]
def convert_schema_to_polars(schema):
pl_schema = {}
for k, v in schema.items():
if v == "String":
pl_schema[k] = pl.String
elif v == "Float64":
pl_schema[k] = pl.Float64
elif v == "Int64":
pl_schema[k] = pl.Int64
return pl_schema
def create_datetime(data, dt_col, format="%m/%d/%Y %I:%M:%S %p"):
# df type is either pandas or polars
df_type = "pandas" if isinstance(data, pd.DataFrame) else "polars"
if "datetime" in str(data[dt_col].dtype).lower():
return data
if df_type == "pandas":
data[dt_col] = pd.to_datetime(data[dt_col], format=format)
elif df_type == "polars":
data = data.with_columns(
pl.col(dt_col).str.strptime(pl.Date, format=format).cast(pl.Datetime)
)
return data
def create_seasons(data, dt_col="Datetime", out_col="Season", prefix=""):
df_type = "pandas" if isinstance(data, pd.DataFrame) else "polars"
out_col = prefix + out_col
spring_start = pd.to_datetime("2018-3-20", format = "%Y-%m-%d").dayofyear
summer_start = pd.to_datetime("2018-6-21", format = "%Y-%m-%d").dayofyear
autumn_start = pd.to_datetime("2018-9-22", format = "%Y-%m-%d").dayofyear
winter_start = pd.to_datetime("2018-12-21", format = "%Y-%m-%d").dayofyear
if df_type == "pandas":
def map_season(date):
if date.dayofyear < spring_start or date.dayofyear >= winter_start:
return "Winter"
elif date.dayofyear >= spring_start and date.dayofyear < summer_start:
return "Spring"
elif date.dayofyear >= summer_start and date.dayofyear < autumn_start:
return "Summer"
elif date.dayofyear >= autumn_start and date.dayofyear < winter_start:
return "Autumn"
data[out_col] = data[dt_col].apply(map_season)
return data
elif df_type == "polars":
def map_season(date):
# for date in dates:
if date.timetuple().tm_yday < spring_start or date.timetuple().tm_yday >= winter_start:
return "Winter"
elif date.timetuple().tm_yday >= spring_start and date.timetuple().tm_yday < summer_start:
return "Spring"
elif date.timetuple().tm_yday >= summer_start and date.timetuple().tm_yday < autumn_start:
return "Summer"
elif date.timetuple().tm_yday >= autumn_start and date.timetuple().tm_yday < winter_start:
return "Autumn"
data = data.with_columns(
pl.col(dt_col).map_elements(map_season, return_dtype=pl.String).alias(out_col)
)
return data
def create_weekend(data, dt_col="Datetime", out_col="is_weekend", prefix=""):
df_type = "pandas" if isinstance(data, pd.DataFrame) else "polars"
out_col = prefix + out_col
if df_type == "pandas":
data[out_col] = (data[dt_col].dt.weekday.isin([5,6])).astype(np.int8)
elif df_type == "polars":
data = data.with_columns(
pl.col(dt_col).dt.weekday().is_in([6,7]).cast(pl.Int8).alias(out_col)
)
return data
def create_holidays(data, dt_col="Datetime", out_col="is_holiday", prefix=""):
df_type = "pandas" if isinstance(data, pd.DataFrame) else "polars"
out_col = prefix + out_col
# The only holiday not included will be new years as I expect a potential affect
HOLIDAYS = [
pd.to_datetime("2016-01-18"), pd.to_datetime("2016-02-15"),
pd.to_datetime("2016-05-30"), pd.to_datetime("2016-07-04"), pd.to_datetime("2016-09-05"),
pd.to_datetime("2016-10-10"), pd.to_datetime("2016-11-11"), pd.to_datetime("2016-11-24"),
# Christmas is variable (depends on what day is actually holiday vs. what day is XMAS)
pd.to_datetime("2016-12-24"), pd.to_datetime("2016-12-25"), pd.to_datetime("2016-12-26"),
pd.to_datetime("2017-01-16"), pd.to_datetime("2017-02-20"),
pd.to_datetime("2017-05-29"), pd.to_datetime("2017-07-04"), pd.to_datetime("2017-09-04"),
pd.to_datetime("2017-10-09"), pd.to_datetime("2017-11-10"), pd.to_datetime("2017-11-23"),
pd.to_datetime("2017-12-24"), pd.to_datetime("2017-12-25"),
pd.to_datetime("2018-01-15"), pd.to_datetime("2018-02-19"),
pd.to_datetime("2018-05-28"), pd.to_datetime("2018-07-04"), pd.to_datetime("2018-09-03"),
pd.to_datetime("2018-10-08"), pd.to_datetime("2018-11-12"), pd.to_datetime("2018-11-22"),
pd.to_datetime("2018-12-24"), pd.to_datetime("2018-12-25"),
]
if df_type == "pandas":
data[out_col] = (data[dt_col].isin(HOLIDAYS)).astype(np.int8)
elif df_type == "polars":
data = data.with_columns(
pl.col(dt_col).dt.datetime().is_in(HOLIDAYS).cast(pl.Int8).alias(out_col)
)
return data
def build_temporal_features(data, dt_col, prefix=""):
df_type = "pandas" if isinstance(data, pd.DataFrame) else "polars"
if df_type == "pandas" and data.index.name == dt_col:
data = data.reset_index()
if df_type == "pandas":
data[prefix+"Year"] = data[dt_col].dt.year.astype(np.int16)
data[prefix+"Month"] = data[dt_col].dt.month.astype(np.int8)
data[prefix+"Day"] = data[dt_col].dt.day.astype(np.int8)
data[prefix+"DayOfYear"] = data[dt_col].dt.dayofyear.astype(np.int16)
data[prefix+"DayOfWeek"] = data[dt_col].dt.dayofweek.astype(np.int8)
else:
data = data.with_columns (**{
prefix+"Year": pl.col(dt_col).dt.year().cast(pl.Int16),
prefix+"Month": pl.col(dt_col).dt.month().cast(pl.Int8),
prefix+"Day": pl.col(dt_col).dt.day().cast(pl.Int8),
prefix+"DayOfYear": pl.col(dt_col).dt.ordinal_day().cast(pl.Int16),
prefix+"DayOfWeek": pl.col(dt_col).dt.weekday().cast(pl.Int8)
})
data = create_seasons(data, dt_col, prefix=prefix)
data = create_weekend(data, dt_col, prefix=prefix)
data = create_holidays(data, dt_col, prefix=prefix)
return data
def agg_and_merge_historical(curr_df, hist_df, col, agg_cols=[], ops=["mean", "max", "min"]):
merge_dict = {}
for agg_col in agg_cols:
describe_tb = hist_df.groupby(col)[agg_col].describe().reset_index()
if col not in merge_dict:
merge_dict[col] = describe_tb[col].values
for op in ops:
merge_col_name = "historical_" + col + "_" + op + "_" + agg_col
if op == "mean":
merge_dict[merge_col_name] = describe_tb["mean"].values
elif op == "max":
merge_dict[merge_col_name] = describe_tb["max"].values
elif op == "min":
merge_dict[merge_col_name] = describe_tb["min"].values
elif op == "median":
merge_dict[merge_col_name] = describe_tb["50%"].values
elif op == "std":
merge_dict[merge_col_name] = describe_tb["std"].values
merge_df = pd.merge(curr_df, pd.DataFrame(merge_dict), on=col, how="left")
return merge_df
def map_vals(data, cols=["Latitude", "Longitude"], label_cols=[], color="red", submap=None, weight=3, radius=1, sample_size=10000, start_loc=[42.1657, -74.9481], zoom_start=6):
cols = cols
df_type = "pandas" if isinstance(data, pd.DataFrame) or isinstance(data, pd.Series) else "polars"
fig = folium.Figure(height=500, width=750)
if submap is None:
map_nyc = folium.Map(
location=start_loc,
zoom_start=zoom_start,
tiles='cartodbpositron',
zoom_control=False,
scrollWheelZoom=False,
dragging=False
)
else:
map_nyc = submap
cols.extend(label_cols)
if df_type == "pandas":
for idx, row in data.loc[:, cols].dropna().sample(sample_size).iterrows():
label = ""
lat, long = row.iloc[0,], row.iloc[1,]
for i, label_col in enumerate(label_cols):
label += label_col + ": " + str(row.iloc[2+i,]) + "\n"
label_params = {"popup": label, "tooltip": label} if len(label_cols) > 0 else {}
folium.CircleMarker(location=[lat, long], radius=radius, weight=weight, color=color, fill_color=color, fill_opacity=0.7, **label_params).add_to(map_nyc)
else:
for row in data[:, cols].drop_nulls().sample(sample_size).rows():
label = ""
lat, long = row[0], row[1]
for i, label_col in enumerate(label_cols):
label += label_col + ": " + str(row[2+i]) + "\n"
label_params = {"popup": label, "tooltip": label} if len(label_cols) > 0 else {}
folium.CircleMarker(location=[lat, long], radius=radius, weight=weight, color=color, fill_color=color, fill_opacity=0.7, **label_params).add_to(map_nyc)
fig.add_child(map_nyc)
return fig, map_nyc
def find_variable_data(soup, curr_var = "Created Date"):
src = "<!doctype html>"
# HTML and head start
src += "<html lang=\"en\">"
src += str(soup.find("head"))
# Body -> content -> container -> row -> variable
src += "<body style=\"background-color: var(--table-odd-background-fill); padding-top: 20px;\">"
src += "<div class=\"content\" style=\"padding-left: 150px; padding-right: 150px; border: 0px !important; \">"
# src += "<div class=\"container\">"
src += "<div class=\"section-items\" style=\"background-color: white;\">"
# src += "<div class=\"row spacing\">"
variables_html = soup.find_all("div", class_="variable")
for var_html in variables_html:
if var_html.text[:len(curr_var)] == curr_var:
parent = var_html.parent
parent['style'] = "border: 0px"
src += str(parent)
break
src += "</div></div>"
# Scripts
for script in soup.find_all("script"):
src += str(script)
# End
src += "</body>"
src += "</html>"
# src = BeautifulSoup(src, 'html.parser').prettify()
src_doc = html.escape(src)
iframe = f'<iframe width="100%" height="1200px" srcdoc="{src_doc}" frameborder="0"></iframe>'
return iframe, src_doc
def plot_autocorr(data, col, apply=None):
time_series = data.loc[:, col].to_frame().copy()
if apply:
time_series[col] = time_series[col].apply(apply)
fig, ax = plt.subplots(2, 1, figsize=(12, 8))
_ = plot_acf(time_series[col], lags=30, ax=ax[0])
_ = plot_pacf(time_series[col], lags=30, method="ols-adjusted", ax=ax[1])
_ = plt.suptitle(f"{col}", y=0.95)
return fig
def adf_test(timeseries):
dftest = adfuller(timeseries, autolag='AIC')
dfoutput = pd.Series(dftest[0:4], index=['Test Statistic','p-value','Lags Used','Number of Observations Used'])
dfoutput['Number of Observations Used'] = dfoutput['Number of Observations Used'].astype(np.int64)
for key,value in dftest[4].items():
dfoutput['Critical Value (%s)'%key] = value
return dfoutput
def kpss_test(timeseries):
kpsstest = kpss(timeseries, regression='ct')
kpss_output = pd.Series(kpsstest[0:3], index=['Test Statistic','p-value','Lags Used'])
for key,value in kpsstest[3].items():
kpss_output['Critical Value (%s)'%key] = value
return kpss_output
def test_stationary(data, var):
adf_df = adf_test(data[var].dropna())
kpss_df = kpss_test(data[var].dropna())
result_df = adf_df.to_frame(name="Augmented-Dickey-Fuller")
result_df["KPSS Test"] = kpss_df
def pass_hypothesis(col):
test_stat, p_val = col.iloc[0], col.iloc[1]
one_p, five_p, ten_p = col.iloc[4], col.iloc[5], col.iloc[6]
if col.name == "KPSS Test":
if test_stat < one_p and p_val < 0.01:
color_fmt = ["background-color: #fc5749; font-weight: bold; color: black"]
elif test_stat < five_p and p_val < 0.05:
color_fmt = ["background-color: #F88379; font-weight: bold; color: black"]
elif test_stat < ten_p and p_val < 0.1:
color_fmt = ["background-color: #ff9f96; font-weight: bold; color: black"]
else:
color_fmt = ["background-color: green; font-weight: bold; color: black"]
else:
if test_stat < one_p and p_val < 0.01:
color_fmt = ["background-color: green; font-weight: bold; color: black"]
elif test_stat < five_p and p_val < 0.05:
color_fmt = ["background-color: greenyellow; font-weight: bold; color: black"]
elif test_stat < ten_p and p_val < 0.1:
color_fmt = ["background-color: lightgreen; font-weight: bold; color: black"]
else:
color_fmt = ["background-color: #fc5749; font-weight: bold; color: black"]
color_fmt.extend(['' for _ in col[1:]])
return color_fmt
result_df.loc["Lags Used",:] = result_df.loc["Lags Used",:].astype(np.int32)
return result_df.style.apply(pass_hypothesis)
def plot_timeseries(data, var, data_name="My", all_vars=[], height=800, width=600, start_date="2017-12-31", end_date="2018-12-31"):
if var == "":
return gr.update()
fig = go.Figure()
fig.add_trace(
go.Scatter(
x=data.index,
y=data[var],
name=var,
customdata=np.dstack((data["Season"].to_numpy(), data.reset_index()["Datetime"].dt.day_name().to_numpy(), data["is_holiday"].astype(bool).to_numpy()))[0],
hovertemplate='<br>value:%{y:.3f} <br>Season: %{customdata[0]} <br>Weekday: %{customdata[1]} <br>Is Holiday: %{customdata[2]}',
)
)
fig.update_layout(
autosize=True,
title=f"{data_name} Time Series by {var}",
xaxis_title='Date',
yaxis_title=var,
hovermode='x unified',
)
fig.update_layout(
autosize=True,
xaxis=dict(
rangeselector=dict(
buttons=list([
dict(count=7, label="1w", step="day", stepmode="backward"),
dict(count=21, label="3w", step="day", stepmode="backward"),
dict(count=1, label="1m", step="month", stepmode="backward"),
dict(count=6, label="6m", step="month", stepmode="backward"),
dict(count=1, label="1y", step="year", stepmode="backward"),
dict(step="all")
])
),
rangeslider=dict(
visible=True,
#
),
type="date",
range=(start_date, end_date),
),
)
return fig
def plot_bivariate(data, x, y, subset=None, trendline=True):
title = f"Scatterplot of {x} vs. {y}"
if subset == "None" or subset is None:
subset = None
height = 450
else:
subset_title = subset.replace(" String","")
title += f" By {subset_title}"
if subset_title in ["Season", "Year"]:
height = 450
else:
height = 800
if trendline:
trendline = "ols"
else:
trendline = None
# Special case to view categorical features
if x in ["Agency", "Borough", "Descriptor"]:
if x == "Agency":
prefix = 'AG'
elif x == "Borough":
prefix = "Borough"
else:
prefix="DG"
categories = [col for col in data.columns if prefix in col]
melt_df = pd.melt(data, id_vars=["Target"], value_vars=categories)
fig = px.scatter(
melt_df,
x="value",
y="Target",
trendline=trendline,
facet_col="variable",
facet_col_wrap=4,
facet_col_spacing=0.05,
title=title
)
height = 800
else:
fig = px.scatter(
data,
x=x, y=y,
trendline=trendline,
facet_col=subset,
facet_col_wrap=4,
facet_col_spacing=0.05,
title=title
)
fig.update_layout(
autosize=True,
height=height,
)
return fig
def plot_seasonality(data, x, y, show_box=True, show_outliers=False):
title = f"{y} by {x}"
if show_box:
if show_outliers:
points = "outliers"
else:
points = "all"
fig = px.box(data, x=x, y=y, points=points, title=title, facet_col_wrap=4, facet_col_spacing=0.05,)
else:
fig = px.strip(data, x=x, y=y, title=title, facet_col_wrap=4, facet_col_spacing=0.05,)
fig.update_layout(
autosize=True,
height=600,
)
return fig
def build_service_data(filename):
# Loading data directly with polars leads to errors
# Some rows end up missing for an unknown reason
# FIX: Load in pandas then convert to polars
service_data_pd = pd.read_csv(filename)
# Quick test to assure the unique key is in fact unique
assert service_data_pd["Unique Key"].nunique() == len(service_data_pd)
# Load from pandas Dataframe
service_data_pd["Incident Zip"] = service_data_pd["Incident Zip"].astype("string")
service_data_pd["BBL"] = service_data_pd["BBL"].astype("string")
service_data = pl.DataFrame(service_data_pd)
# Clear some ram
del service_data_pd
gc.collect()
drop_cols = [
"Unique Key", "Agency Name", "Location Type", "Incident Zip",
"Incident Address", "Street Name", "Cross Street 1",
"Cross Street 2", "Intersection Street 1", "Intersection Street 2",
"Address Type", "City", "Landmark", "Facility Type",
"Status", "Due Date", "Resolution Description",
"Resolution Action Updated Date", "Community Board",
"BBL", "X Coordinate (State Plane)", "Y Coordinate (State Plane)",
"Open Data Channel Type", "Park Facility Name", "Park Borough",
"Vehicle Type", "Taxi Company Borough", "Taxi Pick Up Location",
"Bridge Highway Name", "Bridge Highway Direction", "Road Ramp",
"Bridge Highway Segment", "Location", "Created Year"
]
# Drop columns and create the date variable
service_data = service_data.drop(drop_cols)
service_data = create_datetime(service_data, "Created Date")
service_data = create_datetime(service_data, "Closed Date")
# Group by date to get the number of Created tickets (as target)
sd_grouped = service_data.rename({"Created Date": "Datetime"}).group_by("Datetime").agg(
pl.len().alias("Target"),
).sort(by="Datetime")
# Calculate the number of closed tickets
# Mean diff used to filter service data
# mean_diff = service_data.with_columns(
# diff_created_closed = pl.col("Closed Date") - pl.col("Created Date")
# ).filter((pl.col("Closed Date").dt.year() >= 2016) & (pl.col("Closed Date").dt.year() < 2020))["diff_created_closed"].mean().days
# Mean diff precalculated as
mean_diff = 13
# Create new Closed date with errors filled using the mean diff above
service_data = service_data.with_columns(
Closed_Date_New = pl.when(pl.col("Created Date") - pl.col("Closed Date") > pl.duration(days=1))
.then(pl.col("Created Date") + pl.duration(days=mean_diff))
.otherwise(pl.col("Closed Date")).fill_null(pl.col("Created Date") + pl.duration(days=mean_diff))
)
# Filter tickets such that the closed date < the created date to prevent future data leakage in our dataset
# We want to make sure future data is not accidentally leaked across other points in our data
closed_tickets = service_data.group_by(["Closed_Date_New", "Created Date"]) \
.agg((pl.when(pl.col("Created Date") <= pl.col("Closed_Date_New")).then(1).otherwise(0)).sum().alias("count")) \
.sort("Closed_Date_New") \
.filter((pl.col("Closed_Date_New").dt.year() >= 2016) & (pl.col("Closed_Date_New").dt.year() < 2019)) \
.group_by("Closed_Date_New").agg(pl.col("count").sum().alias("num_closed_tickets"))
# Rename this column to num closed tickets
ct_df = closed_tickets.with_columns(
pl.col("num_closed_tickets")
)
# Concat the new columns into our data
sd_df = pl.concat([sd_grouped, ct_df.drop("Closed_Date_New")], how="horizontal")
assert len(sd_grouped) == len(ct_df)
# CATEGORICAL FEATURE MAPPING
# MAPPING FOR BOROUGH
Borough_Map = {
"Unspecified": "OTHER",
"2017": "OTHER",
None: "OTHER",
"2016": "OTHER"
}
service_data = service_data.with_columns(
pl.col("Borough").replace(Borough_Map)
)
# MAPPING FOR AGENCY
# This mapping was done Manually
Agency_Map = {
"NYPD": "Security", "HPD": "Buildings", "DOT": "Transportation",
"DSNY": "Environment & Sanitation", "DEP": "Environment & Sanitation",
"DOB": "Buildings", "DOE": "Buildings", "DPR": "Parks",
"DOHMH": "Health", "DOF": "Other", "DHS": "Security",
"TLC": "Transportation", "HRA": "Other", "DCA": "Other",
"DFTA": "Other", "EDC": "Other", "DOITT": "Other", "OMB": "Other",
"DCAS": "Other", "NYCEM": "Other", "ACS": "Other", "3-1-1": "Other",
"TAX": "Other", "DCP": "Other", "DORIS": "Other", "FDNY": "Other",
"TAT": "Other", "COIB": "Other", "CEO": "Other", "MOC": "Other",
}
service_data = service_data.with_columns(
pl.col("Agency").replace(Agency_Map).alias("AG") # AG Shorthand for Agency Groups
)
# Mapping for Descriptor using BERTopic
# Store descriptors as pandas dataframe (polars not supported)
# Drop any nan values, and we only care about the unique values
descriptor_docs = service_data["Descriptor"].unique().to_numpy()
# Build our topic mapping using the pretrained BERTopic model
# Load model and get predictions
topic_model = BERTopic.load("models/BERTopic")
topics, probs = topic_model.transform(descriptor_docs)
# Visualize if wanted
# topic_model.visualize_barchart(list(range(-1,6,1)))
# Create a topic to ID map
topic_df = topic_model.get_topic_info()
topic_id_map = {row["Topic"]: row["Name"][2:] for _, row in topic_df.iterrows()}
topic_id_map[-1] = topic_id_map[-1][1:] # Fix for the -1 topic case
# For each document (descriptor string) get a mapping of topics
doc_to_topic_map = defaultdict(str)
for topic_id, doc in zip(topics, descriptor_docs):
topic = topic_id_map[topic_id]
doc_to_topic_map[doc] = topic
service_data = service_data.with_columns(
pl.col("Descriptor").replace(doc_to_topic_map).alias("DG") # DG Shorthand for descriptor Groups
)
# One Hot Encode Features
cat_features = ["AG", "Borough", "DG"]
service_data = service_data.to_dummies(columns=cat_features)
# Group by Date and create our Category Feature Vector
cat_df = service_data.rename({"Created Date": "Datetime"}).group_by("Datetime").agg(
# Categorical Features Sum
pl.col('^AG_.*$').sum(),
pl.col('^Borough_.*$').sum(),
pl.col('^DG_.*$').sum(),
).sort(by="Datetime")
# Concat our category features to our current dataframe
sd_df = pl.concat([sd_df, cat_df.drop("Datetime")], how="horizontal")
# Now that our dataframe is significantly reduced in size
# We can finally convert back to a pandas dataframe
# as pandas is usable across more python packages
sd_df = sd_df.to_pandas()
# Set index to datetime
sd_df = sd_df.set_index("Datetime")
# NOTE we added 7 new rows to our weather df
# These 7 new rows will essentially be our final pred set
# The Target for these rows will be null -> indicating it needs to be predicted
# Add these rows to the service dataframe
preds_df = pd.DataFrame({'Datetime': pd.date_range(start=sd_df.index[-1], periods=8, freq='D')})[1:]
sd_df = pd.concat([sd_df, preds_df.set_index("Datetime")], axis=0)
return sd_df
# Build all weather data from file
def build_weather_data(filename):
# Use pandas to read file
weather_data = pd.read_csv(filename)
# Quickly aggregate Year, Month, Day into a datetime object
# This is because the 311 data uses datetime
weather_data["Datetime"] = weather_data["Year"].astype("str") + "-" + weather_data["Month"].astype("str") + "-" + weather_data["Day"].astype("str")
weather_data = create_datetime(weather_data, "Datetime", format="%Y-%m-%d")
# LOCALIZE
# Pre-recorded min/max values from the service data (so we don't need again)
lat_min = 40.49804421521046
lat_max = 40.91294056699566
long_min = -74.25521082506387
long_max = -73.70038354802529
# Create the conditions for location matching
mincon_lat = weather_data["Latitude"] >= lat_min
maxcon_lat = weather_data["Latitude"] <= lat_max
mincon_long = weather_data["Longitude"] >= long_min
maxcon_long = weather_data["Longitude"] <= long_max
# Localize our data to match the service data
wd_localized = weather_data.loc[mincon_lat & maxcon_lat & mincon_long & maxcon_long]
drop_cols = [
"USAF",
"WBAN",
"StationName",
"State",
"Latitude",
"Longitude"
]
wd_localized = wd_localized.drop(columns=drop_cols)
# AGGREGATE
# Map columns with aggregation method
mean_cols = [
'MeanTemp',
'DewPoint',
'Percipitation',
'WindSpeed',
'Gust',
'SnowDepth',
]
min_cols = [
'MinTemp'
]
max_cols = [
'MaxTemp',
'MaxSustainedWind'
]
round_cols = [
'Rain',
'SnowIce'
]
# Perform Aggregation
mean_df = wd_localized.groupby("Datetime")[mean_cols].mean()
min_df = wd_localized.groupby("Datetime")[min_cols].min()
max_df = wd_localized.groupby("Datetime")[max_cols].max()
round_df = wd_localized.groupby("Datetime")[round_cols].mean().round().astype(np.int8)
wd_full = pd.concat([mean_df, min_df, max_df, round_df], axis=1)
# Add seasonal features
wd_full = build_temporal_features(wd_full, "Datetime")
wd_full["Season"] = wd_full["Season"].astype("category")
wd_full = wd_full.set_index("Datetime")
# We will calculate the imputation for the next 7 days after 12/31/2018
# Along with the 49 missing days
# This will act as our "Weather Forecast"
time_steps = 49 + 7
# Impute Cols
impute_cols = [
'MeanTemp', 'MinTemp', 'MaxTemp', 'DewPoint',
'Percipitation', 'WindSpeed', 'MaxSustainedWind',
'Gust', 'Rain', 'SnowDepth', 'SnowIce',
]
# Mean Vars
mean_vars = ["WindSpeed", "MaxSustainedWind", "Gust", "SnowDepth"]
min_vars = ["SnowIce", "MeanTemp", "MinTemp", "MaxTemp", "DewPoint", "Percipitation"]
max_vars = ["Rain"]
# Use the imported function to create the imputed data
preds_mean = impute_missing_weather(wd_full, strategy="mean", time_steps=time_steps, impute_cols=mean_vars)
preds_min = impute_missing_weather(wd_full, strategy="min", time_steps=time_steps, impute_cols=min_vars)
preds_max = impute_missing_weather(wd_full, strategy="max", time_steps=time_steps, impute_cols=max_vars)
all_preds = pd.concat([preds_mean, preds_min, preds_max], axis=1)
all_preds = build_temporal_features(all_preds.loc[:, impute_cols], "Datetime")
all_preds = all_preds.set_index("Datetime")
wd_curr = wd_full.loc[wd_full["Year"] >= 2016]
wd_df = pd.concat([wd_full, all_preds], axis=0, join="outer")
time_vars = ["Year", "Month", "Day", "DayOfWeek", "DayOfYear", "is_weekend", "is_holiday", "Season"]
wd_df.drop(columns=time_vars)
return wd_df
class MyNaiveImputer():
def __init__(self, data, time_steps=49, freq="D"):
self.data = data.reset_index().copy()
start_date = self.data["Datetime"].max() + pd.Timedelta(days=1)
end_date = start_date + pd.Timedelta(days=time_steps-1)
missing_range = pd.date_range(start_date, end_date, freq="D")
self.missing_df = pd.DataFrame(missing_range, columns=["Datetime"])
self.missing_df = build_temporal_features(self.missing_df, "Datetime")
def impute(self, col, by="DayOfYear", strategy="mean"):
def naive_impute_by(val, impute_X, data, by=by, strategy=strategy):
if strategy.lower() == "mean":
func = pd.core.groupby.DataFrameGroupBy.mean
elif strategy.lower() == "median":
func = pd.core.groupby.DataFrameGroupBy.median
elif strategy.lower() == "max":
func = pd.core.groupby.DataFrameGroupBy.max
elif strategy.lower() == "min":
func = pd.core.groupby.DataFrameGroupBy.min
grouped = func(data.groupby(by)[impute_X])
return grouped[val]
return self.missing_df["DayOfYear"].apply(naive_impute_by, args=(col, self.data, by, strategy))
def impute_all(self, cols, by="DayOfYear", strategy="mean"):
output_df = self.missing_df.copy()
for col in cols:
output_df[col] = self.impute(col, by, strategy)
return output_df
def impute_missing_weather(data, strategy="mean", time_steps=7, impute_cols=impute_cols):
final_imputer = MyNaiveImputer(data, time_steps=time_steps)
preds = final_imputer.impute_all(impute_cols, strategy=strategy).set_index("Datetime")
return preds
def get_feature_importance(data, target, split_date="01-01-2016", print_score=False):
import torch
device = "cuda" if torch.cuda.is_available() else "cpu"
train = data.loc[data.index <= pd.to_datetime(split_date)]
test = data.loc[data.index > pd.to_datetime(split_date)]
if type(target) == str:
X_train, X_test = train.drop(columns=target), test.drop(columns=target)
y_train, y_test = train[target], test[target]
else:
X_train, X_test = train, test
y_train, y_test = target.loc[train.index], target.loc[test.index]
target = str(target.name)
if 'int' in y_train.dtype.name:
# Use binary Classifier
metric = "logloss"
model = xgb.XGBClassifier(
base_score=0.25,
n_estimators=500,
early_stopping_rounds=50,
objective='binary:logistic',
device=device,
max_depth=3,
learning_rate=0.01,
enable_categorical=True,
eval_metric="logloss",
importance_type="gain",
random_state=22,
)
else:
metric = "MAPE"
model = xgb.XGBRegressor(
n_estimators=500,
early_stopping_rounds=50,
objective='reg:squarederror',
device=device,
max_depth=3,
learning_rate=0.01,
enable_categorical=True,
eval_metric="mape",
importance_type="gain",
random_state=22,
)
_ = model.fit(
X_train, y_train,
eval_set=[(X_train, y_train), (X_test, y_test)],
verbose=False
)
fig, ax = plt.subplots()
ax = plot_importance(model, title=f"Feature Importance for {target}", ax=ax)
if print_score:
best_score = str(round(100*model.best_score,2))+"%"
print(f"Best {metric}: {best_score}")
return fig, model
def corr_with_lag(data, target_col, covar, lags=[1], method="pearson"):
data_lagged = pd.DataFrame()
data_lagged["Target"] = data[target_col]
for lag in lags:
new_col = f"lag_{lag}D"
data_lagged[new_col] = data[covar].shift(lag)
return data_lagged.dropna().corr(method=method)
def plot_correlations(data, target, covar, lags=[0,1,2,3,4,5,6,7,10,14,18,21], method="pearson"):
df_corr = corr_with_lag(data, target, covar, lags, method)
mask = np.triu(np.ones_like(df_corr, dtype=bool))
z_dim, x_dim = len(df_corr.to_numpy()), len(df_corr.columns)
y_dim = x_dim
fig = ff.create_annotated_heatmap(
z=df_corr.mask(mask).to_numpy(),
x=df_corr.columns.tolist(),
y=df_corr.columns.tolist(),
colorscale=px.colors.diverging.RdBu,
zmin=-1,
zmax=1,
ygap=2,
xgap=2,
name="",
customdata=np.full((x_dim, y_dim, z_dim), covar),
hovertemplate='%{customdata[0]}<br>%{x} to %{y}<br>Correlation: %{z:.4f}',
showscale=True
)
fig.update_layout(
title_text=f"Correlation Heatmap of Lagged {covar}",
title_x=0.5,
height=600,
xaxis_showgrid=False,
yaxis_showgrid=False,
xaxis_zeroline=False,
yaxis_zeroline=False,
yaxis_autorange='reversed',
template='plotly_white'
)
# fig.update_annotations(font=dict(color="black"))
for i in range(len(fig.layout.annotations)):
if fig.layout.annotations[i].text == 'nan':
fig.layout.annotations[i].text = ""
else:
corr_i = round(float(fig.layout.annotations[i].text), 3)
fig.layout.annotations[i].text = corr_i
if (corr_i > 0.2 and corr_i < 0.5) or (corr_i < -0.2 and corr_i > -0.5):
fig.layout.annotations[i].font.color = "white"
return fig
def plot_all_correlations(data, data_name="weather", method="pearson", width=1392, height=600):
if data_name == "weather":
covars = ["MeanTemp", "MinTemp", "MaxTemp", 'DewPoint', 'Percipitation', 'WindSpeed', 'Gust', 'MaxSustainedWind', "SnowDepth", "SnowIce", "Rain", "Target"]
elif data_name == "service":
covars = [
"num_closed_tickets",
# Agency Group Counts
'AG_Buildings', 'AG_Environment & Sanitation', 'AG_Health',
'AG_Parks', 'AG_Security', 'AG_Transportation',
'AG_Other',
# Borough Counts
'Borough_BRONX', 'Borough_BROOKLYN', 'Borough_MANHATTAN',
'Borough_QUEENS', 'Borough_STATEN ISLAND',
'Borough_OTHER',
# Descriptor Group Counts
'DG_damaged_sign_sidewalk_missing',
'DG_english_emergency_spanish_chinese',
'DG_exemption_commercial_tax_business',
'DG_license_complaint_illegal_violation', 'DG_noise_animal_truck_dead',
'DG_odor_food_air_smoke', 'DG_order_property_inspection_condition',
'DG_water_basin_litter_missed', "Target"
]
df_corr = data.loc[:, covars].corr(method=method)
mask = np.triu(np.ones_like(df_corr, dtype=bool))
fig = ff.create_annotated_heatmap(
z=df_corr.mask(mask).to_numpy(),
x=df_corr.columns.tolist(),
y=df_corr.columns.tolist(),
colorscale=px.colors.diverging.RdBu,
zmin=-1,
zmax=1,
ygap=2,
xgap=2,
name="",
hovertemplate='%{x}-%{y} <br>Correlation: %{z:.4f}',
showscale=True
)
fig.update_layout(
title_text=f"Correlation Heatmap of Weather Variables & Target",
title_x=0.5,
height=600,
width=width,
xaxis_showgrid=False,
yaxis_showgrid=False,
xaxis_zeroline=False,
yaxis_zeroline=False,
yaxis_autorange='reversed',
template='plotly_white'
)
fig.update_annotations(font=dict(color="black"))
for i in range(len(fig.layout.annotations)):
if fig.layout.annotations[i].text == 'nan':
fig.layout.annotations[i].text = ""
else:
corr_i = round(float(fig.layout.annotations[i].text), 3)
fig.layout.annotations[i].text = corr_i
if corr_i > 0.5 or corr_i < -0.5:
fig.layout.annotations[i].font.color = "white"
return fig
def plot_gust_interpolation(data):
fig, ax = plt.subplots(2, 2, figsize=(15,12))
data["Gust_lin"].plot(ax=ax[0][0], color=color_pal[0], title="linear")
data["Gust_spline3"].plot(ax=ax[0][1], color=color_pal[1], title="spline3")
data["Gust_spline5"].plot(ax=ax[1][0], color=color_pal[2], title="spline5")
data["Gust_quad"].plot(ax=ax[1][1], color=color_pal[3], title="quadratic")
curr_fig = plt.gcf()
plt.close()
return curr_fig
def plot_train_split(train, val):
fig = plt.subplots(figsize=(15, 5))
ax = train["Target"].plot(label="Training Set")
val["Target"].plot(label="Validation Set", ax=ax)
ax.axvline('2018-04-01', color='black', ls='--')
ax.legend()
ax.set_title("Train Test Split (2018-04-01)")
curr_fig = plt.gcf()
plt.close()
return curr_fig
def plot_predictions(train, val, preds):
fig = plt.subplots(figsize=(16, 5))
ax = train["Target"].plot(label="Training Set")
val["Target"].plot(label="Validation Set", ax=ax)
val["Prediction"] = preds
val["Prediction"].plot(label="Prediction", ax=ax)
ax.axvline('2018-04-01', color='black', ls='--')
ax.legend()
ax.set_title("Model Prediction for 311 Call Volume")
curr_fig = plt.gcf()
plt.close()
return curr_fig
def plot_final_feature_importance(model):
fig, ax = plt.subplots(figsize=(12,6))
ax = plot_importance(model, max_num_features=20, title=f"Feature Importance for 311 Service Calls", ax=ax)
curr_fig = plt.gcf()
plt.close()
return curr_fig
def predict_recurse(dataset, test, model, features_to_impute=['Target_L1D', 'Target_Diff7D', 'Target_Diff14D'], last_feature='Target_L6D'):
n_steps = len(test)
merged_data = pd.concat([dataset[-14:], test], axis=0)
all_index = merged_data.index
X_test = test.drop(columns="Target")
sd = -6 # Starting point for filling next value
# For each step, get the predictions
for i in range(n_steps-1):
pred = model.predict(X_test)[i]
# For the three features needed, compute the new value
X_test.loc[all_index[sd+i], features_to_impute[0]] = pred
X_test.loc[all_index[sd+i], features_to_impute[1]] = pred - merged_data.loc[all_index[sd+i-7], features_to_impute[1]]
X_test.loc[all_index[sd+i], features_to_impute[2]] = pred - merged_data.loc[all_index[sd+i-14], features_to_impute[2]]
# In the last iteration compute the Lag6D value
if i == 5:
X_test.loc[all_index[sd+i], last_feature] = pred - merged_data.loc[all_index[sd+i-6], last_feature]
final_preds = model.predict(X_test)
return final_preds