Large Scale Data Mining using Genetics-Based Machine Learning

Large Scale Data Mining using
Genetics-Based Machine Learning
Jaume Bacardit Xavier Llorà
School of Computer Science & Google Inc. 1600
School of Bioscience Amphitheatre Pkwy
University Nottingham Mountain View, CA 94043
Nottingham, UK

jqb@cs.nott.ac.uk xllora@gmail.com

GECCO 2012 Tutorial, July 7th 2012

Copyright is held by the author/owner(s).
GECCO’12 Companion, July 7–11, 2012,
Philadelphia, PA, USA. ACM 978-1-4503-1178-6/12/07.

Machine Learning and Data
Mining
 Core of Data Mining  Machine
learning: How to construct programs New
that automatically learn from Instance
experience [Mitchell, 1997]

Training Learning Models Inference
Set Algorithm Engine

Annotated
Instance

What Will We Cover?

• What does large scale mean?
• Evolution as massive parallel processing
• The challenges of data mining
• Kaleidoscopic large scale data mining
• Real examples
• Summary and further directions

WHAT DOES LARGE SCALE
MEAN?
Evolution as massive parallel processing
The challenges of data mining
Kaleidoscopic large scale data mining
Real-world examples
Wrapping up

What Does Large Scale Mean?

• Many scientific disciplines are currently
experiencing a massive “data deluge”
• Vast amounts of data are available thanks to
initiatives such as the human genome project or
the virtual human physiome
• Data mining technologies need to deal with large
volumes of data, scale accordingly, extract
accurate models, and provide new insight
• So, what does large mean?

Large Meaning… Piles of
Records
• Datasets with a high number of records
– This is probably the most visible dimension of large
scale data mining
– GenBank (the
genetic sequences
database from the
NIH) contains (Apr,
2011) more than
135 million gene
sequences and
more than 126
billion nucleotides

Large Meaning… Piles of
Records
• Datasets with a high number of records
– Not all data comes from the natural sciences
– Netflix Prize:
• Generating better movie
recommending methods
from customer ratings
• Training set of 100M ratings
from over 480K customers
on 78K movies
• Data collected from October
1998 and December, 2005
• Competition lasted from
2006 to 2009
• Think big: Twitter, Facebook?

Large Meaning… High
Dimensionality
• High dimensionality domains
– Sometimes each record is characterized by hundreds,
thousands (or even more) features
– Microarray technology (as many
other post-genomic data generation
techniques) can routinely generate
records with tens of thousands of
variables
– Creating each record is usually very
costly, so datasets tend to have a
very small number of records. This
unbalance between number of
records and number of variables is
yet another challenge
(Reinke, 2006, Image licensed under Creative Commons)

Large Meaning… Rare

• Class unbalance
– Challenge to generate accurate classification models
where not all classes are equally represented
– Contact Map prediction
datasets (briefly explained
later in the tutorial) routinely
contain millions of instances
from which less than 2% are
positive examples
– Tissue type identification is
highly unbalance—see figure
(Llora, Priya, Bhargava, 2009)

Large Meaning… Lots of Classes

• Yet another dimension of difficulty
• Reuters-21578 dataset is a text categorization
task with 672 categories
• Very related to the class unbalance problem
• Machine learning methods need to make an extra effort
to make sure that underrepresented data is taken into
account properly

What does large scale mean?

EVOLUTION AS MASSIVE
PARALLEL PROCESSING
Real-world examples
Wrapping up

Evolution and Parallelism

• Evolutionary algorithms are parallelism rich
• A population is data rich (individuals)
• Genetic operators are highly parallel operations

Ind. 1 Ind. 1 Ind. 1 Ind. 1

Ind. 2 Ind. 2 Ind. 2 Ind. 2

Ind. n Ind. n Ind. n Ind. n

evaluation selection crossover

Operations and Their
Dependencies
• No dependencies  embarrassing parallelism
– Fitness evaluation
– Each individual can be evaluated simultaneously
• Weak dependencies  synchronization points
– Crossover
– Once the parents are available the operator can be applied
• Strong dependencies  careful inspection (bottlenecks)
– Selection
– The complete population needs to be available
– The wrong implementation can introduce large serial execution
chunks

Other Perks

• Evaluation can be costly
• Some evolutionary models
– Mimic natural evolution introducing spatial relations (remember
Darwin’s islands?)
– Are model after decentralized models (cellular automata like)
• Based on the nature of evolutionary algorithms and the
above ingredients there multiple parallelization models
has been proposed (Cantu-Paz, 2000; Alba, 2005)

But?

• What about the data?


THE CHALLENGES OF DATA
MINING
Real-world examples
Wrapping up

The Challenges of Data Mining

• We have seen in the previous slides how
evolutionary algorithms have a natural tendency
for parallel processing, hence being suitable for
large-scale data mining
• However, data mining presents a challenge that
goes beyond pure optimization, which is that
evaluation is based on data, not just on a fitness
formula


• Holding the data is the first bottleneck that large-
scale data mining needs to face
– Efficiently parsing the data
– Proper data structures to achieve the minimum
memory footprint
• It may sound like just a matter of programming, but it can
make a difference
• Specially important when using specialized hardware (e.g.
CUDA)
– Optimized publicly available data handling libraries
exist (e.g. the HDF5 library)


• Usually it is not possible to hold all the training
data in memory
– Partition it and use different subsets of data at a time
• Windowing mechanisms, we will talk about them later
• Efficient strategies of use of CUDA technology
– Hold different parts of the data in different machines
• Parallel processing, we will also talk about this later
• Can also data richness become a benefit not a
problem?
– Data-intensive computing


• Preprocessing
– Lot of work in getting high-quality data
– Getting the representation right
– Both require that the data miners and the end users understand
each other
• Classic challenges of machine learning
– Over-learning
• Our models need to have good predictive capacity
– Generating interpretable solution
• Discovering useful new knowledge inside the data


KALEIDOSCOPIC LARGE
SCALE DATA MINING
Real-world examples
Wrapping up

Large Scale Data Mining Using
GBML
• Efficiency enhancement techniques
• Hardware acceleration techniques
• Parallelization models
• Data-intensive computing

Prelude: Efficiency Enhancement

• Review of methods and techniques explicitly
designed for data mining purposes
• Evolutionary computation efficiency
enhancement techniques could also be applied
(and we show some examples of this too)
• For a good tutorial on efficiency enhancement
methods, please see GECCO 2005 Tutorial on
efficiency enhancement by Kumara Sastry at
– http://www.slideshare.net/kknsastry/principled-efficiency-enhancement-techniques

Efficiency Enhancement
Techniques
• Goal: Modify the data mining methods to improve
their efficiency without special/parallel hardware
• Remember:
– An individual can be a rule, or a rule set, or a decision tree…
– Individuals parameters need to be estimated (accuracy,
generality…)
• Included in this category are:
– Windowing mechanisms
– Exploiting regularities in the data
– Fitness surrogates
– Hybrid methods

Windowing Mechanisms

• Classic machine learning concept
– Do we need to use all the training data all the time?
– Using a subset would result in faster evaluations
– How do we select this subset and how often is it changed?
– How accurate the fitness estimation will be? Will it favor
modularity?
• Freitas (2002) proposed a classification of these methods
in three types:
– Individual-wise: Changing the subset of data for each evaluated
solution
– Generation-wise: Changing the subset of data at each generation
of the evolutionary algorithm
– Run-wise: Selecting a single subset of data for a whole run of a
GA

Windowing Mechanisms - ILAS

• Incrementing Learning with Alternating Strata (Bacardit, 2004)
• Generation-wise windowing mechanism
• Training set is divided in non-overlapping strata
• Each GA iteration uses a different strata, using a round-robin
policy (evaluation speedup linearly with the number of strata)
0 Ex/n 2·Ex/n 3·Ex/n Ex

Training set

Iterations
0 Iter
• This mechanism also introduces some extra generalization
pressure, since good solutions need to survive multiple strata

Windowing Mechanisms - ILAS
• How far can we increase the
number of strata?
• Problem with ~260K instances
and 150 strata
• Knowledge learnt on different
strata does not integrate
successfully into a single
solution (if too many are used)
• We have to make sure that
each strata is a good
representation of the overall
training set
• Success model of the number r = #rules in solution, s = #strata,
of strata (Bacardit et al., 2004) p = prob. rule represented in strata,
D = size of the training set

Exploiting Regularities

• The instances in the training set do not usually cover
uniformly the search space
• Instead, there are some recurrent patterns and
regularities, that can be exploited for efficiency purposes
• (Giraldez et al., 2005) proposed a method that
precomputes the possible classifications of a rule
• As they only dealt with discrete/discretized attributes,
they generate a tree structure to efficiently know which
examples belong to each value of each attribute
• Finally the matches of a rule are the intersection of all
these subsets of examples

Exploiting Regularities in the
Data
• Other methods exploit a different regularity:
usually not all attributes are equally important
• Example: Prediction of a Bioinformatics dataset
(Bacardit and Krasnogor, 2009)
• Att Leu-2 ∈ [-0.51,7] and Glu ∈ [0.19,8] and Asp+1 ∈
[-5.01,2.67] and Met+1∈ [-3.98,10] and Pro+2 ∈
[-7,-4.02] and Pro+3 ∈ [-7,-1.89] and Trp+3 ∈ [-8,13]
and Glu+4 ∈ [0.70,5.52] and Lys+4 ∈ [-0.43,4.94] 
alpha
• Only 9 attributes out of 300 were actually in the rule

Data
• Function match (instance x, rule r)
Foreach attribute att in the domain
If att is relevant in rule r and
(x.att < r.att.lower or x.att > r.att.upper)
Return false
EndIf

EndFor
Return true

• Given the previous example of a rule, 293
iterations of this loop are wasted !!

Data
• How to exploit this phenomenon?
• Reordering the attributes in the domain from
specific to general (Butz et al., 2008)
– Afterwards, starting the match process with the most
specific one
– The most specific attributes are usually those that
make the process break. Thus, reducing usually the
number of iterations in the match loop
– Still, in the cases where a whole rule matches, the
irrelevant attributes need to be evaluated

Data
• Could we completely get rid of the irrelevant
attributes?
– The attribute list knowledge representation (ALKR)
(Bacardit, Burke and Krasnogor, 2009)
– This representation automatically identifies which are
the relevant/specific attributes for each rule
– Only tracks information about them

Data
• In ALKR two operators (specialize and
generalize) add or remove attributes from the list
with a given probability, hence exploring the rule-
wise space of the relevant attributes
• ALKR match process is more efficient, however
crossover is costlier and it has two extra
operators
• Since ALKR chromosome only contains relevant
information, the exploration process is more
efficient. On large data sets it managed to
generate better solutions

Fitness Surrogates

• In evolutionary algorithms, we can construct a
function that estimates the evaluation of our
solutions using the training set. This is usually
known as a fitness surrogate
• Two recent works (Orriols et al., 2007) and
(Llorà et al., 2007) use the structural information
extracted from the model building process of
competent genetic algorithms to build such a
function
• Cheap surrogates can help avoid costly
evaluations that tend to dominate execution time

Hybrid Methods

• The Memetic Pittsburgh Learning Classifier
Systems (MPLCS) (Bacardit and Krasnogor,
2009) combines the classic GA exploration
operators with local search (LS) methods.
– The LS operators use information extracted from the
evaluation process
– After evaluating a rule set we know
• Which rules are good and which rules are bad
• Which parts of each rule are good and which parts are bad

Hybrid Methods

• Two kinds of LS operators
– Rule set-wise operator
• Takes N parents (N can be > 2) and generates a single
offspring with the best rules of all of them
– Rule-wise operators that edit rules
• Rule cleaning – drop conditions that misclassify
• Rule splitting – find the exact spot where a rule can be
splitted and the generated rules cleaned
• Rule generalizing –update a rule so it can correctly classify
more examples
• Not only during the learning process. LS methods can
also be used for post-processing the final rule sets
(Franco et al., 2012)

Enough Talk! Where is the Big
Iron?
• Let’s start with a simple hardware acceleration example

Hardware Acceleration
Techniques
• Commodity hardware provides simple vectorized
operations
• Result of the gaming world
• Usually operate over 128 bits (4 floats)
• Vector units are able to execute ops in 1 cycle
• IBM implemented Altivec
• Intel started with MMX and then SSE and derivates
• AMD 3DNow!, 3DNow+!

A Simple Example: XCSlib

• Llora and Sastry (2005) show its usefulness. Also key to
billion bit effort by Golberg, Sastry, and Llora (2007)
• XCSlib version 0.34 (http://xcslib.sourceforge.net/)
– Based on a C++ code base
– Very flexible to modify/add new component
• The first step: Gather the facts
• Need to get a clear picture of the execution profile
– Shark freely available on Mac OS X
– Gprof on Unix systems

XCSlib
• Shark G4 platform profile (same behavior displayed on the AMD
platform)
• The rule matching is conducted by ternary_condition::match

ternary_condition::match
XCSlib
• The main cycle consumer
• Each rule loops to match
• Good candidate for HW
acceleration
• If we accelerate the inner
loop we can drop the time
spent matching

Extending Toward Vector
Instructions

Idea: Loop unroll, using vector operations to
manipulate four integers at once (pack 64
conditions in a single match step)

The Vector-based Matching
(SSE2)

Speedup After Vectorizing

Benefits of caching

Hardware Acceleration On
Steroids
• NVIDIA’s Computer Unified Device Architecture (CUDA)
is a parallel computing architecture that exploits the
capacity within NVIDIA’s Graphic Processor Units
• CUDA runs thousands of threads at the same time 
Single Program, Multiple Data paradigm
• In the last few years GPUs have been extensively used
in the evolutionary computation field
– Many papers and applications are available at
http://www.gpgpgpu.com
• The use of GPGPUs in Machine Learning involves a
greater challenge because it deals with more data but
this also means it is potentially more parallelizable

CUDA memories

• Different types of memory with different access speed
– Global memory (slow and large)
– Shared memory (block-wise; fast but quite small)
– Constant memory (very fast but very small)
• The memory is limited
• The memory copy operations involve a considerable
amount of execution time
• Since we are aiming to work with large scale datasets a
good strategy to minimize the execution time is based on
the memory usage

CUDA in supervised learning
• The match process is the stage
computationally more expensive
• However, performing only the match
inside the GPU means downloading
from the card a structure of size
O(NxM) (N=population size,
M=training set size)
• In most cases we don’t need to
know the specific matches of a
classifier, just how many (reduce
the data)
• Performing the second stage also
inside the GPU allows the system to
reduce the memory traffic to O(N)

CUDA fitness computation for
the BioHEL GBML system
• BioHEL [Bacardit,
Burke and
Krasnogor, 2009]
is a GBML method
designed for
large-scale
datasets
• We recently
extended it with a
CUDA-based
fitness
computation
(Franco,
Krasnogor &
Bacardit, 2010)

Performance of BioHEL using
CUDA
• We used CUDA in a Tesla C1060 card with 4GB of
global memory, and compared the run-time to that of
Intel Xeon E5472 3.0GHz processors

• Biggest speedups obtained in large problems (|T| or
#Att), specially in domains with continuous attributes
• Run time for the largest dataset reduced from 2 weeks
to 8 hours

Integration of CUDA and ILAS
• The speedups of these two techniques can stack one on
top of the other

Parallelization Models

• Coarse-grained parallelism
• Fine-grained parallelism

Coarse-grained Parallelism

• By coarse-grain parallelism we are talking about
executing independently several runs
• As there is no communication, the speedup is
always linear 
• In which situations can we do this?
– Evolutionary algorithms are stochastic methods, we
need to run always our methods several times. If we
have the parallel hardware, this is a trivial way of
gaining efficiency


• There is, however, a more defined way of performing
coarse-grain parallelism: ensemble learning
• These techniques integrate the collective predictions of a
set of models in some principled fashion
• These models can be trained independently


• Ensemble for consensus prediction (Bacardit and
Krasnogor, 2008)
– Similar technique to bagging
1. Evolutionary data mining method is run N times on the original
training set, each of them with a different random seed
2. From each of the N runs, a rule set is generated
3. Exploitation stage: For each new instance, the N models
produce a prediction. The majority class is used as the
ensemble prediction
– Ensembles evaluated on 25 UCI repository datasets
using the Gassist LCS
– In average the ensemble accuracy was 2.6% higher


• Ensemble for consensus prediction
– Prediction of a difficult bioinformatics dataset
– Accuracy increased of ~9% with 25 rule sets

More Corse-Grain

• If evaluation is extremely costly
• Run the same algorithm with the same seed
• Same population everywhere
• Each algorithm only evaluates a chunk of the population
• The fitness estimates are broadcasted (e.g. MPI)
• Minimal communication possible (only the fitness value)
• All algorithms run the same genetic operators on
identical population individuals (as all run using the same
seed)

• The NAX system (Llora, X., Priya, A., and Bhragava,
2007)

In a Picture
Processor 0 Processor 1 Processor p

Initialize Initialize Initialize

Evaluate Chunk Evaluate Chunk Evaluate Chunk

Synchronize Synchronize Synchronize

Selection Selection Selection

Recombination Recombination Recombination

Replacement Replacement Replacement

Done? Done? Done?

Finalize Finalize Finalize

Fine-grained parallelism

• Exploit maximum parallelism
• Impose an spatial topology
• Define neighborhood operators
• GALE (Llora, 2002)
• Easy implementable on shared-memory machines
• Minimizes the computation/communication ratio for
distributed memory implementations

GALE: Topology

• A cell contains 0 or 1 individual
• A cell is surrounded by 8 neighbors
• Subpopulations spatially defined by the
Empty cell Occupied cell (1 ind) adjacent cells

GALE: Merge

• Merge
mate

1. Choose a neighbor

4. Recombine the genetic material

8. Replace the individual

GALE: Split

• Split

1. Replicate and mutate

4. Occupy
– Empty cell with higher
number of neighbors

– Occupied cell with the worst
neighbor (no empty cell
available)

GALE: Survival

• Isolated
• 0-1 Neighbors • psr(ind) fitness proportional
• death  leave cell empty

• Spongy
• 2-6 Neighbors • psr(ind) related to neighbors
• death  leave cell empty

• Crowded
• 7-8 Neighbors • psr(ind) = 0
• death  replace by the best

Data-intensive Computing
• Usually refers to:
– Infrastructure
– Programming techniques/paradigms
• Google made it main stream after their MapReduce model
• Yahoo! provides and open source implementation
– Hadoop (MapReduce)
– HDFS (Hadoop distributed filesystem)
– Mahout (Machine Learning methods)
• Engineered to store petabytes reliably on commodity hardware (fault
tolerant)
• Map: Equivalent to the map operation on functional programming
• Reduce: The reduction phase after maps are computed

Meandre: NCSA’s
Data-Intensive Infrastructure
• Extend the programming limitation of
MapReduce
• Execution Paradigms
– Conventional programs perform computational tasks
by executing a sequence of instructions.
– Data driven execution revolves around the idea of
applying transformation operations to a flow or stream
of data when it is available.

Meandre: The Dataflow
Component
• Data dictates component execution semantics

Inputs Outputs

Component

P

Descriptor in RDF The component
of its behavior implementation

Meandre: Flow (Complex Tasks)
• A flow is a collection of connected components

Read
P Merge
P

Get Show
P
P

Do
P

Dataflow execution

Your Point Being?

• Evolutionary algorithms can be modeled using data-
intensive modeling
• Imagine a stream of individuals being process by
components
• A single model implementation automatically
parallelizable where needed


Real-World Examples
Wrapping up

Real-World Examples

• Example to present
– Protein Structure & Contact Map Prediction (Bacardit et al.,
2009)
– Uncovering new regulators in seed germination (Bassel et al.,
2011)

• A set of LCS applications to Data Mining is collected in
(Bull et al., 2008)

Protein Structure Prediction

• Protein Structure Prediction (PSP) aims to
predict the 3D structure of a protein based on its
primary sequence

3D
Primary Structure
Sequence

Protein Structure Prediction

• Beside the overall 3D PSP (an optimization problem),
several structural aspects can be predicted for each
protein residue
– Coordination number
– Solvent accessibility
– Etc.
• These problems can be modelled in may ways:
– Regression or classification problems
– Low/high number of classes
– Balanced/unbalanced classes
– Adjustable number of attributes
• Ideal benchmarks
– http://icos.cs.nott.ac.uk/datasets/psp_benchmark.html

Contact Map Prediction

• Two residues of a chain are said to be in contact if their
distance is less than a certain threshold
Primary Contact Native State
Sequence

• Contact Map (CM): binary matrix that contains a 1 for a cell if
the residues at the row & column are in contact, 0 otherwise
• This matrix is very sparse, in real proteins there are less than
2% of contacts
• Highly unbalanced dataset

Contact Map Prediction
• (Bacardit et al. 2009) participated in the CASP8 competition
• CASP = Critical Assessment of Techniques for Protein Structure
Prediction. Biannual competition
• Every day, for about three months, the organizers release some
protein sequences for which nobody knows the structure (129
sequences were released in CASP9, in 2010)
• Each prediction group is given three weeks to return their
predictions
• If the machinery is not well oiled, it is not feasible to participate !!
• For CM, prediction groups have to return a list of predicted contacts
(they are not interested in non-contacts) and, for each predicted pair
of contacting residues, a confidence level
• The evaluation for CM ranks this list by the confidence, and
calculates the accuracy of the top L/x predictions (L = length of
chain, x = typically 10)

Contact Map Prediction: Hands on

Training • Training set of 2413 proteins selected to
set represent a broad set of sequences
x50 – 32 million pairs of amino-acids (instances in the
Samples
training set) with less than 2% of real contacts
– Each instance is characterized by up to 631 attributes

x25 • 50 samples of ~660000 examples are
Rule sets generated from the training set. Each sample
contains two no-contact instances for each
contact instance
• The BioHEL GBML method (Bacardit et al.,
2009) was run 25 times on each sample
Consensus • An ensemble of 1250 rule sets (50 samples x
25 seeds) performs the contact maps
predictions using simple consensus voting
• Confidence is computed based on the votes
Predictions
distribution in the ensemble

Results of Contact Map
prediction
• The subset of the most difficult target (Free Modelling
targets) of CASP9 were used to evaluate CM
• Out predictor obtained an average accuracy of 23.6%
• Do you think it is low?
– It is more than 11 times higher than a random prediction
– The predictor was the best sequence-based method in the 2010
competition
• Overall, tackling this problem has forced us to address a
broad range of bottlenecks in DM methods
– Code bottlenecks
– Memory footprint bottlenecks
– Scalability bottlenecks

Functional Network Reconstruction
for seed germination

 Microarray data obtained from seed tissue of
Arabidopsis Thaliana
 122 samples represented by the expression level of
almost 14000 genes
 It had been experimentally determined whether each of
the seeds had germinated or not
 Can we learn to predict germination/dormancy from the
microarray data?
 [Bassel et al., 2011]

Generating rule sets

 BioHEL was able to predict the
outcome of the samples with 93.5%
accuracy (10 x 10-fold cross-
validation
 Learning from a scrambled dataset
(labels randomly assigned to
samples) produced ~50% accuracy

If At1g27595>100.87 and At3g49000>68.13 and At2g40475>55.96  Predict
germination
germination
germination
If At5g54910>45.03 and At4g18975>16.74 and At3g28910>52.76 and
At1g48320>56.80  Predict germination
Everything else  Predict dormancy

Identifying regulators

 Rule building process is stochastic
 Generates different rule sets each time the
system is run
 But if we run the system many times, we
can see some patterns in the rule sets
 Genes appearing quite more frequent than
the rest
 Some associated to dormancy
 Some associated to germination

Known regulators appear with high
frequency in the rules

Generating co-prediction networks
of interactions
• For each of the rules shown before to
be true, all of the conditions in it need
to be true at the same time
– Each rule is expressing an interaction
between certain gens
• From a high number of rule sets we
can identify pairs of genes that co-
occur with high frequency and
generate functional networks
• The network shows different topology
when compared to other type of
network construction methods (e.g. by
gene co-expression)
• Different regions in the network contain
the germination and dormancy genes

Experimental validation
 We have experimentally verified this analysis
 By ordering and planting knockouts for the highly
ranked genes
 We have been able to identify four new regulators of
germination, with different phenotype from the wild
type

Real-world examples

Wrapping Up

Wrapping Up

• We have shown in this tutorial how GBML methods have
high potential for mining large-scale datasets
• They are natural parallel processing machines
• Recent improvements in many dimensions of the
learning process
– Representations
– Learning paradigms
– Inference mechanisms
– Hybridization

The Game Has a New Name

• The exception is becoming norm
– Efficient parallel designs
– Efficiency enhancement methods
– Hardware support (SSE, CUDA, etc.)

• However, all these components cannot be used
blindly, they have to be adjusted properly,
accordingly to the characteristics/dimensions of
the problem

Better Understanding

• Theoretical analysis of the different facets of a GBML
system can help
• Understand better why/when can the components
perform well
• Design robust policies that can take the best of the
techniques at hand
• Provide insight on parameterization of methods
– If we would like the community to use GBML methods, we have
to make them easy to use
• Some work already exists (Butz, 2006; Franco et al.,
2011), but we still have a long road ahead of us

Do not Be Shy

• GBML systems are highly flexible, with good
explanatory power, and can have good
scalability
• Go and give it a shoot!

References
• http://www.ncbi.nlm.nih.gov/Genbank/index.html
• http://www.netflixprize.com/
• V. Reinke, Germline genomics (January 20, 2006), WormBook, ed. The C. elegans
Research Community, WormBook, doi/10.1895/wormbook.1.74.1,
http://www.wormbook.org
• Bernadó, E., Ho, T.K., Domain of Competence of XCS Classifier System in
Complexity Measurement Space, IEEE Transactions on Evolutionary Computation, 9:
82-104, 2005.
• “Physicists brace themselves for lhc ’data avalanche’.”
www.nature.com/news/2008/080722/full/news.2008.967.html
• M. Pop and S. L. Salzberg, “Bioinformatics challenges of new sequencing
technology,” Trends in Genetics, vol. 24, no. 3, pp. 142 – 149, 2008
• http://www.hdfgroup.org/HDF5
• K. Sastry, “Principled Efficiency-Enhancement Techniques”, GECCO-2005 Tutorial
• A.A. Freitas, “Data Mining and Knowledge Discovery with Evolutionary Algorithms”,
Springer-Verlag, 2002
• J. Bacardit, Pittsburgh Genetics-Based Machine Learning in the Data Mining era:
Representations, generalization, and run-time. PhD thesis, Ramon Llull University,
Barcelona, Spain, 2004

References
• Jaume Bacardit, David E. Goldberg, Martin V. Butz, Xavier Llorà and Josep M.
Garrell, Speeding-up Pittsburgh Learning Classifier Systems: Modeling Time and
Accuracy, 8th International Conference on Parallel Problem Solving from Nature -
PPSN VIII
• D. Song, M.I. Heywood and A.N. Zincir-Heywood, Training genetic programming on
half a million patterns: an example from anomaly detection, IEEE Transactions on
Evolutionary Computation, vol. 9, no. 3, pp 225-239, 2005
• Llora, X., Priya, A., and Bhragava, R. (2007), Observer-Invariant Histopathology using
Genetics-Based Machine Learning. Proceedings of the Genetic and Evolutionary
Computation Conference (GECCO 2007) , 2098–2105
• Giráldez R, Aguilar-Ruiz JS, Santos JCR (2005) Knowledge-based fast evaluation for
evolutionary learning. IEEE Transactions on Systems, Man, and Cybernetics, Part C
35(2):254–261
• J. Bacardit, E. K. Burke, and N. Krasnogor. Improving the scalability of rule-based
evolutionary learning. Memetic Computing, 1(1):55-67, 2009
• M. V. Butz, P. L. Lanzi, X. Llorà, and D. Loiacono. An analysis of matching in learning
classifier systems.In GECCO ’08: Proceedings of the 10th annual conference on
Genetic and evolutionary computation, pp. 1349–1356. ACM, 2008.
• Llorà, X., Sastry, K., Yu, T., and Goldberg, D. E. Do not match, inherit: fitness
surrogates for genetics-based machine learning techniques. In Proceedings of the 9th
Annual Conference on Genetic and Evolutionary Computation, pp 1798-1805, ACM,
2007

References
• Orriols-Puig, A., Bernadó-Mansilla, E., Sastry, K., and Goldberg, D. E. Substructrual
surrogates for learning decomposable classification problems: implementation and
first results. 10th International Workshop on Learning Classifier Systems, 2007
• J. Bacardit and N. Krasnogor, Performance and Efficiency of Memetic Pittsburgh
Learning Classifier Systems, Evolutionary Computation Journal, 17(3):307-342, 2009
• G. Wilson and W. Banzhaf, "Linear genetic programming gpgpu on microsoft’s xbox
360," in Proceedings of the 2008 Congress on Evolutionary Computation, pp.
378-385. IEEE Press, 2008
• http://www.gpgpgpu.com/
• J. Bacardit and N. Krasnogor. “Empirical evaluation of ensemble techniques for a
Pittsburgh Learning Classifier System”. Learning Classifier Systems. LNAI 4998, pp.
255-268, 2008, Springer
• http://www.infobiotic.net/PSPbenchmarks/
• J. Bacardit, M. Stout, J.D. Hirst, K. Sastry, X. Llorà and N. Krasnogor. Automated
Alphabet Reduction Method with Evolutionary Algorithms for Protein Structure
Prediction In Proceedings of the 9th Annual Conference on Genetic and Evolutionary
Computation (GECCO2007), pp. 346-353, ACM Press, 2007
• Goldberg, D. E., Sastry, K. and Llora, X. (2007), Toward routine billion-variable
optimization using genetic algorithms. Complexity , 12(3), 27–29.

References
• G. Venturini. SIA: A supervised inductive algorithm with genetic search for learning
attributesbased concepts. In: Brazdil PB (ed) Machine Learning: ECML-93 - Proc. of
theEuropean Conference on Machine Learning, Springer-Verlag, Berlin, Heidelberg, pp
280–296, 1993
• J. Rissanen J. Modeling by shortest data description. Automatica vol. 14:465–471, 1978
• L. Bull, E. Bernadó-Mansilla and J. Holmes (editors), Learning Classifier Systems in
Data Mining. Springer, 2008
• Alba, E., Ed. Parallel Metaheuristics. Wiley, 2007.
• Cantu-Paz, E. Efficient and Accurate Parallel Genetic Algorithms. Springer, 2000.
• Llora, X. E2K: evolution to knowledge. SIGEVOlution 1, 3 (2006), 10–17.
• Llora, X. Genetic Based Machine Learning using Fine-grained Parallelism for Data
Mining. PhD thesis, Enginyeria i Arquitectura La Salle. Ramon Llull University,
Barcelona, February, 2002.RFC2413, The Dublin Core Metadata Initiative, 2008.
• Llora, X., Acs, B., Auvil, L., Capitanu, B., Welge, M., and Goldberg, D. E. Meandre:
Semantic-driven data-intensive ﬂows in the clouds. In Proceedings of the 4th IEEE
International Conference on e-Science (2008), IEEE press, pp. 238–245.
• M. Butz, Rule-Based Evolutionary Online Learning Systems: A Principled Approach to
LCS Analysis and Design, Studies in Fuzziness and Soft Computing, vol 109. Springe,r,
2006

References
• Hadoop (http://hadoop.apache.org/core/)
• Meandre (http://seasr.org/meandre)
• Dean, J. & Ghemawat, S. MapReduce: Simplified Data Processing in Large Clusters,
OSDI'04: Sixth Symposium on Operating System Design and Implementation,San
Francisco, CA, December, 2004.
• Franco, M. A., Krasnogor, N., and Bacardit, J. Speeding up the evaluation of
evolutionary learning systems using GPGPUs. In Proceedings of the 12th Annual
Conference on Genetic and Evolutionary Computation - GECCO '10. pp 1039-1046,
2010
• A. Verma, X. Llora, S. Venkataram, D.E. Goldberg and R. Campbell, Scaling eCGA
Model Building via Data Intensive Computing, In Proceedings of WCCI2010, IEEE
press, pp 1-8, 2010
• J. Bacardit, M. Stout, J.D. Hirst, A. Valencia, R.E. Smith and N. Krasnogor. Automated
Alphabet Reduction for Protein Datasets. BMC Bioinformatics 10:6, 2009
• M. Franco, N. Krasnogor and J. Bacardit. Modelling the Initialisation Stage of the ALKR
representation for Discrete Domains and GABIL encoding, In Proceedings of the 13th
Annual Conference on Genetic and Evolutionary Computation - GECCO ’11, pp.
1291-1298. ACM, 2011
• M. Franco, N. Krasnogor and J. Bacardit. Post-processing operators for decision lists, ,
In Proceedings of the 14th Annual Conference on Genetic and Evolutionary
Computation - GECCO ’12, pp. 1291-1298. ACM, 2012

Large Scale Data Mining using
Genetics-Based Machine Learning
Jaume Bacardit Xavier Llorà
School of Computer Science & Google Inc. 1600
School of Bioscience Amphitheatre Pkwy
University Nottingham Mountain View, CA 94043
Nottingham, UK

jqb@cs.nott.ac.uk xllora@gmail.com

GECCO 2012 Tutorial, July 7th 2012

Large Scale Data Mining using Genetics-Based Machine Learning

More Related Content

Viewers also liked (14)

Similar to Large Scale Data Mining using Genetics-Based Machine Learning (20)

More from jaumebp (6)

Recently uploaded (20)

Large Scale Data Mining using Genetics-Based Machine Learning

Editor's Notes