Slide 1
Slide 1 text
Staying MAD : in-database analytics with
MADlib
Ike Okonkwo (@ikeondata)
Zipfian Academy
3.3.14
Thursday, March 6, 14
Slide 2
Slide 2 text
What is MADlib ?
• Started out as academic project
• UC Berkeley
• University of Wisconsin
• University of Florida
• Industry backing from Pivotal / EMC
• Aims to become ‘CRAN’ for databases
• Currently runs on Greenplum DB / PostgreSQL
• Provides scalable analytics in SQL DBMS
Thursday, March 6, 14
Slide 3
Slide 3 text
In-Database Analytics ?
• Performing advanced analytics and machine
directly in the database
• MADlib implements machine learning algorithms in
SQL
• python drivers for complex tasks that require
multiple iterations on data - MCMC, Gradient
Descent
• Calls optimized C/C++ Linear Algebra libraries for
matrix math
Thursday, March 6, 14
Slide 4
Slide 4 text
Being MAD
Thursday, March 6, 14
Slide 5
Slide 5 text
Being MAD
• Magnetic - consume all types of data
Thursday, March 6, 14
Slide 6
Slide 6 text
Being MAD
• Magnetic - consume all types of data
• Agile - modeling, ETL, iterating
Thursday, March 6, 14
Slide 7
Slide 7 text
Being MAD
• Magnetic - consume all types of data
• Agile - modeling, ETL, iterating
• Deep - statistical and machine learning algorithms
Thursday, March 6, 14
Slide 8
Slide 8 text
Being MAD
• Magnetic - consume all types of data
• Agile - modeling, ETL, iterating
• Deep - statistical and machine learning algorithms
• library of parallel and scalable tools
Thursday, March 6, 14
Slide 9
Slide 9 text
Traditional BI / Analytics
Database / Data Warehouse Analytics
Reports
Scripts
Thursday, March 6, 14
Slide 10
Slide 10 text
Traditional BI / Analytics
Database / Data Warehouse Analytics
Reports
Scripts
Thursday, March 6, 14
Slide 11
Slide 11 text
MADlib
Database / Data Warehouse
Analytics
Thursday, March 6, 14
Slide 12
Slide 12 text
MADlib Tools
• Supervised Learning
• Linear Regression, Logistic Regression, Naive Bayes,Decision Trees, SVM
• Unsupervised Learning
• K-means, SVD, LDA, Association Rules
• Support Modules
• Sparse Matrices, Array Operation, Conjugate Gradient Optimization
Thursday, March 6, 14
Slide 13
Slide 13 text
Under the hood
Thursday, March 6, 14
Slide 14
Slide 14 text
Under the hood
Lasso (L1)
Ridge (L2)
Linear Regression
SVM
Loss Functions
MADlib uses Stochastic Gradient Descent under the hood to
implement these models
Thursday, March 6, 14
Slide 15
Slide 15 text
Open Source Interfaces
• PyMADlib
• PivotalR
Thursday, March 6, 14
Slide 16
Slide 16 text
Other Approaches...
• Hadoop Stack
• Mahout, Cascading, Cascalog, MapReduce
• BDA Stack
• Spark
• Others
• Graphlab, Vowpal Wabbit
• SAS HPA
• $$$$
Thursday, March 6, 14
Slide 17
Slide 17 text
Want to learn more
• Pivotal Open Source Hub meetup group
• Topic - Big Data Analytics : Scalable machine learning using open-source
tools
• Where : Pivotal Labs, SF
• When : 3.4.14 - 6.30pm
Thursday, March 6, 14
Slide 18
Slide 18 text
References
• http://www.slideshare.net/jdegoes/indatabase-predictive-analytics
• http://www.slideshare.net/SrivatsanRamanujam/py-ma-dlibdatadaytexas
• http://www.slideshare.net/SrivatsanRamanujam/python-powered-data-science-at-pivotal-
pydata-2013?v=qf1&b=&from_search=9
• http://www.slideshare.net/SarahAerni/data-science-as-a-commodity-use-madlib-r-other-oss-
tools-for-data-science-from-pivotal-open-source-hub-meetup?v=qf1&b=&from_search=6
• http://www.slideshare.net/christangrant/mad-skills-9012267?v=qf1&b=&from_search=11
• http://www.eecs.berkeley.edu/Pubs/TechRpts/2012/EECS-2012-38.pdf
• http://sacan.biomed.drexel.edu/vldb2012/program/?
volno=vol5no12&pid=935&downloadslides=1
• Introduction to Statistical Learning, Hastie, Tibshirani, et al
Thursday, March 6, 14
Slide 19
Slide 19 text
Thursday, March 6, 14
Slide 20
Slide 20 text
Backup Slides
• array operations implemented as UDFs
• UDA and window aggregates
• Common use case : Legacy Data Warehouse with
no flexibility
• supports statistical text analysis - feature
extraction, string matching (n-grams), Viterbi and
MCMC inference
Thursday, March 6, 14