Transcript Data Mining Slides - San Diego Supercomputer Center

Intro to Data Mining
1
Natasha Balac, Ph.D.
Predictive Analytics Center of Excellence, Director
San Diego Supercomputer Center
University of California, San Diego
Necessity is the Mother of Invention

Problem

Data explosion

Automated data collection tools and mature database technology lead to
tremendous amounts of data stored in databases, data warehouses and
other information repositories

“We are drowning in data, but starving for knowledge!” (John Naisbitt,
1982)
2
Necessity is the Mother of Invention

Solution

Predictive Analytics or Data Mining

Extraction or “mining” of interesting knowledge (rules, regularities,
patterns, constraints) from data in large databases

Data -driven discovery and modeling of hidden patterns in large
volumes of data

Extraction of implicit, previously unknown and unexpected, potentially
extremely useful information from data
3
Predictive Analytics
Analytical Tools
Top-Down
Methodology
Surface
Shallow
Data
Bottom-Up
Methodology
4
Hidden
SQL tools for simple
queries and reporting
Statistical & OLAP
tools for summaries
and analysis
Data Mining methods
for knowledge
discovery
Query Reporting OLAP
Data Mining
Extraction of
data - detailed
and/or
summarized
Analysis,
summaries,
Trends
Information
Analysis
Who purchased
the product in
the last 2
quarters?
What is an
average income
of the buyers per
quarter by
district?
Discovery of
hidden patterns,
information,
predicting future
trends
Insight
knowledge and
prediction
Which customers
are likely to buy a
similar product in
the future and
why?
5
DM Enables Predictive Analytics
Role of Software
Data mining
Proactive
Predictive Analytics
Interactive
OLAP
Ad-hoc reporting
Passive
Canned reporting
Presentatio
n
Exploration
Discovery
Business
Insight
What Is Data Mining?

Data mining:

7
Extraction of interesting (non-trivial, implicit,
previously unknown and potentially useful)
information or patterns from data in large databases
Data Mining is NOT


Data Warehousing
(Deductive) query processing






8
SQL/ Reporting
Software Agents
Expert Systems
Online Analytical Processing (OLAP)
Statistical Analysis Tool
Data visualization
Machine Learning Techniques
Technical basis for data mining: algorithms for
acquiring structural descriptions from examples

Methods originate from artificial intelligence,
statistics, and research on databases

9
Multidisciplinary Field
Database
Technology
Machine
Learning
Artificial
Intelligence
10
Statistics
Data Mining
Visualization
Other
Disciplines
Multidisciplinary Field


Database technology
Artificial Intelligence






11
Machine Learning including Neural Networks
Statistics
Pattern recognition
Knowledge-based systems/acquisition
High-performance computing
Data visualization
History of Data Mining
12
History



Emerged late 1980s
Flourished –1990s
Roots traced back along three family lines



13
Classical Statistics
Artificial Intelligence
Machine Learning
Statistics

Foundation of most DM technologies



14
Regression analysis, standard
distribution/deviation/variance, cluster
analysis, confidence intervals
Building blocks
Significant role in today’s data mining –
but alone is not powerful enough
Artificial Intelligence




15
Heuristics vs. Statistics
Human-thought-like processing
Requires vast computer processing power
Supercomputers
Machine Learning

Union of statistics and AI

Blends AI heuristics with advanced statistical
analysis
Machine Learning – let computer programs



16
learn about data they study - make different
decisions based on the quality of studied data
using statistics for fundamental concepts and
adding more advanced AI heuristics and
algorithms
Data Mining




Adoption of the Machine learning techniques
to the real world problems
Union: Statistics, AI, Machine learning
Used to find previously hidden trends or
patterns
Finding increasing acceptance in science and
business areas which need to analyze large
amount of data to discover trends which could
not be found otherwise
17
Terminology







18
Gold Mining
Knowledge mining from databases
Knowledge extraction
Data/pattern analysis
Knowledge Discovery Databases or KDD
Information harvesting
Business intelligence
What does Data Mining Do?
Explores
Your Data
Finds
Patterns
Performs
Predictions
What can we do with Data Mining?



Exploratory Data Analysis
Predictive Modeling: Classification and Regression
Descriptive Modeling


Discovering Patterns and Rules




20
Cluster analysis/segmentation
Association/Dependency rules
Sequential patterns
Temporal sequences
Deviation detection
CRISP-DM - Cross Industry
Standard Process for Data Mining
CRISP-DM Process Model
Data Mining Applications




Science: Chemistry, Physics, Medicine
 Biochemical analysis, remote sensors on a satellite, Telescopes – star
galaxy classification, medical image analysis
Bioscience
 Sequence-based analysis, protein structure and function prediction,
protein family classification, microarray gene expression
Pharmaceutical companies, Insurance and Health care, Medicine
 Drug development, identify successful medical therapies, claims
analysis, fraudulent behavior, medical diagnostic tools, predict office
visits
Financial Industry, Banks, Businesses, E-commerce
 Stock and investment analysis, identify loyal customers vs. risky
customer, predict customer spending, risk management, sales
forecasting
22
Data Mining Applications

Market analysis and management


Risk analysis and management


Forecasting, customer retention, improved underwriting, quality
control, competitive analysis (Banking and Credit Card scoring)
Fraud detection and management

23
Target marketing, customer relation management, market basket
analysis, cross selling, market segmentation (Credit Card scoring,
Personalization & Customer Profiling )
Fraud, waste and abuse including: illegal practices, waste, payment
error, non-compliance, incorrect billing practices
Data Mining Applications

Sports and Entertainment


Astronomy


24
IBM Advanced Scout analyzed NBA game statistics (shots
blocked, assists, and fouls) to gain competitive advantage for
New York Knicks and Miami Heat
JPL and the Palomar Observatory discovered 22 quasars with
the help of data mining
Campaign Management and Database Marketing
Data Mining Tasks

Concept/Class description: Characterization and
discrimination


Generalize, summarize, and contrast data characteristics, e.g., dry
vs. wet regions; “normal” vs. fraudulent behavior
Association (correlation and causality)

Multi-dimensional interactions and associations
age(X, “20-29”) ^ income(X, “60-90K”)  buys(X, “TV”)
Hospital(area code) ^ procedure(X) ->claim (type) ^ claim(cost)
25
Data Mining Tasks

Classification and Prediction

Finding models (functions) that describe and distinguish
classes or concepts for future prediction

Example: classify countries based on climate, or
classify cars based on gas mileage, fraud based on
claims information

Presentation:


26
If-THEN rules, decision-tree, classification rule, neural network
Prediction: Predict some unknown or missing numerical
values
Data Mining Tasks

Cluster analysis

Class label is unknown: Group data to form new
classes


27
Example: cluster claims or providers to find distribution
patterns of unusual behavior
Clustering based on the principle: maximizing the
intra-class similarity and minimizing the interclass
similarity
Data Mining Tasks

Outlier analysis

Outlier: a data object that does not comply with the general
behavior of the data

Mostly considered as noise or exception, but is quite useful
in fraud detection, rare events analysis

Trend and evolution analysis

Trend and deviation: regression analysis

Sequential pattern mining, periodicity analysis
28
KDD Process
Database
Selection
Transformation
Data
Preparation
Training
Data
Data
Mining
Evaluation,
Verification
29
Model,
Patterns
Learning and Modeling Methods





Decision Tree Induction (C4.5, J48)
Regression Tree Induction (CART, MP5)
Multivariate Regression Tree (MARS)
Clustering (K-means, EM, Cobweb)
Artificial Neural Networks (Backpropagation,
Recurrent)


30
Support Vector Machines (SVM)
Various other models
TAXONOMY



Predictive Methods
 Use some variables to predict some unknown
or future values of other variables
Descriptive Methods
 Find human –interpretable patterns that
describe the data
Supervised vs. Unsupervised

Labeled vs. unlabeled data
Data Mining Challenges



32
Computationally expensive to investigate all
possibilities
Dealing with noise/missing information and errors
in data
Mining methodology and user interaction

Mining different kinds of knowledge in databases

Incorporation of background knowledge

Handling noise and incomplete data

Pattern evaluation: the interestingness problem

Expression and visualization of data mining results
Data Mining Heuristics and Guide





Choosing appropriate attributes/input
representation
Finding the minimal attribute space
Finding adequate evaluation function(s)
Extracting meaningful information
Model evaluation accuracy vs. overfitting
33
Available Data Mining Tools








34
COTs:
IBM Intelligent Miner
SAS Enterprise Miner
Oracle ODM
Microstrategy
Microsoft DBMiner
Pentaho
Matlab
Teradata







Open Source:
WEKA
KNIME
Orange
RapidMiner
NLTK
R
Rattle
Data Mining Trends


Many data mining problems involve large,
complex databases, complicated modeling
techniques and substantial computer processing
Scalability is very important due to the rapid
growth in the amount the data and the need to
build and deploy the models at rapid rates
Needs for Scalable High
Performance Data Mining




The size of a database as well as factors in
building, testing, validation and deploying a data
mining solution
Taking advantage of parallel database/file system
systems and additional CPUs
Work with more data, build more models, and
improve their accuracy by simply adding
additional CPUs
Build a good data mining model as quickly as
possible!
Scalable High Performance Data
Mining



One solution: run parallel data mining
algorithms and parallel DBMSs on parallel
hardware
Once the model is build – still need for parallel
computing to apply the model to a large amount
of data
Scoring: data mining model applied to a batch
of data, record or events at the time; require
that a scalable solution be deployed
Data mining applications on Gordon
DM Suites
MathWorks
Octave
Computational Packages with DM tools
Others as
Requested
Libraries for building tools
38
Summary

Discovering interesting patterns from large amounts
of data

CRISP-DM Industry standard

Learn from the past


Predict for the future


39
High quality, evidence based decisions
Prevent future instances of fraud, waste & abuse
React to changing circumstances

Current models, continuous learning
Thank you!
Questions:
[email protected]
40
An Introduction to the Gordon Architecture
San Diego Supercomputer Center
Gordon Design Partnership
Funding and Oversight from the Office of
Cyberinfrastructure
Design, Deployment, Support, Management
System Integrator
Sandy Bridge processors, flash memory,
motherboards
vSMP Foundation Memory Aggregation Software
3D Torus Subnet Manager, IB Switches, HCAs
Gordon is not about FLOPS, but …
A conservative estimate
puts Gordon in the top 50
on the Top 500 list
Access to Big Data Comes with a Latency Penalty
1.00E+00
Data Oasis
Lustre 4PB PFS
1.00E-01
64 I/O nodes
300 TB Intel SSD
(lower is
better)
Latency (seconds)
1.00E-02
1.00E-03
Quick Path
Interconnect
10’s of GB
1.00E-04
1.00E-05
1.00E-06
1.00E-07
1.00E-08
QDR InfiniBand
Interconnect
100’s of GB
L3 Cache
MB
L1 Cache
KB
1.00E-09
1.00E-05
L2 Cache
KB
1.00E-03
1.00E-01
1.00E+01
DDR3
Memory
10’s of GB
1.00E+03
1.00E+05
1.00E+07
Data Capacity (GB)
(higher is
better)
1.00E+09
Access to Big Data Comes with a Latency Penalty
1.00E+00
Data Oasis
Lustre 4PB PFS
1.00E-01
64 I/O nodes
300 TB Intel SSD
(lower is
better)
Latency (seconds)
1.00E-02
1.00E-03
Quick Path
Interconnect
10’s of GB
1.00E-04
1.00E-05
1.00E-06
1.00E-07
1.00E-08
QDR InfiniBand
Interconnect
100’s of GB
L3 Cache
MB
L1 Cache
KB
1.00E-09
1.00E-05
L2 Cache
KB
1.00E-03
1.00E-01
1.00E+01
DDR3
Memory
10’s of GB
1.00E+03
1.00E+05
1.00E+07
Data Capacity (GB)
(higher is
better)
1.00E+09
Flash Drives are a Good Fit for Data Intensive
Computing
Flash Drive
Typical
HDD
Good for
Data
Intensive
Apps
< .1 ms
10 ms
✔
250 /170 MB/s
100 MB/s
✔
35,000/
2000
100
✔
2-5 W
6-10 W
✔
1M hours
1M hours
-
Price/GB
$2/GB
$.50/GB
-
Endurance
2-10PB
N/A
✔
Latency
Bandwidth (r/w)
IOPS (r/w)
Power consumption
MTBF
Total Cost of Ownership
*** The jury is still out ***
.
Apart from the differences between HDD and SSD it is not common to find
local storage “close” to the compute. We have found this to be attractive in
our Trestles cluster, which has local flash on the compute, but is used for
traditional HPC applications (not high IOPS).
Flash Glossary of Terms
Term
Definition
NAND
Physical silicon MOSFET storage array. The flash storage media.
SLC: Single-level cell
Type of NAND storage device. Lower storage density for the same price. Higher
endurance and performance than MLC. 1 bits of data per cell.
MLC: Multi-level cell
Type of NAND storage device. Higher storage density for the same price. Lower
endurance and performance than SLC. 2 bits of data per cell.
eMLC: Enterprise MLC
Form of eMLC that uses high quality NAND, and additional controller software to
achieve higher performance and endurance.
IOPS: I/O operations per second
Key storage benchmark that is relevant for data intensive applications. The ability to
sustain high IOPS provides performance for applications that exhibit high random
access data patterns.
Wear Leveling
The process of distributing write operations uniformly over all of the blocks in the flash
memory chips to preserve the life of the NAND. Controlled by the SSD controller
software.
Write Amplification Factor
Ratio of actual erase operations to the minimum required to erase the data. Greater
than 1 is bad. Less than 1 is possible with compression (though most HPC data is
not compressible)
Overprovisioning
Space on the flash drive not available to the user that is used for garbage collection,
wear leveling, and containing bad blocks. Overprovisioning can increase endurance.
Program/Erase (P/E) cycle
A flash memory cell must be erased before it can be written to. This causes cell wear
out and is what sets the endurance of an SSD. Minimum erase size for SSD’s is a
block.
Endurance
How much data can be written to an SSD before it ceases to function properly. Set
by the P/E cycle count. This is lower for MLC due to the number of programmable
Gordon System Specification
Intel Sandy Bridge Compute node
Sockets & Cores
2 & 12*
Clock speed
2.0 *
DIMM slots per socket
4
DRAM capacity
64 GB
Intel Flash I/O Node
NAND flash SSD drives
16 or more
SSD capacity per drive/Capacity per node
256 GB / 4 TB *
Bandwidth per drive (r/w)
250 MB/s / 170 MB/s
IOPS (r/w)
35,000 / 2000
SMP Super-Node (via vSMP)
Compute nodes / I/O Nodes
32 / 2
Addressable DRAM
2 TB
Addressable memory including flash
10 TB
Gordon
Compute Nodes
1,024
Total compute cores
12,288
Peak performance
200 TF *
Aggregate memory
64 TB DRAM; 256 TB flash
InfiniBand Interconnect
Aggregate torus BW
9.2 TB/s
Type
Dual-Rail QDR InfiniBand
Link Bandwidth
8 GB/s (bidirectional)
Latency (min-max)
1.25 µs – 2.5 µs
Lustre-based Disk I/O Subsystem
Total storage
4 PB (raw)
* May be revised
I/O bandwidth
100 GB/s
Gordon Applications Software
chemistry
adf
amber
gamess
gaussian
gromacs
lammps
namd
nwchem
distributed
computing
globus
Hadoop
MapReduce
visualization
idl
NCL
paraview
tecplot
visit
VTK
genomics
abyss
blast
hmmer
soapdenovo
velvet
compilers/languages
gcc, intel, pgi
MATLAB, Octave, R
PGAS (UPC)
DB2, PostgreSQL
data mining
IntelligentMiner
RapidMiner
RATTLE
Weka
libraries
ATLAS
BLACS
fftw
HDF5
Hypre
SPRNG
superLU
MapReduce Paradigmatic Example:
string counting

Scheduler: manage threads, initiate data split and call Map

Map: count strings, output key=string & value=count

Scheduler: re-partitions keys & values

Reduce: sum up counts
User defines keys & values
50
User defined functions
MR provides parallelization,concurrency,
and intermediate data functions (by key&value)
Matlab and Data Analysis

Mathematical and matrix operations

Interactive or scripts

Comes with tools (scripts) for data analysis,


e.g. clustering, neural networks, SVMs, …
Uses MKL (Intels math kernel library) for Matrix calculations
with threads
51
Matlab in HPC setting

Distributed toolbox provides distribute/gather
functions

In a nutshell:

Create job object: createMatlabPoolJob(scheduler information)

Create tasks for that job createTask(job,@myfunction,#tasks,{parameters..})

In your code: spmd
D=codistributed(X); or D=codistributed.build(X);
< statements>
end;
52
Matlab in vSMP setting

vSMP submission indicates threads
In submission script,
set environment variables for MKL
In matlab code:
setenv('MKL_NUM_THREADS', getenv(number_of_procs);

No programming changes necessary, but programming
considerations exist
53
Matlab: matrix multiplication

In vSMP: Y=X’NxN * XNxN
8 threads
time(s)
32 threads
N=10K
Gb=2
54
20K
6.5
30K
14
40K
25
Matrix size
50K
40
Matlab: matrix inversion

In vSMP: Y=inv(XNxN + I*0.0001)
32 threads
time(s)
time(s)
16 threads
N=10K
Gb=2
55
20K
6.5
30K
14
Matrix size
40K
25
50K
40
Matlab and Data Mining Case Study

Kmeans clustering

Assign each point to one of a few clusters so that total
distance to center is minimized

Options: distance function, number of clusters, initial cluster
centers, number of iterations, stopping criteria
56
Matlab original Kmeans Script
1. Difference_by_col=X(:,1)-Cluster_Means(1,1)
XNxP
each row
is a point in
Rp
00100 0200 …00100100010000010001000100
0
00002 0000 …000030010001000001010000000
00000 0100 …10000000100010000 00100 0000
…
…
…
…
0 0 0000 0400 0100 0100 0200100 0200 …00100
Cluster_MeansMxP
0 0 1 0 0 0 0 .5 0 0 … 0 0 1 0 0 .5 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0
1 0 0 0 0 .5 0 0 0 … 1 0 0 0 .4 0 0 0 1 0 0 0 1 0 0 0 0 0 0 1 0 0
…
…
0 0 0000 0400 0100 0100 0200100 0200 …00
2. square difference
3. sum as you loop across cols to get Distances to cluster center
Works better for large N small P
57
Matlab Kmeans Script altered
1. Difference_by_row=X(1,:)-Cluster_Means(1,:)
XNxP
each row
is a point in
Rp
00100 0200 …00100100010000010001000100
00002 0000 …000030010001000001010000000
00000 0100 …10000000100010000 00100 0000
…
…
…
…
0 0 0000 0400 0100 0100 0200100 0200 …00100
Cluster_MeansMxP
0 0 1 0 0 0 0 .5 0 0 … 0 0 1 0 0 .5 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0
1 0 0 0 0 .5 0 0 0 … 1 0 0 0 .4 0 0 0 1 0 0 0 1 0 0 0 0 0 0 1 0 0
…
…
0 0 0000 0400 0100 0100 0200100 0200 …00
2. dot(difference_by_row)
3. loop across rows to get Distances
Works better for large P and dot( ) will use threads
58
Matlab Kmeans Benchmarks

Kmeans on 10,000,000 entries from NYTimes articles
(http://archive.ics.uci.edu/ml/datasets/Bag+of+Words)

Running as full data matrix ~ 45K articles x102K words,

Each cell holds word count (double float)

about 37Gb in Matlab, total memory for script about 61Gb

Kmeans (original) runtime ~ 50 hours

Kmeans (altered) runtime ~ 10 hours, 8 threads
59
Matlab Kmeans Results
cluster means shown
with coordinates
determining fontsize
7 viable clusters found
60
MapReduce Framework

A library for distributed computing

Started by Google, gaining popularity

Various implementations: Hadoop (distributed), Phoenix (threaded)
User outputs keys & values
61
User defined functions
MR provides parallelization,concurrency,
and intermediate data functions (by key&value)
MapReduce Paradigmatic Example:
string counting

Scheduler: manage threads, initiate data split and call Map

Map: count strings, output key=string & value=count

Scheduler: re-partitions keys & values

Reduce: sum up counts
User defines keys & values
62
User defined functions
MR provides parallelization,concurrency,
and intermediate data functions (by key&value)
MapReduce Kmeans clustering

C-code for Kmeans (sample code with MapReduce)

Use 10,000,000 entries from NYTimes articles

Running as full data matrix ~ 45Kx102K ints, about 20 Gb
total in vSMP

Running time about 20 min, 32 threads
(but not a completely fair comparison to Matlab kmeans)
63
Case Study: Matrix Factorization
(work by R. Anil & C. Elkan, UCSD, CSE for KDD 2011 competition)

Given large N sparse data matrix
sparse
missing
data
XNxP
customer
ratings
items rated
***1** *2** **1**1***1*
* 2 * * * * * * * * 3 * * 1 * * * ** * * *
****0*5**1** 1*******1*
…
…
…
…
…
…
…
* * **** *4** *1** *1** *
Find vectors U,V such that X ~ ∑ f(U • V’) + penalty
64
Case Study: Matrix Factorization

C code with pthreads

For different f(U • V’) functions
65
Function
Time (s)
1 iteration
Memory
Dash node
Sigmoidal
74
29Gb
non-vSMP
Alternating Least
Squares
673
15Gb
non-vSMP
Log-linear
1110
70Gb
vSMP
Additional Methods Slides
66
Decision Tree Induction

Method for approximating discrete-valued
functions



67
robust to noisy/missing data
can learn non-linear relationships
inductive bias towards shorter trees
Decision Tree Induction

Applications:






68
medical diagnosis – ex. heart disease
analysis of complex chemical compounds
classifying equipment malfunction
risk of loan applicants
Boston housing project – price prediction
fraud detection
DT for Medical Diagnosis and Prognosis
Heart Disease
<= 91
Class 2:
Minimum systolic blood
pressure over a 24-hour
period following
admission to the hospital
> 91
Age of Patient
<=62.5
>62.5
Early death
Class 1:
Was there sinus
tachycardia?
Survivors
NO
Beriman et. al, 1984
69
YES
Class 1:
Class 2:
Survivors
Early death
Regression Tree Induction

Why Regression tree?

Ability to:



70
Predict continuous variable
Model conditional effects
Model uncertainty
Regression Trees



71
Quinlan, 1992
Continuous goal
variables
Induction by means of
an efficient recursive
partitioning algorithm
Uses linear regression
to select internal nodes
Clustering
 Basic idea: Group similar things together
 Unsupervised Learning – Useful when no
other info is available
 K-means
 Partitioning instances into k disjoint clusters
 Measure of similarity
72
Clustering
X
X
X
73
X
Kmeans Results from 10 million
NYTimes articles
cluster means shown
with coordinates
determining fontsize
7 viable clusters found
74
Artificial Neural Networks (ANNs)
Network of many simple units
•
Main Components
•
•Inputs
•Hidden
layers
•Outputs
Adjusting weights of
connections
•
Backpropagation
•
75
Support Vector Machines (SVM)



Blend of linear modeling and instance based learning
SVM select a small number of critical boundary instances
called support vectors from each class and build a linear
discriminant function that separates them as widely as
possible
They transcend the limitations of linear boundaries by
making it practical to include extra nonlinear terms in the
calculations

76
making it possible to form quadratic, cubic, higher-order decision
boundaries
Support vectors

Resilient to overfitting because they learn a
particular linear decision boundary



The instances closest to the maximum margin
hyperplane are called support vectors
Important observation: the support vectors define
the maximum margin hyperplane

77
The maximum margin hyperplane
All other instances can be deleted without changing the
position and orientation of the hyperplane
SVM: finding the maximum
margin hyperplane
78
Data Miners: Past and Present

Traditional approaches have been for DM
experts: “White-coat PhD statisticians”


DM tools also fairly expensive
Today: approach is designed for those with
some Database/Analytics skills


DM built into DB, easy to use GUI, Workflows
Many jobs available from Statistical analyst to
Data Scientist!
Data scientist:
The hot new gig in tech


80
Article in Fortune –
“The unemployment rate in the U.S.
continues to be abysmal (9.1% in July),
but the tech world has spawned a new
kind of highly skilled, nerdy-cool job that
companies are scrambling to fill: data
scientist”