Data Workflows for Machine Learning - SF Bay Area ML

Data Workﬂows for
Machine Learning:

 
SF Bay Area ML Meetup

2014-04-09

!
Paco Nathan 
@pacoid

http://liber118.com/pxn/
meetup.com/SF-Bayarea-Machine-Learning/events/173759442/

BTW, since it’s April 9th…

Happy #IoTDay !!

http://iotday.org/

!
(any idea of what this means?)

Why is this talk here?

Machine Learning in production apps is less and less about
algorithms (even though that work is quite fun and vital)

!
Performing real work is more about:

! socializing a problem within an organization

! feature engineering (“Beyond Product Managers”)

! tournaments in CI/CD environments

! operationalizing high-ROI apps at scale

! etc.

!
So I’ll just crawl out on a limb and state that leveraging great
frameworks to build data workﬂows is more important than  
chasing after diminishing returns on highly nuanced algorithms.

!
Because Interwebs!

Data Workflows for Machine Learning

Middleware has been evolving for Big Data, and there are some
great examples — we’ll review several. Process has been evolving
too, right along with the use cases.

!
Popular frameworks typically provide some Machine Learning
capabilities within their core components, or at least among their
major use cases.

!
Let’s consider features from Enterprise DataWorkflows as a basis  
for what’s needed in Data Workflows for Machine Learning.

!
Their requirements for scale, robustness, cost trade-offs,
interdisciplinary teams, etc., serve as guides in general.

Caveat Auditor

I won’t claim to be expert with each of the frameworks and
environments described in this talk. Expert with a few of them
perhaps, but more to the point: embroiled in many use cases.

!
This talk attempts to deﬁne a “scorecard” for evaluating
important ML data workﬂow features:  
what’s needed for use cases, compare and contrast of what’s
available, plus some indication of which frameworks are likely  
to be best for a given scenario.

!
Seriously, this is a work in progress.

Outline

• Definition: Machine Learning

• Definition: Data Workflows

• A whole bunch o’ examples across several platforms

• Nine points to discuss, leading up to a scorecard

• A crazy little thing called PMML

• Questions, comments, flying tomatoes…

Data Workﬂows for
Machine Learning:

 
Frame the question…

!
“A Basis for What’s Needed”

“Machine learning algorithms can figure out how to perform 
important tasks by generalizing from examples.This is often 
feasible and cost-effective where manual programming is not.  
As more data becomes available, more ambitious problems  
can be tackled. As a result, machine learning is widely used  
in computer science and other fields.”

Pedro Domingos, U Washington 
A Few Useful Things to Know about Machine Learning
Definition: Machine Learning
• overfitting (variance)

• underfitting (bias)

• “perfect classifier” (no free lunch)
[learning as generalization] = 
[representation] + [evaluation] + [optimization]

Deﬁnition: Machine Learning … Conceptually 
!
!
1. real-world data

2. graph theory for representation

3. convert to sparse matrix for production work  
leveraging abstract algebra + func programming

4. cost-effective parallel processing for ML apps  
at scale, live use cases

Deﬁnition: Machine Learning … Conceptually 
!
!
1. real-world data

2. graph theory for representation

3. convert to sparse matrix for production work  
leveraging abstract algebra + func programming

4. cost-effective parallel processing for ML apps  
at scale, live use cases
[representation]
[evaluation]
[optimization]
[generalization]
f(x): loss function

g(z): regularization term

discovery
discovery
modeling
modeling
integration
integration
appsapps
systems
systems
business process,

stakeholder
data prep, discovery,
modeling, etc.
software engineering,
automation
systems engineering,
availability
data
science
Data
Scientist
App Dev
Ops
Domain
Expert
Deﬁnition: Machine Learning … Interdisciplinary Teams, Needs × Roles

Definition: Machine Learning … Process, Feature Engineering
evaluationrepresentation optimization use cases
feature engineering
data
sources
data
sources
data prep
pipeline
train
test
hold-outs
learners
classifiers
classifiers
classifiers
data
sources
scoring

Definition: Machine Learning … Process, Tournaments
feature engineering tournaments
data
sources
data
sources
data prep
pipeline
train
test
hold-outs
learners
classifiers
classifiers
classifiers
data
sources
scoring
quantify and measure:
• benefit?
• risk?
• operational costs?
• obtain other data?
• improve metadata?
• refine representation?
• improve optimization?
• iterate with stakeholders?
• can models (inference) inform
first principles approaches?

monoid (an alien life form, courtesy of abstract algebra):

! binary associative operator

! closure within a set of objects

! unique identity element

what are those good for?

! composable functions in workflows

! compiler “hints” on steroids, to parallelize

! reassemble results minimizing bottlenecks at scale

! reusable components, like Lego blocks

! think: Docker for business logic

Monoidify! Monoids as a Design Principle for Efficient
MapReduce Algorithms 
Jimmy Lin, U Maryland/Twitter

kudos to Oscar Boykin, Sam Ritchie, et al.
Definition: Machine Learning … Invasion of the Monoids

Definition: Machine Learning
data
sources
data
sources
data prep
pipeline
train
test
hold-outs
learners
classifiers
classifiers
classifiers
data
sources
scoring
• benefit?
• risk?
discovery
discovery
modeling
modeling
integration
integration
appsapps
systems
systems
data
science
Data
Scientist
App Dev
Ops
Domain
Expert
introduced
capability

Definition
data
sources
data
sources
data prep
pipeline
train
test
hold-outs
learners
classifiers
classifiers
classifiers
data
sources
scoring
• benefit?
• risk?
discovery
discovery
modeling
modeling
integration
integration
appsapps
systems
systems
data
science
Data
Scientist
App Dev
Ops
Domain
Expert
introduced
capability
Can we fit the required  
process into generalized  
workflow definitions?

Deﬁnition: Data Workﬂows

Middleware, effectively, is evolving for Big Data and Machine Learning… 
The following design pattern — a DAG — shows up in many places,  
via many frameworks:
ETL
data
prep
predictive
model
data
sources
end
uses


deﬁnition of a typical Enterprise workﬂow which crosses through
multiple departments, languages, and technologies…
ETL
data
prep
predictive
model
data
sources
end
uses
ANSI SQL for ETL


ETL
data
prep
predictive
model
data
sources
end
usesJava, Pig, etc., for business logic


ETL
data
prep
predictive
model
data
sources
end
uses
SAS for predictive models

ETL
data
prep
predictive
model
data
sources
end
uses
SAS for predictive models
ANSI SQL for ETL most of the licensing costs…

ETL
data
prep
predictive
model
data
sources
end
usesJava, Pig, etc.,
most of the project costs…

ETL
data
prep
predictive
model
data
sources
end
usesJava, Pig, etc.,
most of the project costs…
Something emerges  
to ﬁll these needs,  
for instance…

ETL
data
prep
predictive
model
data
sources
end
uses
Lingual:
DW → ANSI SQL
Pattern:
SAS, R, etc. → PMML
business logic in Java,
Clojure, Scala, etc.
sink taps for
Memcached, HBase,
MongoDB, etc.
source taps for
Cassandra, JDBC,
Splunk, etc.

For example, Cascading and related projects implement the following
components, based on 100% open source: cascading.org
a compiler sees it all… 
one connected DAG:
• troubleshooting
• exception handling
• notiﬁcations
• some optimizations

ETL
data
prep
predictive
model
data
sources
end
uses
Lingual:
DW → ANSI SQL
Pattern:
sink taps for
Memcached, HBase,
MongoDB, etc.
source taps for
Cassandra, JDBC,
Splunk, etc.
Cascading allows multiple departments to combine their workﬂow components
into an integrated app – one among many, typically – based on 100% open source
!
!
FlowDef flowDef = FlowDef.flowDef()!
.setName( "etl" )!
.addSource( "example.employee", emplTap )!
.addSource( "example.sales", salesTap )!
.addSink( "results", resultsTap );!
!
SQLPlanner sqlPlanner = new SQLPlanner()!
.setSql( sqlStatement );!
!
flowDef.addAssemblyPlanner( sqlPlanner );!
!
!
cascading.org

ETL
data
prep
predictive
model
data
sources
end
uses
Lingual:
DW → ANSI SQL
Pattern:
sink taps for
Memcached, HBase,
MongoDB, etc.
source taps for
Cassandra, JDBC,
Splunk, etc.
Cascading allows multiple departments to combine their workﬂow components
into an integrated app – one among many, typically – based on 100% open source
!
!
FlowDef flowDef = FlowDef.flowDef()!
.setName( "classifier" )!
.addSource( "input", inputTap )!
.addSink( "classify", classifyTap );!
!
PMMLPlanner pmmlPlanner = new PMMLPlanner()!
.setPMMLInput( new File( pmmlModel ) )!
.retainOnlyActiveIncomingFields();!
!
flowDef.addAssemblyPlanner( pmmlPlanner );!
!
!

Enterprise DataWorkﬂows with Cascading

O’Reilly, 2013

shop.oreilly.com/product/
0636920028536.do
ETL
data
prep
predictive
model
data
sources
end
uses


O’Reilly, 2013

0636920028536.do
ETL
data
prep
predictive
model
data
sources
end
uses
This begs an update…


O’Reilly, 2013

0636920028536.do
ETL
data
prep
predictive
model
data
sources
end
uses
Because…

Data Workflows for Machine Learning - SF Bay Area ML

Data Workﬂows for
Machine Learning:

 
Examples…

!
“Neque per exemplum”

Example: KNIME

“a user-friendly graphical workbench for the entire analysis 
process: data access, data transformation, initial investigation, 
powerful predictive analytics, visualisation and reporting.”

• large number of integrations (over 1000 modules)

• ranked #1 in customer satisfaction among  
open source analytics frameworks

• visual editing of reusable modules

• leverage prior work in R, Perl, etc.

• Eclipse integration

• easily extended for new integrations
knime.org

Example: Python stack

Python has much to offer – ranging across an  
organization, no just for the analytics staff

• ipython.org
• pandas.pydata.org
• scikit-learn.org
• numpy.org
• scipy.org
• code.google.com/p/augustus
• continuum.io
• nltk.org
• matplotlib.org

Example: Julia

“a high-level, high-performance dynamic programming language 
for technical computing, with syntax that is familiar to users of 
other technical computing environments”

• signiﬁcantly faster than most alternatives

• built to leverage parallelism, cloud computing

• still relatively new — one to watch!
importall Base!
!
type BubbleSort <: Sort.Algorithm end!
!
function sort!(v::AbstractVector, lo::Int, hi::Int, ::BubbleSort, o::Sort.Ordering)!
    while true!
        clean = true!
        for i = lo:hi-1!
            if Sort.lt(o, v[i+1], v[i])!
                v[i+1], v[i] = v[i], v[i+1]!
                clean = false!
            end!
        end!
        clean && break!
    end!
    return v!
end!
julialang.org

Example: Summingbird

“a library that lets you write streaming MapReduce programs 
that look like native Scala or Java collection transformations  
and execute them on a number of well-known distributed 
MapReduce platforms like Storm and Scalding.”

• switch between Storm, Scalding (Hadoop)

• Spark support is in progress

• leverage Algebird, Storehaus, Matrix API, etc.
github.com/twitter/summingbird
def wordCount[P <: Platform[P]]!
(source: Producer[P, String], store: P#Store[String, Long]) =!
source.flatMap { sentence => !
toWords(sentence).map(_ -> 1L)!
}.sumByKey(store)

Example: Scalding

import com.twitter.scalding._!
!
class WordCount(args : Args) extends Job(args) {!
Tsv(args("doc"),!
('doc_id, 'text),!
skipHeader = true)!
.read!
.flatMap('text -> 'token) {!
text : String => text.split("[ [](),.]")!
}!
.groupBy('token) { _.size('count) }!
.write(Tsv(args("wc"), writeHeader = true))!
}
github.com/twitter/scalding

Example: Scalding

• extends the Scala collections API so that distributed lists
become “pipes” backed by Cascading

• code is compact, easy to understand

• nearly 1:1 between elements of conceptual ﬂow diagram  
and function calls

• extensive libraries are available for linear algebra, abstract
algebra, machine learning – e.g., Matrix API, Algebird, etc.

• signiﬁcant investments by Twitter, Etsy, eBay, etc.

• less learning curve than Cascalog

• build scripts… those take a while to run :
github.com/twitter/scalding

Example: Cascalog

(ns impatient.core!
  (:use [cascalog.api]!
        [cascalog.more-taps :only (hfs-delimited)])!
  (:require [clojure.string :as s]!
            [cascalog.ops :as c])!
  (:gen-class))!
!
(defmapcatop split [line]!
  "reads in a line of string and splits it by regex"!
  (s/split line #"[[](),.)s]+"))!
!
(defn -main [in out & args]!
  (?<- (hfs-delimited out)!
       [?word ?count]!
       ((hfs-delimited in :skip-header? true) _ ?line)!
       (split ?line :> ?word)!
       (c/count ?count)))!
!
; Paul Lam!
; github.com/Quantisan/Impatient
cascalog.org

Example: Cascalog

• implements Datalog in Clojure, with predicates backed  
by Cascading – for a highly declarative language

• run ad-hoc queries from the Clojure REPL – 
approx. 10:1 code reduction compared with SQL

• composable subqueries, used for test-driven
development (TDD) practices at scale

• Leiningen build: simple, no surprises, in Clojure itself

• more new deployments than other Cascading DSLs –  
Climate Corp is largest use case: 90% Clojure/Cascalog

• has a learning curve, limited number of Clojure
developers

• aggregators are the magic, and those take effort to learn
cascalog.org

Example: Apache Spark

in-memory cluster computing,  
by Matei Zaharia, RamVenkataraman, et al.

• intended to make data analytics fast to write and to run

• load data into memory and query it repeatedly, much
more quickly than with Hadoop

• APIs in Scala, Java and Python, shells in Scala and Python

• Shark (Hive on Spark), Spark Streaming (like Storm), etc.

• integrations with MLbase, Summingbird (in progress)

// word count!
val file = spark.textFile(“hdfs://…”)!
val counts = file.flatMap(line => line.split(” “))!
.map(word => (word, 1))!
.reduceByKey(_ + _)!
counts.saveAsTextFile(“hdfs://…”)
spark-project.org

Example: MLbase

“distributed machine learning made easy”

• sponsored by UC Berkeley EECS AMP Lab

• MLlib – common algorithms, low-level, written atop Spark

• MLI – feature extraction, algorithm dev atop MLlib

• ML Optimizer – automates model selection,  
compiler/optimizer

• see article: 
http://strata.oreilly.com/2013/02/mlbase-scalable-
machine-learning-made-accessible.html
!
mlbase.org
data = load("hdfs://path/to/als_clinical")!
// the features are stored in columns 2-10!
X = data[, 2 to 10]!
y = data[, 1]!
model = do_classify(y, X)

Example: Spark SQL

“blurs the lines between RDDs and relational tables”

intermix SQL commands to querying external data,
along with complex analytics, in a single app:

• allows SQL extensions based on MLlib

• Shark is being migrated to Spark SQL

Spark SQL: Manipulating Structured Data Using Spark 
Michael Armbrust, Reynold Xin (2014-03-24) 
databricks.com/blog/2014/03/26/Spark-
SQL-manipulating-structured-data-using-
Spark.html 
people.apache.org/~pwendell/catalyst-
docs/sql-programming-guide.html

!
// data can be extracted from existing sources
// e.g., Apache Hive
val trainingDataTable = sql("""
SELECT e.action
u.age,
u.latitude,
u.logitude
FROM Users u
JOIN Events e
ON u.userId = e.userId""")
!
// since `sql` returns an RDD, the results above
// can be used in MLlib
val trainingData = trainingDataTable.map { row =>
val features = Array[Double](row(1), row(2), row(3))
LabeledPoint(row(0), features)
}
!
val model =
new LogisticRegressionWithSGD().run(trainingData)
spark-project.org

Example: Titan

distributed graph database,  
by Matthias Broecheler, Marko Rodriguez, et al.

• scalable graph database optimized for storing and
querying graphs

• supports hundreds of billions of vertices and edges

• transactional database that can support thousands of
concurrent users executing complex graph traversals

• supports search through Lucene, ElasticSearch

• can be backed by HBase, Cassandra, BerkeleyDB

• TinkerPop native graph stack with Gremlin, etc.
// who is hercules' paternal grandfather?!
g.V('name','hercules').out('father').out('father').name
thinkaurelius.github.io/titan

Example: MBrace

“a .NET based software stack that enables easy large-scale 
distributed computation and big data analysis.”

• declarative and concise distributed algorithms in  
F# asynchronous workﬂows

• scalable for apps in private datacenters or public  
clouds, e.g.,Windows Azure or Amazon EC2

• tooling for interactive, REPL-style deployment,
monitoring, management, debugging inVisual Studio

• leverages monads (similar to Summingbird)

• MapReduce as a library, but many patterns beyond
m-brace.net
let rec mapReduce (map: 'T -> Cloud<'R>)!
(reduce: 'R -> 'R -> Cloud<'R>)!
(identity: 'R)!
(input: 'T list) =!
cloud {!
match input with!
| [] -> return identity!
| [value] -> return! map value!
| _ ->!
let left, right = List.split input!
let! r1, r2 =!
(mapReduce map reduce identity left)!
<||>!
(mapReduce map reduce identity right)!
return! reduce r1 r2!
}

Data Workﬂows for
Machine Learning:

 
Update the deﬁnitions…

!
“Nine Points to Discuss”

Data Workflows

Formally speaking, workflows include people (cross-dept)  
and define oversight for exceptional data

…otherwise, data flow or pipeline would be more apt

…examples include Cascading traps, with exceptional tuples  
routed to a review organization, e.g., Customer Support
includes people, defines
oversight for exceptional data
Hadoop
Cluster
source
tap
source
tap sink
tap
trap
tap
customer
profile DBsCustomer
Prefs
logs
logs
Logs
Data
Workflow
Cache
Customers
Support
Web
App
Reporting
Analytics
Cubes
sink
tap
Modeling PMML
1

Data Workﬂows

Footnote –

see also an excellent article related to this point  
and the next:

!
Therbligs for data science:  
A nuts and bolts framework for accelerating data work

Abe Gong @Jawbone

blog.abegong.com/2014/03/therbligs-for-data-science.html

strataconf.com/strata2014/public/schedule/detail/32291

brighttalk.com/webcast/9059/105119

!
!
!
includes people, deﬁnes
oversight for exceptional data
1+

Data Workﬂows

Workﬂows impose a separation of concerns, allowing  
for multiple abstraction layers in the tech stack

…specify what is required, not how to accomplish it

…articulating the business logic

… Cascading leverages pattern language

…related notions from Knuth, literate programming

…not unlike BPM/BPEL for Big Data

…examples: IPython, R Markdown, etc.

separation of concerns, allows
for literate programming
2

Data Workflows

Footnote –

while discussing separation of concerns and a general
design pattern, let’s consider data workflows in a
broader context of probabilistic programming, since
business data is heterogeneous and structured…  
the data for every domain is heterogeneous.

Paraphrasing Ginsberg:

I’ve seen the best minds of my generation destroyed by
madness, dragging themselves through quagmires of large
LCD screens filled with Intellij debuggers and database
command lines, yearning to fit real-world data into their
preferred “deterministic” tools…
separation of concerns, allows
for literate programming
2+
Probabilistic Programming:  
Why,What, How,When 
Beau Cronin 
Strata SC (2014) 
speakerdeck.com/beaucronin/
probabilistic-programming-
strata-santa-clara-2014
Why Probabilistic Programming  
Matters 
Rob Zinkov 
Convex Optimized (2012-06-27) 
zinkov.com/posts/
2012-06-27-why-prob-
programming-matters/

Data Workﬂows

Multiple abstraction layers in the tech stack are  
needed, but still emerging

…feedback loops based on machine data

…optimizers feed on abstraction

…metadata accounting:

• track data lineage

• propagate schema

• model / feature usage

• ensemble performance

…app history accounting:

• util stats, mixing workloads

• heavy-hitters, bottlenecks

• throughput, latency
multiple abstraction layers  
for metadata, feedback,  
and optimization
Cluster
Scheduler
Planners/
Optimizers
Clusters
DSLs
Machine
Data
• app history
• util stats
• bottlenecksMixed
Topologies
Business
Process
Reusable
Components
Portable
Models
• data lineage
• schema propagation
• feature selection
• tournaments
3

Data Workﬂows
Multiple
needed, but still emerging

…feedback loops based on
…optimizers feed on abstraction
…metadata accounting:

• track data lineage
• propagate
• model / feature usage

• ensemble performance

…app history accounting:

• util stats, mixing workloads

• heavy-hitters, bottlenecks

• throughput, latency
multiple abstraction layers  
for metadata, feedback,  
and optimization
Cluster
Scheduler
Planners/
Optimizers
Clusters
DSLs
Machine
Data
• app history
• util stats
Topologies
Business
Process
Reusable
Components
Portable
Models
• data lineage
• tournaments
3
Um, will compilers  
ever look like this??

Data Workflows

Workflows must provide for test, which is no simple  
matter

…testing is required on three abstraction layers:

• model portability/evaluation, ensembles, tournaments

• TDD in reusable components and business process

• continuous integration / continuous deployment

…examples: Cascalog for TDD, PMML for model eval

…still so much more to improve

…keep in mind that workflows involve people, too

testing: model evaluation,
TDD, app deployment
4
Cluster
Scheduler
Planners/
Optimizers
Clusters
DSLs
Machine
Data
• app history
• util stats
Topologies
Business
Process
Reusable
Components
Portable
Models
• data lineage
• tournaments

Data Workflows

Workflows enable system integration

…future-proof integrations and scale-out

…build components and apps, not jobs and command lines

…allow compiler optimizations across the DAG, 
i.e., cross-dept contributions

…minimize operationalization costs:

• troubleshooting, debugging at scale

• exception handling

• notifications, instrumentation

…examples: KNIME, Cascading, etc.

!
future-proof system
integration, scale-out, ops
Hadoop
Cluster
source
tap
source
tap sink
tap
trap
tap
customer
profile DBsCustomer
Prefs
logs
logs
Logs
Data
Workflow
Cache
Customers
Support
Web
App
Reporting
Analytics
Cubes
sink
tap
Modeling PMML
5

Data Workflows

Visualizing workflows, what a great idea.

…the practical basis for:

• collaboration

• rapid prototyping

• component reuse

…examples: KNIME wins best in category

…Cascading generates flow diagrams,  
which are a nice start

visualizing allows people  
to collaborate through code
6

Data Workflows

Abstract algebra, containerizing workflow metadata
…monoids, semigroups, etc., allow for reusable components  
with well-defined properties for running in parallel at scale

…let business process be agnostic about underlying topologies,  
with analogy to Linux containers (Docker, etc.)

…compose functions, take advantage of sparsity (monoids)

…because “data is fully functional” – FP for s/w eng benefits

…aggregators are almost always “magic”, now we can solve  
for common use cases

…read Monoidify! Monoids as a Design Principle for Efficient
MapReduce Algorithms by Jimmy Lin

…Cascading introduced some notions of this circa 2007,  
but didn’t make it explicit

…examples: Algebird in Summingbird, Simmer, Spark, etc.
abstract algebra and
functional programming
containerize business process
7

Data Workﬂows

Footnote –

see also two excellent talks related to this point:

!
Algebra for Analytics 
speakerdeck.com/johnynek/algebra-for-analytics 
Oscar Boykin, Strata SC (2014)

Add ALL theThings: 
Abstract Algebra Meets Analytics 
infoq.com/presentations/abstract-algebra-analytics 
Avi Bryant, Strange Loop (2013)
abstract algebra and
functional programming
7+

Data Workﬂows

Workﬂow needs vary in time, and need to blend time

…something something batch, something something low-latency

…scheduling batch isn’t hard; scheduling low-latency is  
computationally brutal (see the Omega paper)

…because people like saying “Lambda Architecture”,  
it gives them goosebumps, or something

…because real-time means so many different things

…because batch windows are so 1964

…examples: Summingbird, Oryx
blend results from different
time scales: batch plus low-
latency
8
Big Data 
Nathan Marz, James Warren 
manning.com/marz

Data Workflows

Workflows may define contexts in which model selection  
possibly becomes a compiler problem

…for example, see Boyd, Parikh, et al.

…ultimately optimizing for loss function + regularization term

…perhaps not ready for prime-time immediately

…examples: MLbase
optimize learners in context,
to make model selection
potentially a compiler problem
9
f(x): loss function
g(z): regularization term

Data Workﬂows

For one of the best papers about what workﬂows really 
truly require, see Out of the Tar Pit by Moseley and Marks
bonus

Nine Points for Data Workflows — a wish list

1. includes people, defines oversight for exceptional data
2. separation of concerns, allows for literate programming
3. multiple abstraction layers for metadata, feedback, and
optimization
4. testing: model evaluation, TDD, app deployment
5. future-proof system integration, scale-out, ops
6. visualizing workflows allows people to collaborate
through code
7. abstract algebra and functional programming
8. blend results from different time scales:  
batch plus low-latency
9. optimize learners in context, to make model  
selection potentially a compiler problem
Cluster
Scheduler
Planners/
Optimizers
Clusters
DSLs
Machine
Data
• app history
• util stats
Topologies
Business
Process
Reusable
Components
Portable
Models
• data lineage
• tournaments

Nine Points for Data Workﬂows — a scorecard

Spark,
MLbase
Oryx Summing!
bird
Cascalog Cascading
!
KNIME Py Data R
Markdown
MBrace
includes people,
exceptional data
separation of
concerns
multiple  
abstraction layers
testing in depth
future-proof  
system
integration
visualize to collab
can haz monoids
blends batch +  
“real-time”
optimize learners  
in context
can haz PMML ✔ ✔ ✔ ✔ ✔

Data Workﬂows for
Machine Learning:

 
PMML…

!
“Oh, by the way…”

• established XML standard for predictive model markup

• organized by Data Mining Group (DMG), since 1997  
http://dmg.org/

• members: IBM, SAS, Visa, FICO, Equifax, NASA, 
Microstrategy, Microsoft, etc.

• PMML concepts for metadata, ensembles, etc., translate  
directly into workﬂow abstractions, e.g., Cascading

!
“PMML is the leading standard for statistical and data mining models and 
supported by over 20 vendors and organizations.With PMML, it is easy  
to develop a model on one system using one application and deploy the 
model on another system using another application.”
PMML – an industry standard
wikipedia.org/wiki/Predictive_Model_Markup_Language

• Association Rules: AssociationModel element

• Cluster Models: ClusteringModel element

• Decision Trees: TreeModel element

• Naïve Bayes Classiﬁers: NaiveBayesModel element

• Neural Networks: NeuralNetwork element

• Regression: RegressionModel and GeneralRegressionModel elements

• Rulesets: RuleSetModel element

• Sequences: SequenceModel element

• SupportVector Machines: SupportVectorMachineModel element

• Text Models: TextModel element

• Time Series: TimeSeriesModel element
PMML – model coverage
ibm.com/developerworks/industry/library/ind-PMML2/

!
PMML in Action 
Alex Guazzelli,Wen-Ching Lin,Tridivesh Jena 
amazon.com/dp/1470003244

!
See also excellent resources at:

zementis.com/pmml.htm
PMML – further study

Hadoop
Cluster
source
tap
source
tap sink
tap
trap
tap
customer
profile DBsCustomer
Prefs
logs
logs
Logs
Data
Workflow
Cache
Customers
Support
Web
App
Reporting
Analytics
Cubes
sink
tap
Modeling PMML
Pattern – model scoring

• a PMML library for Cascading workﬂows
• migrate workloads: SAS,Teradata, etc.,  
exporting predictive models as PMML

• great open source tools – R, Weka,  
KNIME, Matlab, RapidMiner, etc.

• leverage PMML as another kind of DSL

• only for scoring models, not training

Pattern – model scoring

• originally intended as a general purpose, tuple-oriented
model scoring engine for JVM-based frameworks

• library plugs into Cascading (Hadoop), and ostensibly
Cascalog, Scalding, Storm, Spark, Summingbird, etc.

• dev forum: 
https://groups.google.com/forum/#!forum/pattern-user

• original GitHub repo: 
https://github.com/ceteri/pattern/

• somehow, the other fork became deeply intertwined  
with Cascading dependencies…

• work in progress to integrate with other frameworks,  
add models, and also train models at scale using Spark

## train a RandomForest model!
!
f <- as.formula("as.factor(label) ~ .")!
fit <- randomForest(f, data_train, ntree=50)!
!
## test the model on the holdout test set!
!
print(fit$importance)!
print(fit)!
!
predicted <- predict(fit, data)!
data$predicted <- predicted!
confuse <- table(pred = predicted, true = data[,1])!
print(confuse)!
!
## export predicted labels to TSV!
!
write.table(data, file=paste(dat_folder, "sample.tsv", sep="/"),  
quote=FALSE, sep="t", row.names=FALSE)!
!
## export RF model to PMML!
!
saveXML(pmml(fit), file=paste(dat_folder, "sample.rf.xml", sep="/"))!
Pattern – create a model in R

<?xml version="1.0"?>!
<PMML version="4.0" xmlns="http://www.dmg.org/PMML-4_0"!
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"!
xsi:schemaLocation="http://www.dmg.org/PMML-4_0 
http://www.dmg.org/v4-0/pmml-4-0.xsd">!
<Header copyright="Copyright (c)2012 Concurrent, Inc." description="Random Forest Tree Model">!
  <Extension name="user" value="ceteri" extender="Rattle/PMML"/>!
  <Application name="Rattle/PMML" version="1.2.30"/>!
  <Timestamp>2012-10-22 19:39:28</Timestamp>!
</Header>!
<DataDictionary numberOfFields="4">!
  <DataField name="label" optype="categorical" dataType="string">!
   <Value value="0"/>!
   <Value value="1"/>!
  </DataField>!
  <DataField name="var0" optype="continuous" dataType="double"/>!
</DataDictionary>!
<MiningModel modelName="randomForest_Model" functionName="classification">!
  <MiningSchema>!
   <MiningField name="label" usageType="predicted"/>!
   <MiningField name="var0" usageType="active"/>!
  </MiningSchema>!
  <Segmentation multipleModelMethod="majorityVote">!
   <Segment id="1">!
    <True/>!
    <TreeModel modelName="randomForest_Model" functionName="classification" algorithmName="randomForest" splitCharacteristic="binarySplit">!
     <MiningSchema>!
      <MiningField name="label" usageType="predicted"/>!
     </MiningSchema>!
...!
Pattern – capture model parameters as PMML

public static void main( String[] args ) throws RuntimeException {!
String inputPath = args[ 0 ];!
String classifyPath = args[ 1 ];!
// set up the config properties!
Properties properties = new Properties();!
AppProps.setApplicationJarClass( properties, Main.class );!
HadoopFlowConnector flowConnector = new HadoopFlowConnector( properties );!
// create source and sink taps!
Tap inputTap = new Hfs( new TextDelimited( true, "t" ), inputPath );!
Tap classifyTap = new Hfs( new TextDelimited( true, "t" ), classifyPath );!
// handle command line options!
OptionParser optParser = new OptionParser();!
optParser.accepts( "pmml" ).withRequiredArg();!
OptionSet options = optParser.parse( args );!
!
// connect the taps, pipes, etc., into a flow!
FlowDef flowDef = FlowDef.flowDef().setName( "classify" )!
.addSource( "input", inputTap )!
.addSink( "classify", classifyTap );!
!
if( options.hasArgument( "pmml" ) ) {!
String pmmlPath = (String) options.valuesOf( "pmml" ).get( 0 );!
PMMLPlanner pmmlPlanner = new PMMLPlanner()!
.setPMMLInput( new File( pmmlPath ) )!
.retainOnlyActiveIncomingFields()!
.setDefaultPredictedField( new Fields( "predict", Double.class ) ); // default value if missing from the model!
flowDef.addAssemblyPlanner( pmmlPlanner );!
}!
!
// write a DOT file and run the flow!
Flow classifyFlow = flowConnector.connect( flowDef );!
classifyFlow.writeDOT( "dot/classify.dot" );!
classifyFlow.complete();!
}
Pattern – score a model, within an app

Customer
Orders
Classify
Scored
Orders
GroupBy
token
Count
PMML
Model
M R
Failure
Traps
Assert
Confusion
Matrix
Pattern – score a model, using pre-deﬁned Cascading app

## run an RF classifier at scale!
!
hadoop jar build/libs/pattern.jar !
data/sample.tsv out/classify out/trap !
--pmml data/sample.rf.xml!
!
!
Pattern – score a model, using pre-deﬁned Cascading app

Data Workﬂows for
Machine Learning:

 
Summary

!
“feedback and discussion”

 
?

PMML – what’s needed?

in the language standard:

! data preparation

! data sources/sinks as parameters

! track data lineage

! more monoid, less XML

! handle data exceptions => TDD

! updates/active learning?

! tournaments?

!
in the workflow frameworks:

! include feature engineering w/ model?

! support evaluation better

! build ensembles
data
sources
data
sources
data prep
pipeline
train
test
hold-outs
learners
classifiers
classifiers
classifiers
data
sources
scoring
• benefit?
• risk?

Thank you, and…

stay tuned for

Just Enough Math

(Summer 2014 w/ Allen Day)

!
monthly newsletter for updates, events,
conference summaries, etc.:

liber118.com/pxn/

Data Workflows for Machine Learning - SF Bay Area ML

More Related Content

What's hot (20)

Viewers also liked (8)

Similar to Data Workflows for Machine Learning - SF Bay Area ML (20)

More from Paco Nathan (20)

Recently uploaded (20)

Data Workflows for Machine Learning - SF Bay Area ML