Agile Data Science 2.0

Building Full Stack Data Analytics Applications with Kafka and Spark
Agile Data Science 2.0
http://bit.ly/agile_data_slides_6

Russell Jurney
2
Data Engineer
Data Scientist
Visualization Software Engineer
85%
85%
85%
Writer
85%
Teacher
50%
Russell Jurney is a veteran data
scientist and thought leader. He
coined the term Agile Data Science in
the book of that name from O’Reilly
in 2012, which outlines the first agile
development methodology for data
science. Russell has constructed
numerous full-stack analytics
products over the past ten years and
now works with clients helping them
extract value from their data assets.
Russell Jurney
Skill
Principal Consultant at Data Syndrome
Russell Jurney
Data Syndrome, LLC
Email : russell.jurney@gmail.com
Web : datasyndrome.com
Principal Consultant

Lorem Ipsum dolor siamet suame this placeholder for text can simply
random text. It has roots in a piece of classical. variazioni deiwords which
whichhtly. ven on your zuniga merida della is not denis.
Product Consulting
We build analytics products and systems
consisting of big data viz, predictions,
recommendations, reports and search.
Corporate Training
We offer training courses for data
scientists and engineers and data
science teams,
Video Training
We offer video training courses that rapidly
acclimate you with a technology and
technique.

Agile Data Science 2.0 4
What makes data science “agile data science”?
Theory

Agile Manifesto
5
Four principles for software engineering

Yes. Building applications is a fundamental skill for today’s data scientist.
Data Products or Data Science?

Data Products
or
Data Science?

If someone else has to start over and rebuild it, it ain’t agile.
Big Data or Data Science?

Goal of
Methodology
The goal of agile data science in <140 characters: to
document and guide exploratory data analysis to
discover and follow the critical path to a compelling
product.

Agile Data Science Manifesto
10
Seven Principles for Agile Data Science
Discover and pursue the critical path to a killer product
Iterate, iterate, iterate: tables, charts, reports, predictions1.
Integrate the tyrannical opinion of data in product management4.
Get Meta. Describe the process, not just the end-state7.
Ship intermediate output. Even failed experiments have output2.
Climb up and down the data-value pyramid as we work5.
Prototype experiments over implementing tasks3.
6.

Iterate, iterate, iterate: tables, charts, reports, predictions
11

Ship intermediate output. Even failed experiments have output.
12
—>

Prototype experiments over implementing tasks
13

Climb up and down the data-value pyramid
14
People will pay more for the things towards the top, but you need the things
on the bottom to have the things above. They are foundational. See:
Maslow’s Theory of Needs.
Data Value Pyramid

Discover and pursue the critical path to a killer product
15
In analytics, the end-goal moves or is complex in nature, so we model as a
network of tasks rather than as a strictly linear process.
Critical Path

Get Meta. Describe the process, not just the end state
16

Doing Agile Wrong
17
iterating between sprints, not within them is a bad idea.

Data Science Teams
18
The roles in a data science team are many, necessitating generalists

Data Science Releases
19
Snowballing releases towards a broader audience

Decorate your environment by visualizing your dataset!
Collage

Things we use to build the apps
Tools

Agile Data Science 2.0 Stack
22
Apache Spark Apache Kafka MongoDB
Batch and Realtime
Realtime Queue Document Store
Flask
Simple Web App
Example of a high productivity stack for “big” data applications
ElasticSearch
Search
d3.js
Visualization

Flow of Data Processing
23
Tools and processes in collecting, refining, publishing and decorating data
{“hello”: “world”}

Architecture of Agile Data Science
24
How the system comes together

Data Syndrome: Agile Data Science 2.0
Apache Spark Ecosystem
25
HDFS, Amazon S3, Spark, Spark SQL, Spark MLlib, Spark Streaming
/

SQL or dataflow programming?
Programming Models

Describing what you want and letting the planner figure out how
SQL
SELECT associations2.object_id,
associations2.term_id, associations2.cat_ID,
associations2.term_taxonomy_id 
FROM (SELECT objects_tags.object_id,
objects_tags.term_id, wp_cb_tags2cats.cat_ID,
categories.term_taxonomy_id 
FROM (SELECT
wp_term_relationships.object_id,
wp_term_taxonomy.term_id,
wp_term_taxonomy.term_taxonomy_id 
FROM wp_term_relationships 
LEFT JOIN wp_term_taxonomy ON
wp_term_relationships.term_taxonomy_id =
ORDER BY object_id ASC, term_id ASC)  
AS objects_tags 
LEFT JOIN wp_cb_tags2cats ON
objects_tags.term_id = wp_cb_tags2cats.tag_ID 
LEFT JOIN (SELECT
wp_term_relationships.object_id,
wp_term_taxonomy.term_id as cat_ID,
FROM wp_term_relationships 
LEFT JOIN wp_term_taxonomy ON
wp_term_relationships.term_taxonomy_id =
WHERE wp_term_taxonomy.taxonomy =
'category' 
GROUP BY object_id, cat_ID,
term_taxonomy_id 
ORDER BY object_id, cat_ID,
term_taxonomy_id)  
AS categories on wp_cb_tags2cats.cat_ID
= categories.term_id 
WHERE objects_tags.term_id =
wp_cb_tags2cats.tag_ID 
GROUP BY object_id, term_id, cat_ID,
term_taxonomy_id 
ORDER BY object_id ASC, term_id ASC, cat_ID
ASC)  
AS associations2 
LEFT JOIN categories ON associations2.object_id
= categories.object_id 
WHERE associations2.cat_ID <> categories.cat_ID 
GROUP BY object_id, term_id, cat_ID,
term_taxonomy_id 
ORDER BY object_id, term_id, cat_ID,
term_taxonomy_id

Flowing data through operations to effect change
Dataflow Programming

The best of both worlds!
SQL AND
Dataflow
Programming
# Flights that were late arriving... 
late_arrivals =
on_time_dataframe.filter(on_time_dataframe.ArrD
elayMinutes > 0) 
total_late_arrivals = late_arrivals.count() 
 
# Flights that left late but made up time to
arrive on time... 
on_time_heros = on_time_dataframe.filter( 
(on_time_dataframe.DepDelayMinutes > 0) 
& 
(on_time_dataframe.ArrDelayMinutes <= 0) 
) 
total_on_time_heros = on_time_heros.count() 
 
# Get the percentage of flights that are late,
rounded to 1 decimal place 
pct_late = round((total_late_arrivals /
(total_flights * 1.0)) * 100, 1) 
 
print("Total flights:
{:,}".format(total_flights)) 
print("Late departures:
{:,}".format(total_late_departures)) 
print("Late arrivals:
{:,}".format(total_late_arrivals)) 
print("Recoveries:
{:,}".format(total_on_time_heros)) 
print("Percentage Late: {}%".format(pct_late)) 
 
# Why are flights late? Lets look at some
delayed flights and the delay causes 
late_flights = spark.sql(""" 
SELECT 
ArrDelayMinutes, 
WeatherDelay, 
CarrierDelay, 
NASDelay, 
SecurityDelay, 
LateAircraftDelay 
FROM 
on_time_performance 
WHERE 
WeatherDelay IS NOT NULL 
OR 
CarrierDelay IS NOT NULL 
OR 
NASDelay IS NOT NULL 
OR 
SecurityDelay IS NOT NULL 
OR 
LateAircraftDelay IS NOT NULL 
ORDER BY 
FlightDate 
""") 
late_flights.sample(False, 0.01).show()
# Calculate the percentage contribution to delay for each source 
total_delays = spark.sql(""" 
SELECT 
ROUND(SUM(WeatherDelay)/SUM(ArrDelayMinutes) * 100, 1) AS
pct_weather_delay, 
ROUND(SUM(CarrierDelay)/SUM(ArrDelayMinutes) * 100, 1) AS
pct_carrier_delay, 
ROUND(SUM(NASDelay)/SUM(ArrDelayMinutes) * 100, 1) AS pct_nas_delay, 
ROUND(SUM(SecurityDelay)/SUM(ArrDelayMinutes) * 100, 1) AS
pct_security_delay, 
ROUND(SUM(LateAircraftDelay)/SUM(ArrDelayMinutes) * 100, 1) AS
pct_late_aircraft_delay 
FROM on_time_performance 
""") 
total_delays.show() 
 
# Generate a histogram of the weather and carrier delays 
weather_delay_histogram = on_time_dataframe 
.select("WeatherDelay") 
.rdd 
.flatMap(lambda x: x) 
.histogram(10) 
 
print("{}n{}".format(weather_delay_histogram[0],
weather_delay_histogram[1])) 
 
# Eyeball the first to define our buckets 
weather_delay_histogram = on_time_dataframe 
.select("WeatherDelay") 
.rdd 
.flatMap(lambda x: x) 
.histogram([1, 15, 30, 60, 120, 240, 480, 720, 24*60.0]) 
print(weather_delay_histogram)
# Transform the data into something easily consumed by d3 
record = {'key': 1, 'data': []} 
for label, count in zip(weather_delay_histogram[0],
weather_delay_histogram[1]): 
record['data'].append( 
{ 
'label': label, 
'count': count 
} 
) 
 
# Save to Mongo directly, since this is a Tuple not a dataframe or RDD 
from pymongo import MongoClient 
client = MongoClient() 
client.relato.weather_delay_histogram.insert_one(record)

Batch processing still dominant for charts, reports, search indexes and
training predictive models
Apache Airflow
for
Batch Scheduling

FAA on-time performance data
Data

Collect and Serialize Events in JSON
I never regret using JSON
32

FAA On-Time Performance Records
95% of commercial flights
33http://www.transtats.bts.gov/Fields.asp?table_id=236

FAA On-Time Performance Records
95% of commercial flights
34
"Year","Quarter","Month","DayofMonth","DayOfWeek","FlightDate","UniqueCarrier","AirlineID","Carrier","TailNum","FlightNum",
"OriginAirportID","OriginAirportSeqID","OriginCityMarketID","Origin","OriginCityName","OriginState","OriginStateFips",
"OriginStateName","OriginWac","DestAirportID","DestAirportSeqID","DestCityMarketID","Dest","DestCityName","DestState",
"DestStateFips","DestStateName","DestWac","CRSDepTime","DepTime","DepDelay","DepDelayMinutes","DepDel15","DepartureDelayGroups",
"DepTimeBlk","TaxiOut","WheelsOff","WheelsOn","TaxiIn","CRSArrTime","ArrTime","ArrDelay","ArrDelayMinutes","ArrDel15",
"ArrivalDelayGroups","ArrTimeBlk","Cancelled","CancellationCode","Diverted","CRSElapsedTime","ActualElapsedTime","AirTime",
"Flights","Distance","DistanceGroup","CarrierDelay","WeatherDelay","NASDelay","SecurityDelay","LateAircraftDelay",
"FirstDepTime","TotalAddGTime","LongestAddGTime","DivAirportLandings","DivReachedDest","DivActualElapsedTime","DivArrDelay",
"DivDistance","Div1Airport","Div1AirportID","Div1AirportSeqID","Div1WheelsOn","Div1TotalGTime","Div1LongestGTime",
"Div1WheelsOff","Div1TailNum","Div2Airport","Div2AirportID","Div2AirportSeqID","Div2WheelsOn","Div2TotalGTime",
"Div2LongestGTime","Div2WheelsOff","Div2TailNum","Div3Airport","Div3AirportID","Div3AirportSeqID","Div3WheelsOn",
"Div3TotalGTime","Div3LongestGTime","Div3WheelsOff","Div3TailNum","Div4Airport","Div4AirportID","Div4AirportSeqID",
"Div4WheelsOn","Div4TotalGTime","Div4LongestGTime","Div4WheelsOff","Div4TailNum","Div5Airport","Div5AirportID",
"Div5AirportSeqID","Div5WheelsOn","Div5TotalGTime","Div5LongestGTime","Div5WheelsOff","Div5TailNum"

openflights.org Database
Airports, Airlines, Routes
35

Scraping the FAA Registry
Airplane Data by Tail Number
36

Wikipedia Airlines Entries
Descriptions of Airlines
37

National Centers for Environmental Information
Historical Weather Observations
38

How to store it and how to fetch it…
Data Structure
and
Access Patterns

First Order Form
Storing groups of documents in key/value stores
40
Key/Value and some Document Stores
Partition Tolerance
No Single Master

First Order Form
Storing groups of documents in key/value stores
41

Second Order Form
Using column stores and compound keys to enable range scans
42
Range Scans for Fun and Profit
One or more Masters

Second Order Form
Using column stores and compound keys to enable range scans
43
# Achieve everything many access patterns via key composition:
Requirement: SELECT FLIGHTS WHERE FlightDate < X AND FlightDate> Y
Compose Key: TailNum + FlightDate, ex. N16954-2015-12-16
Range Scan: FROM: N16954-2015-12-01
TO: N16954-2015-01-01

Third Order Form
Using databases with B-Trees to query and analyze raw records
44
SELECT COUNT(*), SUM(*)
…
FROM …
GROUP BY …
WHERE …

Working our way up the data value pyramid
Climbing the Stack

Starting by “plumbing” the system from end to end
Plumbing

Publishing Flight Records
Plumbing our master records through to the web
47

Publishing Flight Records to MongoDB
48
import pymongo 
import pymongo_spark 
# Important: activate pymongo_spark. 
pymongo_spark.activate() 
# Load the parquet file
on_time_dataframe = spark.read.parquet('data/on_time_performance.parquet') 
# Convert to RDD of dicts and save to MongoDB 
as_dict = on_time_dataframe.rdd.map(lambda row: row.asDict())
as_dict.saveToMongoDB(‘mongodb://localhost:27017/agile_data_science.on_time_performance')

Publishing Flight Records to ElasticSearch
49
# Load the parquet file 
 
# Save the DataFrame to Elasticsearch 
on_time_dataframe.write.format("org.elasticsearch.spark.sql") 
.option("es.resource","agile_data_science/on_time_performance") 
.option("es.batch.size.entries","100") 
.mode("overwrite") 
.save()

Putting Records on the Web
50
from flask import Flask, render_template, request 
from pymongo import MongoClient 
from bson import json_util 
 
# Set up Flask and Mongo 
app = Flask(__name__) 
client = MongoClient() 
 
# Controller: Fetch an email and display it 
@app.route("/on_time_performance") 
def on_time_performance(): 
 
carrier = request.args.get('Carrier') 
flight_date = request.args.get('FlightDate') 
flight_num = request.args.get('FlightNum') 
 
flight = client.agile_data_science.on_time_performance.find_one({ 
'Carrier': carrier, 
'FlightDate': flight_date, 
'FlightNum': int(flight_num) 
}) 
 
return json_util.dumps(flight) 
 
if __name__ == "__main__": 
app.run(debug=True)

51

52

Getting to know your data
Tables and Charts

Tables in PySpark
Back end development in PySpark
54
# Load the parquet file 
 
# Use SQL to look at the total flights by month across 2015 
on_time_dataframe.registerTempTable("on_time_dataframe") 
total_flights_by_month = spark.sql( 
"""SELECT Month, Year, COUNT(*) AS total_flights 
FROM on_time_dataframe 
GROUP BY Year, Month 
ORDER BY Year, Month""" 
) 
 
# This map/asDict trick makes the rows print a little prettier. It is optional. 
flights_chart_data = total_flights_by_month.map(lambda row: row.asDict()) 
flights_chart_data.collect() 
 
# Save chart to MongoDB 
import pymongo_spark 
pymongo_spark.activate() 
flights_chart_data.saveToMongoDB( 
'mongodb://localhost:27017/agile_data_science.flights_by_month' 
)

Tables in Flask and Jinja2
Front end development in Flask: controller and template
55
# Controller: Fetch a flight table 
@app.route("/total_flights") 
def total_flights(): 
total_flights = client.agile_data_science.flights_by_month.find({},  
sort = [ 
('Year', 1), 
('Month', 1) 
]) 
return render_template('total_flights.html', total_flights=total_flights)
{% extends "layout.html" %} 
{% block body %} 
<div> 
<p class="lead">Total Flights by Month</p> 
<table class="table table-condensed table-striped" style="width: 200px;"> 
<thead> 
<th>Month</th> 
<th>Total Flights</th> 
</thead> 
<tbody> 
{% for month in total_flights %} 
<tr> 
<td>{{month.Month}}</td> 
<td>{{month.total_flights}}</td> 
</tr> 
{% endfor %} 
</tbody> 
</table> 
</div> 
{% endblock %}

Tables
Visualizing data
56

Charts in Flask and d3.js
Visualizing data with JSON and d3.js
57
# Serve the chart's data via an asynchronous request (formerly known as 'AJAX') 
@app.route("/total_flights.json") 
def total_flights_json(): 
sort = [ 
('Year', 1), 
('Month', 1) 
]) 
return json_util.dumps(total_flights, ensure_ascii=False)
var width = 960, 
height = 350; 
 
var y = d3.scale.linear() 
.range([height, 0]); 
// We define the domain once we get our data in d3.json, below 
 
var chart = d3.select(".chart") 
.attr("width", width) 
.attr("height", height); 
 
d3.json("/total_flights.json", function(data) { 
 
var defaultColor = 'steelblue'; 
var modeColor = '#4CA9F5'; 
 
var maxY = d3.max(data, function(d) { return d.total_flights; }); 
y.domain([0, maxY]); 
 
var varColor = function(d, i) { 
if(d['total_flights'] == maxY) { return modeColor; } 
else { return defaultColor; } 
} 
var barWidth = width / data.length; 
var bar = chart.selectAll("g") 
.data(data) 
.enter() 
.append("g") 
.attr("transform", function(d, i) { return "translate(" + i * barWidth + ",0)"; }); 
 
bar.append("rect") 
.attr("y", function(d) { return y(d.total_flights); }) 
.attr("height", function(d) { return height - y(d.total_flights); }) 
.attr("width", barWidth - 1) 
.style("fill", varColor); 
 
bar.append("text") 
.attr("x", barWidth / 2) 
.attr("y", function(d) { return y(d.total_flights) + 3; }) 
.attr("dy", ".75em") 
.text(function(d) { return d.total_flights; }); 
});

Charts
Visualizing data
58

59
sort = [ 
('Year', 1), 
('Month', 1) 
]) 
var width = 960,
height = 350;
var y = d3.scale.linear()
.range([height, 0]);
// We define the domain once we get our data in d3.json, below
var chart = d3.select(".chart")
.attr("width", width)
.attr("height", height);
d3.json("/total_flights.json", function(data) {
y.domain([0, d3.max(data, function(d) { return d.total_flights; })]);
var barWidth = width / data.length;
var bar = chart.selectAll("g")
.data(data)
.enter()
.append("g")
.attr("transform", function(d, i) {
return "translate(" + i * barWidth + ",0)"; });
bar.append("rect")
.attr("y", function(d) { return y(d.total_flights); })
.attr("height", function(d) { return height - y(d.total_flights); })
.attr("width", barWidth - 1);
bar.append("text")
.attr("x", barWidth / 2)
.attr("y", function(d) { return y(d.total_flights) + 3; })
.attr("dy", ".75em")
.text(function(d) { return d.total_flights; });
});

Charts
Visualizing data
60

61
sort = [ 
('Year', 1), 
('Month', 1) 
]) 
var width = 960, 
height = 350; 
 
var y = d3.scale.linear() 
.range([height, 0]); 
// We define the domain once we get our data in d3.json, below 
 
var chart = d3.select(".chart") 
.attr("width", width) 
.attr("height", height); 
 
d3.json("/total_flights.json", function(data) { 
 
var defaultColor = 'steelblue'; 
var modeColor = '#4CA9F5'; 
 
var maxY = d3.max(data, function(d) { return d.total_flights; }); 
y.domain([0, maxY]); 
 
var varColor = function(d, i) { 
if(d['total_flights'] == maxY) { return modeColor; } 
else { return defaultColor; } 
} 
var barWidth = width / data.length; 
var bar = chart.selectAll("g") 
.data(data) 
.enter() 
.append("g") 
.attr("transform", function(d, i) { return "translate(" + i * barWidth + ",0)"; }); 
 
bar.append("rect") 
.attr("y", function(d) { return y(d.total_flights); }) 
.attr("height", function(d) { return height - y(d.total_flights); }) 
.attr("width", barWidth - 1) 
.style("fill", varColor); 
 
bar.append("text") 
.attr("x", barWidth / 2) 
.attr("y", function(d) { return y(d.total_flights) + 3; }) 
.attr("dy", ".75em") 
.text(function(d) { return d.total_flights; }); 
});

Charts
Visualizing data
62

Exploring your data through interaction
Reports

Creating Interactive Ontologies from Semi-Structured Data
Extracting and visualizing entities
64

Home Page
Extracting and decorating entities
65

Airline Entity @ /airline/DL
Extracting and decorating entities
66

Summarizing all Airline Entities @ /airplanes
Describing entities in aggregate
67

Summarizing Airlines 2.0
68

69

Summarizing Airlines 3.0 - Entity Resolution Problem
70
...
AIRBUS
AIRBUS INDUSTRIE
...
EMBRAER
EMBRAER S A
...
GULFSTREAM AEROSPACE
GULFSTREAM AEROSPACE CORP
...
MCDONNELL DOUGLAS
MCDONNELL DOUGLAS AIRCRAFT CO
MCDONNELL DOUGLAS CORPORATION
...

Summarizing Airlines 3.0 - Entity Resolution Hack
71
# Detect the longest common beginning string in a pair of strings
def longest_common_beginning(s1, s2):
if s1 == s2:
return s1
min_length = min(len(s1), len(s2))
i = 0
while i < min_length:
if s1[i] == s2[i]:
i += 1
else:
break
return s1[0:i]
# Compare two manufacturers, returning a tuple describing the result
def compare_manufacturers(mfrs):
mfr1 = mfrs[0]
mfr2 = mfrs[1]
lcb = longest_common_beginning(mfr1, mfr2)
len_lcb = len(lcb)
record = {
'mfr1': mfr1,
'mfr2': mfr2,
'lcb': lcb,
'len_lcb': len_lcb,
'eq': mfr1 == mfr2
}
return record
# Pair every unique instance of Manufacturer field with every
# other for comparison
comparison_pairs = manufacturer_variety.join(manufacturer_variety)
# Do the comparisons
comparisons = comparison_pairs.rdd.map(compare_manufacturers)
# Matches have > 5 starting chars in common
matches = comparisons.filter(lambda f: f['eq'] == False and f['len_lcb'] > 5)
#
# Now we create a mapping of duplicate keys from their raw value
# to the one we're going to use
#
# 1) Group the matches by the longest common beginning ('lcb')
common_lcbs = matches.groupBy(lambda x: x['lcb'])
# 2) Emit the raw value for each side of the match along with the key, our 'lcb'
mfr1_map = common_lcbs.map(
lambda x: [(y['mfr1'], x[0]) for y in x[1]]).flatMap(lambda x: x)
mfr2_map = common_lcbs.map(
lambda x: [(y['mfr2'], x[0]) for y in x[1]]).flatMap(lambda x: x)
# 3) Combine the two sides of the comparison's records
map_with_dupes = mfr1_map.union(mfr2_map)
# 4) Remove duplicates
mfr_dedupe_mapping = map_with_dupes.distinct()
# 5) Convert mapping to dataframe to join to airplanes dataframe
mapping_dataframe = mfr_dedupe_mapping.toDF()
# 6) Give the mapping column names
mapping_dataframe.registerTempTable("mapping_dataframe")
mapping_dataframe = spark.sql(
"SELECT _1 AS Raw, _2 AS NewManufacturer FROM mapping_dataframe"
)

72

Searching Flights
Using ElasticSearch to search for flights
73

Predicting the future for fun and profit
Predictions

Back End Design
Deep Storage and Spark vs Kafka and Spark Streaming
75
/
Batch Realtime
Historical Data
Train Model Apply Model
Realtime Data

Data Syndrome: Agile Data Science 2.0 76
jQuery in the web client submits a form to create the prediction request, and
then polls another url every few seconds until the prediction is ready. The
request generates a Kafka event, which a Spark Streaming worker processes
by applying the model we trained in batch. Having done so, it inserts a record
for the prediction in MongoDB, where the Flask app sends it to the web client
the next time it polls the server
Front End Design
/flights/delays/predict/classify_realtime/

User Interface
Where the user submits prediction requests
77

String Vectorization
From properties of items to vector format
78

scikit-learn was 166. Spark MLlib is very powerful!
http://bit.ly/train_model_spark
190 Line Model
# !/usr/bin/env python 
 
import sys, os, re 
 
# Pass date and base path to main() from airflow 
def main(base_path): 
 
# Default to "." 
try: base_path 
except NameError: base_path = "." 
if not base_path: 
base_path = "." 
 
APP_NAME = "train_spark_mllib_model.py" 
 
# If there is no SparkSession, create the environment 
try: 
sc and spark 
except NameError as e: 
import findspark 
findspark.init() 
import pyspark 
import pyspark.sql 
 
sc = pyspark.SparkContext() 
spark = pyspark.sql.SparkSession(sc).builder.appName(APP_NAME).getOrCreate() 
 
# 
# { 
# "ArrDelay":5.0,"CRSArrTime":"2015-12-31T03:20:00.000-08:00","CRSDepTime":"2015-12-31T03:05:00.000-08:00", 
# "Carrier":"WN","DayOfMonth":31,"DayOfWeek":4,"DayOfYear":365,"DepDelay":14.0,"Dest":"SAN","Distance":368.0, 
# "FlightDate":"2015-12-30T16:00:00.000-08:00","FlightNum":"6109","Origin":"TUS" 
# } 
# 
from pyspark.sql.types import StringType, IntegerType, FloatType, DoubleType, DateType, TimestampType 
from pyspark.sql.types import StructType, StructField 
from pyspark.sql.functions import udf 
 
schema = StructType([ 
StructField("ArrDelay", DoubleType(), True), # "ArrDelay":5.0 
StructField("CRSArrTime", TimestampType(), True), # "CRSArrTime":"2015-12-31T03:20:00.000-08:00" 
StructField("CRSDepTime", TimestampType(), True), # "CRSDepTime":"2015-12-31T03:05:00.000-08:00" 
StructField("Carrier", StringType(), True), # "Carrier":"WN" 
StructField("DayOfMonth", IntegerType(), True), # "DayOfMonth":31 
StructField("DayOfWeek", IntegerType(), True), # "DayOfWeek":4 
StructField("DayOfYear", IntegerType(), True), # "DayOfYear":365 
StructField("DepDelay", DoubleType(), True), # "DepDelay":14.0 
StructField("Dest", StringType(), True), # "Dest":"SAN" 
StructField("Distance", DoubleType(), True), # "Distance":368.0 
StructField("FlightDate", DateType(), True), # "FlightDate":"2015-12-30T16:00:00.000-08:00" 
StructField("FlightNum", StringType(), True), # "FlightNum":"6109" 
StructField("Origin", StringType(), True), # "Origin":"TUS" 
]) 
 
input_path = "{}/data/simple_flight_delay_features.jsonl.bz2".format( 
base_path 
) 
features = spark.read.json(input_path, schema=schema) 
features.first() 
 
# 
# Check for nulls in features before using Spark ML 
# 
null_counts = [(column, features.where(features[column].isNull()).count()) for column in features.columns] 
cols_with_nulls = filter(lambda x: x[1] > 0, null_counts) 
print(list(cols_with_nulls)) 
 
# 
# Add a Route variable to replace FlightNum 
# 
from pyspark.sql.functions import lit, concat 
features_with_route = features.withColumn( 
'Route', 
concat( 
features.Origin, 
lit('-'), 
features.Dest 
) 
) 
features_with_route.show(6) 
 
# 
# Use pysmark.ml.feature.Bucketizer to bucketize ArrDelay into on-time, slightly late, very late (0, 1, 2) 
# 
from pyspark.ml.feature import Bucketizer 
 
# Setup the Bucketizer 
splits = [-float("inf"), -15.0, 0, 30.0, float("inf")] 
arrival_bucketizer = Bucketizer( 
splits=splits, 
inputCol="ArrDelay", 
outputCol="ArrDelayBucket" 
) 
 
# Save the bucketizer 
arrival_bucketizer_path = "{}/models/arrival_bucketizer_2.0.bin".format(base_path) 
arrival_bucketizer.write().overwrite().save(arrival_bucketizer_path) 
 
# Apply the bucketizer 
ml_bucketized_features = arrival_bucketizer.transform(features_with_route) 
ml_bucketized_features.select("ArrDelay", "ArrDelayBucket").show() 
 
# 
# Extract features tools in with pyspark.ml.feature 
# 
from pyspark.ml.feature import StringIndexer, VectorAssembler 
 
# Turn category fields into indexes 
for column in ["Carrier", "Origin", "Dest", "Route"]: 
string_indexer = StringIndexer( 
inputCol=column, 
outputCol=column + "_index" 
) 
 
string_indexer_model = string_indexer.fit(ml_bucketized_features) 
ml_bucketized_features = string_indexer_model.transform(ml_bucketized_features) 
 
# Drop the original column 
ml_bucketized_features = ml_bucketized_features.drop(column) 
 
# Save the pipeline model 
string_indexer_output_path = "{}/models/string_indexer_model_{}.bin".format( 
base_path, 
column 
) 
string_indexer_model.write().overwrite().save(string_indexer_output_path) 
 
# Combine continuous, numeric fields with indexes of nominal ones 
# ...into one feature vector 
numeric_columns = [ 
"DepDelay", "Distance", 
"DayOfMonth", "DayOfWeek", 
"DayOfYear"] 
index_columns = ["Carrier_index", "Origin_index", 
"Dest_index", "Route_index"] 
vector_assembler = VectorAssembler( 
inputCols=numeric_columns + index_columns, 
outputCol="Features_vec" 
) 
final_vectorized_features = vector_assembler.transform(ml_bucketized_features) 
 
# Save the numeric vector assembler 
vector_assembler_path = "{}/models/numeric_vector_assembler.bin".format(base_path) 
vector_assembler.write().overwrite().save(vector_assembler_path) 
 
# Drop the index columns 
for column in index_columns: 
final_vectorized_features = final_vectorized_features.drop(column) 
 
# Inspect the finalized features 
final_vectorized_features.show() 
 
# Instantiate and fit random forest classifier on all the data 
from pyspark.ml.classification import RandomForestClassifier 
rfc = RandomForestClassifier( 
featuresCol="Features_vec", 
labelCol="ArrDelayBucket", 
predictionCol="Prediction", 
maxBins=4657, 
maxMemoryInMB=1024 
) 
model = rfc.fit(final_vectorized_features) 
 
# Save the new model over the old one 
model_output_path = "{}/models/spark_random_forest_classifier.flight_delays.5.0.bin".format( 
base_path 
) 
model.write().overwrite().save(model_output_path) 
 
# Evaluate model using test data 
predictions = model.transform(final_vectorized_features) 
 
from pyspark.ml.evaluation import MulticlassClassificationEvaluator 
evaluator = MulticlassClassificationEvaluator( 
metricName="accuracy" 
) 
accuracy = evaluator.evaluate(predictions) 
print("Accuracy = {}".format(accuracy)) 
 
# Check the distribution of predictions 
predictions.groupBy("Prediction").count().show() 
 
# Check a sample 
predictions.sample(False, 0.001, 18).orderBy("CRSDepTime").show(6) 
 
if __name__ == "__main__": 
main(sys.argv[1])

Initializing the Environment
Setting up the environment…
80
 
 
 
# Default to "." 
try: base_path 
except NameError: base_path = "." 
if not base_path: 
base_path = "." 
 
 
try: 
sc and spark 
import findspark 
findspark.init() 
import pyspark 
 
spark = pyspark.sql.SparkSession(sc).builder.appName(APP_NAME).getOrCreate()

Loading the Training Data
Using DataFrames to load structured data…
81
 
]) 
 
input_path = "{}/data/simple_flight_delay_features.jsonl.bz2".format( 
base_path 
) 
features.first()

Checking for Nulls
Checking the data for null values that would crash Spark MLlib
82
# 
# Check for nulls in features before using Spark ML 
# 
null_counts = [(column, features.where(features[column].isNull()).count()) for column in features.columns] 
cols_with_nulls = filter(lambda x: x[1] > 0, null_counts) 
print(list(cols_with_nulls))

Adding a Feature
Using DataFrame.withColumn to add a Route feature to the data…
83
# 
# 
'Route', 
concat( 
features.Origin, 
lit('-'), 
features.Dest 
) 
) 
features_with_route.show(6)

Bucketizing the Prediction Column
Using Bucketizer to convert a continuous variable to a nominal one…
84
# 
# 
 
splits=splits, 
) 
 
# Save the bucketizer 
 
# Apply the bucketizer 
ml_bucketized_features = arrival_bucketizer.transform(features_with_route) 
ml_bucketized_features.select("ArrDelay", "ArrDelayBucket").show()

StringIndexing the String Columns
Using StringIndexer to convert nominal fields to numeric ones…
85
# 
# 
 
inputCol=column, 
) 
 
 
# Drop the original column 
ml_bucketized_features = ml_bucketized_features.drop(column) 
 
string_indexer_output_path = "{}/models/string_indexer_model_{}.bin".format( 
base_path, 
column 
) 
string_indexer_model.write().overwrite().save(string_indexer_output_path)

Vectorizing the Numeric Columns
Combining the numeric fields with VectorAssembler…
86
"DayOfYear"] 
) 
 
 
 
final_vectorized_features.show()

Training the Classifier Model
Creating and training a RandomForestClassifier model
87
maxBins=4657, 
maxMemoryInMB=1024 
) 
model = rfc.fit(final_vectorized_features) 
 
model_output_path = "{}/models/spark_random_forest_classifier.flight_delays.5.0.bin".format( 
base_path 
) 
model.write().overwrite().save(model_output_path)

Evaluating the Classifier Model
Using MultiClassificationEvaluator to check the accuracy of the model…
88
predictions = model.transform(final_vectorized_features) 
 
metricName="accuracy" 
) 
accuracy = evaluator.evaluate(predictions) 
print("Accuracy = {}".format(accuracy)) 
 
# Check the distribution of predictions 
predictions.groupBy("Prediction").count().show() 
 
# Check a sample 
predictions.sample(False, 0.001, 18).orderBy("CRSDepTime").show(6)

Running Main
Just what it looks like…
89
if __name__ == "__main__": 
main(sys.argv[1])

Using the model in realtime via Spark Streaming!
Deploying the Model

Deployment of Model Training to Airflow
Using Airflow to batch schedule the training of our model
91
training_dag = DAG(
'agile_data_science_batch_prediction_model_training',
default_args=default_args
)
# We use the same two commands for all our PySpark tasks
pyspark_bash_command = """
spark-submit --master {{ params.master }}
{{ params.base_path }}/{{ params.filename }}
{{ params.base_path }}
"""
pyspark_date_bash_command = """
spark-submit --master {{ params.master }}
{{ params.base_path }}/{{ params.filename }}
{{ ds }} {{ params.base_path }}
"""
# Gather the training data for our classifier
extract_features_operator = BashOperator(
task_id = "pyspark_extract_features",
bash_command = pyspark_bash_command,
params = {
"master": "local[8]",
"filename": "ch08/extract_features.py",
"base_path": "{}/".format(PROJECT_HOME)
},
dag=training_dag
)
# Train and persist the classifier model
train_classifier_model_operator = BashOperator(
task_id = "pyspark_train_classifier_model",
bash_command = pyspark_bash_command,
params = {
"master": "local[8]",
"filename": "ch08/train_spark_mllib_model.py",
"base_path": "{}/".format(PROJECT_HOME)
},
dag=training_dag
)
# The model training depends on the feature extraction
train_classifier_model_operator.set_upstream(extract_features_operator)

Deployment of Model Training to Airflow
Using Airflow to batch schedule the training of our model
92

Loading the Models
Loading the models we trained in batch to reproduce the data pipeline
93
# ch08/make_predictions_streaming.py
# Load the arrival delay bucketizer 
arrival_bucketizer = Bucketizer.load(arrival_bucketizer_path) 
 
# Load all the string field vectorizer pipelines into a dict 
from pyspark.ml.feature import StringIndexerModel 
 
string_indexer_models = {} 
for column in ["Carrier", "DayOfMonth", "DayOfWeek", "DayOfYear", 
"Origin", "Dest", "Route"]: 
string_indexer_model_path = "{}/models/string_indexer_model_{}.bin".format( 
base_path, 
column 
) 
string_indexer_model = StringIndexerModel.load(string_indexer_model_path) 
string_indexer_models[column] = string_indexer_model

Loading the Models
Loading the models we trained in batch to reproduce the data pipeline
94
# ch08/make_predictions_streaming.py
# Load the numeric vector assembler 
from pyspark.ml.feature import VectorAssembler 
vector_assembler = VectorAssembler.load(vector_assembler_path) 
 
# Load the classifier model 
from pyspark.ml.classification import RandomForestClassifier,
RandomForestClassificationModel 
random_forest_model_path = "{}/models/spark_random_forest_classifier.flight_delays.
5.0.bin".format( 
base_path 
) 
rfc = RandomForestClassificationModel.load( 
random_forest_model_path

Connecting to Kafka
Creating a direct stream to the Kafka queue containing our prediction requests
95
# 
# Process Prediction Requests in Streaming 
#
from pyspark.streaming.kafka import KafkaUtils 
 
stream = KafkaUtils.createDirectStream( 
ssc, 
[PREDICTION_TOPIC], 
{ 
"metadata.broker.list": BROKERS, 
"group.id": "0", 
} 
) 
 
object_stream = stream.map(lambda x: json.loads(x[1])) 
object_stream.pprint()

Repeating the Pipeline
Running the prediction requests through the same data flow as the training data
96
row_stream = object_stream.map( 
lambda x: Row( 
FlightDate=iso8601.parse_date(x['FlightDate']), 
Origin=x['Origin'], 
Distance=x['Distance'], 
DayOfMonth=x['DayOfMonth'], 
DayOfYear=x['DayOfYear'], 
UUID=x['UUID'], 
DepDelay=x['DepDelay'], 
DayOfWeek=x['DayOfWeek'], 
FlightNum=x['FlightNum'], 
Dest=x['Dest'], 
Timestamp=iso8601.parse_date(x['Timestamp']), 
Carrier=x['Carrier'] 
) 
) 
row_stream.pprint()
# Do the classification and store to Mongo 
row_stream.foreachRDD(classify_prediction_requests) 
 
ssc.start() 
ssc.awaitTermination()

97
def classify_prediction_requests(rdd): 
 
from pyspark.sql.types import StringType, IntegerType, DoubleType, DateType, TimestampType 
 
prediction_request_schema = StructType([ 
StructField("Carrier", StringType(), True), 
StructField("DayOfMonth", IntegerType(), True), 
StructField("DayOfWeek", IntegerType(), True), 
StructField("DayOfYear", IntegerType(), True), 
StructField("DepDelay", DoubleType(), True), 
StructField("Dest", StringType(), True), 
StructField("Distance", DoubleType(), True), 
StructField("FlightDate", DateType(), True), 
StructField("FlightNum", StringType(), True), 
StructField("Origin", StringType(), True), 
StructField("Timestamp", TimestampType(), True), 
StructField("UUID", StringType(), True), 
]) 
 
prediction_requests_df = spark.createDataFrame(rdd, schema=prediction_request_schema) 
prediction_requests_df.show()
prediction_requests_with_route = prediction_requests_df.withColumn( 
'Route', 
concat( 
prediction_requests_df.Origin, 
lit('-'), 
prediction_requests_df.Dest 
) 
) 
prediction_requests_with_route.show(6)
...

98
for column in ["Carrier", "DayOfMonth", "DayOfWeek", "DayOfYear", 
"Origin", "Dest", "Route"]: 
string_indexer_model = string_indexer_models[column] 
prediction_requests_with_route = string_indexer_model.transform(prediction_requests_with_route) 
 
# Vectorize numeric columns: DepDelay, Distance and index columns 
final_vectorized_features = vector_assembler.transform(prediction_requests_with_route) 
 
# Inspect the vectors 
 
# Drop the individual index columns 
index_columns = ["Carrier_index", "DayOfMonth_index", "DayOfWeek_index", "DayOfYear_index", 
"Origin_index", "Dest_index", "Route_index"] 
 
 
# Make the prediction 
predictions = rfc.transform(final_vectorized_features)
# Drop the features vector and prediction metadata to give the original fields 
predictions = predictions.drop("Features_vec") 
final_predictions = predictions.drop("indices").drop("values").drop("rawPrediction").drop("probability") 
 
# Inspect the output 
final_predictions.show()

Storing to Mongo
Putting the result where our web application can access it
99
# Store to Mongo 
if final_predictions.count() > 0: 
final_predictions.rdd.map(lambda x: x.asDict()).saveToMongoDB( 
"mongodb://localhost:27017/agile_data_science.flight_delay_classification_response" 
)

Final Result of Deployed Prediction
The bang for the buck!
100

Experimental setup for iteratively improving the predictive model
Improving the Model

Experiment Setup
Necessary to improve model
102

155 additional lines to setup an experiment
and add 3 new features to improvement the model
http://bit.ly/improved_model_spark
345 L.O.C.
 
import json 
import datetime, iso8601 
from tabulate import tabulate 
 
 
try: 
sc and spark 
import findspark 
findspark.init() 
import pyspark 
 
spark = pyspark.sql.SparkSession(sc).builder.appName(APP_NAME).getOrCreate() 
 
# 
# { 
# "ArrDelay":5.0,"CRSArrTime":"2015-12-31T03:20:00.000-08:00","CRSDepTime":"2015-12-31T03:05:00.000-08:00", 
# "Carrier":"WN","DayOfMonth":31,"DayOfWeek":4,"DayOfYear":365,"DepDelay":14.0,"Dest":"SAN","Distance":368.0, 
# "FlightDate":"2015-12-30T16:00:00.000-08:00","FlightNum":"6109","Origin":"TUS" 
# } 
# 
 
]) 
 
input_path = "{}/data/simple_flight_delay_features.json".format( 
base_path 
) 
features.first() 
 
# 
# 
'Route', 
concat( 
features.Origin, 
lit('-'), 
features.Dest 
) 
) 
features_with_route.show(6) 
 
# 
# Add the hour of day of scheduled arrival/departure 
# 
from pyspark.sql.functions import hour 
features_with_hour = features_with_route.withColumn( 
"CRSDepHourOfDay", 
hour(features.CRSDepTime) 
) 
features_with_hour = features_with_hour.withColumn( 
"CRSArrHourOfDay", 
hour(features.CRSArrTime) 
) 
features_with_hour.select("CRSDepTime", "CRSDepHourOfDay", "CRSArrTime", "CRSArrHourOfDay").show() 
 
# 
# 
 
splits=splits, 
) 
 
# Save the model 
 
# Apply the model 
ml_bucketized_features = arrival_bucketizer.transform(features_with_hour) 
ml_bucketized_features.select("ArrDelay", "ArrDelayBucket").show() 
 
# 
# 
 
inputCol=column, 
) 
 
string_indexer_output_path = "{}/models/string_indexer_model_3.0.{}.bin".format( 
base_path, 
column 
) 
string_indexer_model.write().overwrite().save(string_indexer_output_path) 
 
"DayOfYear", "CRSDepHourOfDay", 
"CRSArrHourOfDay"] 
) 
 
vector_assembler_path = "{}/models/numeric_vector_assembler_3.0.bin".format(base_path) 
 
 
 
# 
# Cross validate, train and evaluate classifier: loop 5 times for 4 metrics 
# 
 
from collections import defaultdict 
scores = defaultdict(list) 
feature_importances = defaultdict(list) 
metric_names = ["accuracy", "weightedPrecision", "weightedRecall", "f1"] 
split_count = 3 
 
for i in range(1, split_count + 1): 
print("nRun {} out of {} of test/train splits in cross validation...".format( 
i, 
split_count, 
) 
) 
 
# Test/train split 
training_data, test_data = final_vectorized_features.randomSplit([0.8, 0.2]) 
 
maxBins=4657, 
) 
model = rfc.fit(training_data) 
 
model_output_path = "{}/models/spark_random_forest_classifier.flight_delays.baseline.bin".format( 
base_path 
) 
model.write().overwrite().save(model_output_path) 
 
predictions = model.transform(test_data) 
 
# Evaluate this split's results for each metric 
for metric_name in metric_names: 
metricName=metric_name 
) 
score = evaluator.evaluate(predictions) 
 
scores[metric_name].append(score) 
print("{} = {}".format(metric_name, score)) 
 
# 
# Collect feature importances 
# 
feature_names = vector_assembler.getInputCols() 
feature_importance_list = model.featureImportances 
for feature_name, feature_importance in zip(feature_names, feature_importance_list): 
feature_importances[feature_name].append(feature_importance) 
 
# 
# Evaluate average and STD of each metric and print a table 
# 
import numpy as np 
score_averages = defaultdict(float) 
 
# Compute the table data 
average_stds = [] # ha 
metric_scores = scores[metric_name] 
 
average_accuracy = sum(metric_scores) / len(metric_scores) 
score_averages[metric_name] = average_accuracy 
 
std_accuracy = np.std(metric_scores) 
 
average_stds.append((metric_name, average_accuracy, std_accuracy)) 
 
# Print the table 
print("nExperiment Log") 
print("--------------") 
print(tabulate(average_stds, headers=["Metric", "Average", "STD"])) 
 
# 
# Persist the score to a sccore log that exists between runs 
# 
import pickle
# Load the score log or initialize an empty one 
try: 
score_log_filename = "{}/models/score_log.pickle".format(base_path) 
score_log = pickle.load(open(score_log_filename, "rb")) 
if not isinstance(score_log, list): 
score_log = [] 
except IOError: 
score_log = [] 
 
# Compute the existing score log entry 
score_log_entry = {metric_name: score_averages[metric_name] for metric_name in metric_names} 
 
# Compute and display the change in score for each metric 
try: 
last_log = score_log[-1] 
except (IndexError, TypeError, AttributeError): 
last_log = score_log_entry 
 
experiment_report = [] 
run_delta = score_log_entry[metric_name] - last_log[metric_name] 
experiment_report.append((metric_name, run_delta)) 
 
print("nExperiment Report") 
print("-----------------") 
print(tabulate(experiment_report, headers=["Metric", "Score"])) 
 
# Append the existing average scores to the log 
score_log.append(score_log_entry) 
 
# Persist the log for next run 
pickle.dump(score_log, open(score_log_filename, "wb")) 
 
# 
# Analyze and report feature importance changes 
# 
 
# Compute averages for each feature 
feature_importance_entry = defaultdict(float) 
for feature_name, value_list in feature_importances.items(): 
average_importance = sum(value_list) / len(value_list) 
feature_importance_entry[feature_name] = average_importance 
 
# Sort the feature importances in descending order and print 
import operator 
sorted_feature_importances = sorted( 
feature_importance_entry.items(), 
key=operator.itemgetter(1), 
reverse=True 
) 
 
print("nFeature Importances") 
print("-------------------") 
print(tabulate(sorted_feature_importances, headers=['Name', 'Importance'])) 
 
# 
# Compare this run's feature importances with the previous run's 
# 
 
# Load the feature importance log or initialize an empty one 
try: 
feature_log_filename = "{}/models/feature_log.pickle".format(base_path) 
feature_log = pickle.load(open(feature_log_filename, "rb")) 
if not isinstance(feature_log, list): 
feature_log = [] 
except IOError: 
feature_log = [] 
 
# Compute and display the change in score for each feature 
try: 
last_feature_log = feature_log[-1] 
except (IndexError, TypeError, AttributeError): 
last_feature_log = defaultdict(float) 
for feature_name, importance in feature_importance_entry.items(): 
last_feature_log[feature_name] = importance 
 
# Compute the deltas 
feature_deltas = {} 
for feature_name in feature_importances.keys(): 
run_delta = feature_importance_entry[feature_name] - last_feature_log[feature_name] 
feature_deltas[feature_name] = run_delta 
 
# Sort feature deltas, biggest change first 
import operator 
sorted_feature_deltas = sorted( 
feature_deltas.items(), 
key=operator.itemgetter(1), 
reverse=True 
) 
 
# Display sorted feature deltas 
print("nFeature Importance Delta Report") 
print("-------------------------------") 
print(tabulate(sorted_feature_deltas, headers=["Feature", "Delta"])) 
 
# Append the existing average deltas to the log 
feature_log.append(feature_importance_entry) 
 
# Persist the log for next run 
pickle.dump(feature_log, open(feature_log_filename, "wb")) 
 
if __name__ == "__main__": 
main(sys.argv[1])

Creating an Experiment
Cross validate, train and evaluate classifier
104
from collections import defaultdict
scores = defaultdict(list)
feature_importances = defaultdict(list)
metric_names = ["accuracy", "weightedPrecision", "weightedRecall", "f1"]
split_count = 3
for i in range(1, split_count + 1):
print("nRun {} out of {} of test/train splits in cross validation...".format(
i,
split_count,
)
)
# Test/train split
training_data, test_data = final_vectorized_features.randomSplit([0.8, 0.2])
# Instantiate and fit random forest classifier on all the data
from pyspark.ml.classification import RandomForestClassifier
rfc = RandomForestClassifier(
featuresCol="Features_vec",
labelCol="ArrDelayBucket",
predictionCol="Prediction",
maxBins=4657,
)
model = rfc.fit(training_data)

105
# Save the new model over the old one
model_output_path = "{}/models/spark_random_forest_classifier.flight_delays.baseline.bin".format(
base_path
)
model.write().overwrite().save(model_output_path)
# Evaluate model using test data
predictions = model.transform(test_data)
# Evaluate this split's results for each metric
from pyspark.ml.evaluation import MulticlassClassificationEvaluator
for metric_name in metric_names:
evaluator = MulticlassClassificationEvaluator(
labelCol="ArrDelayBucket",
predictionCol="Prediction",
metricName=metric_name
)
score = evaluator.evaluate(predictions)
scores[metric_name].append(score)
print("{} = {}".format(metric_name, score))

106
#
# Collect feature importances
#
feature_names = vector_assembler.getInputCols()
feature_importance_list = model.featureImportances
for feature_name, feature_importance in zip(feature_names, feature_importance_list):
feature_importances[feature_name].append(feature_importance)
#
# Evaluate average and STD of each metric and print a table
#
import numpy as np
score_averages = defaultdict(float)
# Compute the table data
average_stds = [] # ha
metric_scores = scores[metric_name]
average_accuracy = sum(metric_scores) / len(metric_scores)
score_averages[metric_name] = average_accuracy
std_accuracy = np.std(metric_scores)
average_stds.append((metric_name, average_accuracy, std_accuracy))

107
# Print the table
print("nExperiment Log")
print("--------------")
print(tabulate(average_stds, headers=["Metric", "Average", "STD"]))
#
# Persist the score to a sccore log that exists between runs
#
import pickle
# Load the score log or initialize an empty one
try:
score_log_filename = "{}/models/score_log.pickle".format(base_path)
score_log = pickle.load(open(score_log_filename, "rb"))
if not isinstance(score_log, list):
score_log = []
except IOError:
score_log = []
# Compute the existing score log entry
score_log_entry = {metric_name: score_averages[metric_name] for metric_name in metric_names}

108
# Compute and display the change in score for each metric
try:
last_log = score_log[-1]
except (IndexError, TypeError, AttributeError):
last_log = score_log_entry
experiment_report = []
run_delta = score_log_entry[metric_name] - last_log[metric_name]
experiment_report.append((metric_name, run_delta))
print("nExperiment Report")
print("-----------------")
print(tabulate(experiment_report, headers=["Metric", "Score"]))
# Append the existing average scores to the log
score_log.append(score_log_entry)
# Persist the log for next run
pickle.dump(score_log, open(score_log_filename, "wb"))

109
#
# Analyze and report feature importance changes
#
# Compute averages for each feature
feature_importance_entry = defaultdict(float)
for feature_name, value_list in feature_importances.items():
average_importance = sum(value_list) / len(value_list)
feature_importance_entry[feature_name] = average_importance
# Sort the feature importances in descending order and print
import operator
sorted_feature_importances = sorted(
feature_importance_entry.items(),
key=operator.itemgetter(1),
reverse=True
)
print("nFeature Importances")
print("-------------------")
print(tabulate(sorted_feature_importances, headers=['Name', 'Importance']))

110
#
# Compare this run's feature importances with the previous run's
#
# Load the feature importance log or initialize an empty one
try:
feature_log_filename = "{}/models/feature_log.pickle".format(base_path)
feature_log = pickle.load(open(feature_log_filename, "rb"))
if not isinstance(feature_log, list):
feature_log = []
except IOError:
feature_log = []
# Compute and display the change in score for each feature
try:
last_feature_log = feature_log[-1]
except (IndexError, TypeError, AttributeError):
last_feature_log = defaultdict(float)
for feature_name, importance in feature_importance_entry.items():
last_feature_log[feature_name] = importance

111
# Compute the deltas
feature_deltas = {}
for feature_name in feature_importances.keys():
run_delta = feature_importance_entry[feature_name] - last_feature_log[feature_name]
feature_deltas[feature_name] = run_delta
# Sort feature deltas, biggest change first
import operator
sorted_feature_deltas = sorted(
feature_deltas.items(),
key=operator.itemgetter(1),
reverse=True
)
# Display sorted feature deltas
print("nFeature Importance Delta Report")
print("-------------------------------")
print(tabulate(sorted_feature_deltas, headers=["Feature", "Delta"]))
# Append the existing average deltas to the log
feature_log.append(feature_importance_entry)
# Persist the log for next run
pickle.dump(feature_log, open(feature_log_filename, "wb"))

Running an Experiment
112
Experiment Log
--------------
Metric Average STD
----------------- --------- -----------
accuracy 0.594443 0.000382382
weightedPrecision 0.642419 0.00352101
weightedRecall 0.594443 0.000382382
f1 0.522397 0.000438121

Comparing Experiments
113
Experiment Report
-----------------
Metric Score
----------------- -----------
accuracy 0.00300548
weightedPrecision -0.00592227
weightedRecall 0.00300548
f1 -0.0105553

Feature Importances
114
Feature Importances
-------------------
Name Importance
--------------- ------------
DepDelay 0.882216
Route_index 0.0571401
Origin_index 0.0142741
Distance 0.0134583
Dest_index 0.00745796
DayOfYear 0.00544761
Carrier_index 0.00454088
DayOfMonth 9.31109e-05
DayOfWeek 5.2597e-05

Comparing Feature Importances
115
Feature Importance Delta Report
-------------------------------
Feature Delta
--------------- ------------
CRSArrHourOfDay 0.00982592
DepDelay 0.00671702
CRSDepHourOfDay 0.0012717
DayOfWeek -2.72792e-05
DayOfMonth -8.81635e-05
Origin_index -0.00152788
Distance -0.00188322
Route_index -0.00214609
Dest_index -0.0032327
DayOfYear -0.00377184
Carrier_index -0.00513747

Next steps for learning more about Agile Data Science 2.0
Next Steps

Building Full-Stack Data Analytics Applications with Spark
http://bit.ly/agile_data_science
Available Now on O’Reilly Safari: http://bit.ly/agile_data_safari

Realtime Predictive
Analytics
Rapidly learn to build entire predictive systems driven by
Kafka, PySpark, Speak Streaming, Spark MLlib and with a web
front-end using Python/Flask and JQuery.
Available for purchase at http://datasyndrome.com/video

Data Syndrome Russell Jurney
Principal Consultant
Email : rjurney@datasyndrome.com
Web : datasyndrome.com
Data Syndrome, LLC
Product Consulting
We build analytics products
and systems consisting of
big data viz, predictions,
recommendations, reports
and search.
Corporate Training
We offer training courses
for data scientists and
engineers and data
science teams,
Video Training
We offer video training
courses that rapidly
acclimate you with a
technology and technique.

Agile Data Science 2.0

More Related Content

What's hot (18)

Similar to Agile Data Science 2.0 (20)

Recently uploaded (20)

Agile Data Science 2.0