Distributed Decision Tree Induction

A Review of Distributed Decision
Tree Induction

Gregory Gay
West Virginia University
greg@greggay.com

Data Mining 101

● Data mining = finding
patterns in heaps of
data
● Classification = Given
data and a series of
possible labels,
gather evidence and
assign a label to each
piece of data.
2

Decision Tree Induction
● These algorithms greedily and recursively
select the attribute that provides the most
information gain.
● This attribute is used to partition the data set
into subsets around a particular value.
● Stops when the leaf node in the tree only
contains data with the proper class label.

3

Splitting the Data
● Split data based on the (attribute,value) pair
that maximizes:
●

● I(T) is an impurity measure. This is usually
entropy(C4.5) or gini(SMP):

4

Climbing Mount Everest

● The success of data
mining is its downfall.
● Fast, accurate results
expected on huge
data sets.
● Parallelism of data
and/or the algorithm.

5

Paradigm A: Data Parallelism
● Given the large size of these data sets,
gathering all of the data in a centralized location
or in a single file is neither desirable nor always
feasible because of the bandwidth and storage
requirements.
● Also, data may be fragmented in order to
address privacy and security constraints.
● In these cases, there is a need for algorithms
that can learn from fragmented data.
● Two methods – horizontal, vertical
6

Horizontal Fragmentation
● The entries are equally divided into a number of
subsets equal to the number of different storage
sites.
● The learner uses frequency counts to find the
attribute that yields the most information gain to
further partition the set of examples.
● Given: |E|= # examples in the data set T , |A|=
#attributes, V= max(values per attribute),
M=#sites, N=#classes, size(D)=#nodes in tree D
● Time complexity = |E||A|size(D)
7
● Communication complexity = N|A|VMsize(D)

Vertical Fragmentation
● Individual attributes or subsets of attributes,
along with a list of (value, identifier) pairs for
each are distributed to sites.
● Can suffer from load imbalance and poor
scalability.
● Given: |E|= # examples in the data set T , |A|=
#attributes, V= max(values per attribute),
M=#sites, N=#classes, size(D)=#nodes in tree
D
● Time complexity = |E||A|Msize(D)
8
● Comm. complexity = (|E|+N|A|V)M size(D)

When Distributed Data Wins
● It would be possible to collect those fragments
and reassemble them in a centralized location.
So, why we should leave the data at distributed
sites?
● Privacy or security - learner must perform solely
on statistical summaries.
● Distributed versions of algorithms compare
favorably with the corresponding centralized
technique whenever its communication cost is
less than the cost of collecting all of the data in
one place. 9

Paradigm B: Task Parallelism
● The construction of decision trees in parallel.
● Single process begins the construction process.
When the #child nodes = #available
processors, the nodes are split among them.
Each processor then proceeds with the tree
construction algorithm.
● Implementing parallel algorithms for decision
tree classification is a difficult task for multiple
reasons.

10

Paradigm C (B.5?): Hybrid
Parallelism
● Most parallel algorithm don’t practice strict task
parallelism
● Instead – parallelized algorithm over split data.
● Some (parallelized C4.5), switch between
data/task parallelism when communication cost
is too high.
● Others always operate in both modes (SMP,
INDUS).

11

Algorithm: C4.5 Parallel
● Parallel implementation of Ross Quinlan’s C4.5
decision tree algorithm.
● Data parallelism at the beginning, task
parallelism at the lower nodes of the tree.
● Horizontal fragmentation scheme. Uniform load
balance is achieved by equally distributing data
among the processors and using a breadth-first
strategy to build the actual decision tree. When
a process finished exploring it nodes, it sends a
request for more nodes.
12

● Each processor is responsible for building its
own attribute lists and class list.
● The continuous attribute lists are globally
sorted. Each processor has a list of sorted
values.
● Before evaluating the possible split points in
their entries, the distributions must reflect the
data entries assigned to the other processors.
– Discrete – store the data everywhere
– Continuous - the gain calculated based on the
distributions before and after the split point.
● After evaluating all of the local data, the
processors communicate among themselves in
order to find the best split. 13

● Horizontal data fragmentation as long as the
#examples covered by the nodes > a pre-
defined threshold (when communication cost
for building a node > cost of building it locally
plus the cost of moving the set of examples
between processors)
● When #data entries < threshold, they are
evenly distributed among the processors.

14

Algorithm: SMP Tree Classifier
● SMP clusters = shared-memory nodes with a
small number of processors (usually 2-8)
connected together with a high-speed
interconnect.
● Two-tiered architecture where a combination of
shared-memory and distributed-memory
algorithms can be deployed.
● The training dataset is horizontally fragmented
across the SMP nodes so each node carries
out tree construction using an equal subset of
the overall data set. 15

● Attribute lists are dynamically scheduled to take
advantage of the parallel light-weight threads
● One of these threads is designated as a master
thread. Is responsible for processing the
attribute lists as well as exchanging data and
information between SMP nodes.
● The numerical attribute lists are globally sorted
such that the first node has the first portion of
the data set
● When a tree node is split into its children, the
local attribute lists are partitioned among those
children. Thus, no communication costs are
incurred due to the attribute lists as the tree is
16
grown in each SMP node.

● Finding global best split causes synchronization
● The algorithm will proceed in a breadth-first
manner, processing all of the leaf nodes before
processing any of the new child nodes.
● Before finding the new best split point, attribute
lists inserted into a queue shared by all of the
threads running on the SMP node. This queue
is used to schedule attribute lists to threads.
● Each leaf node calculates and compares its
split points with all of those along the leaf nodes
● SMP node 0 can compute the overall best split
point for all attributes. It will broadcast this point
to all of the SMP nodes.
17

INDUS
● INDUS is a multi-agent system used for
knowledge acquisition from data sources
● Can deal with horizontal or vertical data, only
needs statistical summaries of the data.
● Designed to provide a unified query interface
over a set of distributed, heterogeneous, and
autonomous data sources as if the dataset
were a table.
● These queries are driven by an internal tree
learner that uses them whenever a new node
needs to be added to the tree. 18

Summary
● We want to classify data. Decision tree
classifiers are a good way to do it.
● We need to classify BIG data sets.
● We can improve performance by parallelizing
the data (horizontally or vertically) or the
algorithm. Often both.
● We looked at a few algorithms to do this.
● They all perform well, show a promising
improvement over standard methods.
19

Now you can climb the mountain!

20

Questions? Comments?
● Paper: http://greggay.com/pdf/dsys1.pdf
● greg@greggay.com
● http://twitter.com/Greg4cr
● http://facebook.com/greg.gay

21

Distributed Decision Tree Induction

More Related Content

What's hot

Similar to Distributed Decision Tree Induction

More from gregoryg

Recently uploaded

Distributed Decision Tree Induction