Analysis on different Data mining Techniques and algorithms used in IOT

Shweta Bhatia Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 5, Issue 11, (Part - 1) November 2015, pp.82-85
www.ijera.com 82 |
P a g e
Analysis on different Data mining Techniques and algorithms
used in IOT
Shweta Bhatia1
, Sweety Patel2
1 Assistant Professor, Shri Ramkrishna Institute of Computer Education &Applied Sciences, Sarvajanik
Education Society, Athwagate, Surat, India
2 Assistant Professor, S.V.Patel College of Management and Computer Application, sumul dairy raod, surat.
Abstract
In this paper, we discusses about five functionalities of data mining in IOT that affects the performance and that
are: Data anomaly detection, Data clustering, Data classification, feature selection, time series prediction. Some
important algorithm has also been reviewed here of each functionalities that show advantages and limitations as
well as some new algorithm that are in research direction. Here we had represent knowledge view of data
mining in IOT.
I. Introduction
The Internet of Things (IOT) and its related
technologies can seamlessly integrate classical
networks with network instruments and devices. The
data in the Internet of Things can be categorized into
several types: RFID data stream, address identifiers,
descriptive data, positional data, environment data
and sensor network data etc. [1]. Today, IOT brings
the great challenges for managing, analysing and
mining data. In IOT systems, data quality
management is a critical technology to provide high-
quality and trusted data to business-level analysis,
optimization and decision making. In order to
improve quality of data, anomaly detection
techniques are widely used to remove noises and
inaccurate data. For anomaly detection, having more
data means it’s easier to detect an unusual event
against the background of normal events [3].
Data Clustering refers to grouping of data based
on specific features and its value. In IOT, Data
clustering is an intermediate step for identifying
patterns from the collected data. It’s most common
process in unsupervised machine learning. Clustering
methods are divided into 4 major categories such as:
partitioning methods, hierarchical methods, density-
based methods and grid based methods. Other
clustering techniques also exist such as: fuzzy
clustering, artificial neural network and generic
algorithms.
The problem of Data classification is stated as:
given a set of training data points along with
associated label for an unlabelled test instances.
Classification algorithm contain 2 phases: Training
phase and Testing phase. On the basis of training data
set, segmentation is done which encodes knowledge
about the structure of the groups in form of target
variable. Thus classification problem is referred to as
supervised learning.
The feature selection is the process used to
recognized pattern and allows you to identify
attributes that affect quality index the most. After
some initial level of experiment feature selection is
preferable, identify what are attributes that affects a
specific problem most and then perform data
classification, time series prediction or anomaly
detection more easily as it reduce the dimensionality
in mining the problem. Features selection is to find a
satisfactory feature subset from the candidate feature
set, so that to reach an optimal classification accuracy
and computing complexity control.
A time series is collection of temporal data
objects, which includes characteristics such as: large
data size, high dimensionality, and updating
continuously. Representation, similarity measures
and indexing are 3 components of time series task
relies on. Time series representation reduces the
dimension and it divides into 3 categories: model
based representation, non-adaptive data
representation and adaptive data representation. The
similarity measure is carried out in proper manner
such as: research directions include subsequence
matching and full subsequence matching. The
indexing of time series is linked with representation
and similar measure tools [2].
II. Anomaly detection algorithms
Anomaly detection algorithm could be either
global or local. Role of different algorithms that can
be used for IOT with data anomaly are nearest
neighbour-based anomaly detection, clustering based
anomaly detection, statistical anomaly detection, and
spectral anomaly detection.
NN based algorithms assign the anomaly score
of data occurrences relative to their neighborhood.
Nearest-neighbor (NN) based anomaly detection are
broadly used in areas such as: for finding similar
RESEARCH ARTICLE OPEN ACCESS

www.ijera.com 83 | P a g e
patches in images, HTM based applications for IT
analytics, wireless sensor networks, etc.
NN based algorithms are: 1. k-NN Global
Anomaly Score 2. Local Outlier Factor (LOF) 3.
Connectivity based Outlier Factor (COF) 4. Local
Outlier Probability (LoOP) 5. Influenced Outlierness
(INFLO) 6. Local Correlation Integral (LOCI) [5].
1. k-NN Global anomaly score: Score is the distance
to the k-th neighbor and Score is the average distance
of k neighbors. k-NN is simplest method for
classification used in data mining that indirectly
associate with IOT.
2. LOF: Local Outlier Factor: Most prominent AD
algorithm and is able to find local anomalies. Efforts
is O(n2
). For example: compute the local density:
LRDmin(p) = 1/ (∑ oԑNmin (p) reach_dist min (p,o) / |
Nmin(P) | )
Statistical anomaly detection is quickly
becoming a required capability in new world of the
IOT. This is traceable for many special cases such as:
continuous signal, discrete event timing, and user log
files. Statistical anomaly detection is that you encode
the patterns of what is normal as a probabilistic
model. From benefit point of view, probabilistic
models come with built-in measure of anomaly in
terms of determining what is anomalous and this
models come with a way of learning from observed
data that is called a training algorithm. For instance: a
rate model for web traffic anomaly detector can do a
very good at predicting traffic during late December,
even though it was trained only on last week of
November and first week of December [9].
The key property of probabilistic is the constraint that
the probability of all possible things has to sum to
one. On basis of this constraint, training algorithm
model concentrating probability around what is
normal and thus making the modelled probability of
anomalies lower.
Spectral techniques attempt to find an
approximate of data using combination of attributes
that capture bulk of variations in data. This technique
automatically perform dimensionality reduction and
suitable for handling high dimensional data sets. This
can be used with unsupervised setting. The
disadvantage of techniques are useful only if the
normal and anomalous instances are separable in
lower dimensional embedding of the data and it
contains high computational complexity.
III. Data Clustering, Classification and
Feature Selection Algorithm
For real time IOT data stream, fast density based
clustering is required which includes density based
data stream clustering. This clustering grouped as
density-grid based method and density based
microclustering method. The main advantage of
density-grid approach is its fast processing time that
is independent of the number of data points and
dependent only on number of cells. On other hand,
density based microclustering method keep summary
of clusters in microclusters and form a final clusters
from them. By using advantage of both method an
author in paper [4], proposed real experiment with
HDC (hybrid density-based clustering for data
stream)-stream algorithm. HDC-Stream only searches
in potential list and if it cannot find the suitable
microcluster, the data point is mapped to the grid,
which keeps the outlier buffer. In future, author will
focus on distributed HDC-stream density-based
clustering to improve performance in IOT.
Classification widely used algorithms for IOT
while mining a data on internet are: 1. C4.5 or
Decision tree, 2. k-nearest neighbour algorithm, 3.
support vector machine, 4. the apriori algorithm, 5.
AdaBoost algorithm. These classification algorithm
can be implemented on different types of data sets
and on basis of performance these algorithm also
used to detect the natural disasters like cloud
bursting, earth quake, etc.
1. On the basis of feature values, decision tree
classifies instances. For a given set S cases, C4.5 first
grows an initial A. tree using divide-and-conquer
algorithm as follow: tree is leaf labelled if all cases
belongs to same class S or S is small. B. Otherwise,
select test based on single attribute.
Some limitation this algorithm pertains are: Empty
branches, insignificant branches and over fitting.
Most decision tree algorithm cannot perform well
with problem that require diagonal partitioning.
2. The most common role of data mining is to find
frequent itemsets from transaction datasets and derive
association rules. Once itemsets are obtained, it’s
upfront to generate association rules. To achieve this
Apriori algorithm is helpful. This algorithm is
assumes that items within transaction or itemsets are
sorted in lexicographic order. The Apriori algorithm
generally perform in 2steps join and prune step. It
then calculates frequency only for those candidates
generated by scanning the database.
As growing pressure for classification of data in
urgency situation: data classification for breach
response, for e-discovery, for business unity as
moving towards cloud [8].

There are general 3 classes of feature selection
algorithms: 1. Filter methods, 2. Wrapper methods
and 3. Embedded methods [12].
1. To assign a scoring to each feature, filter selection
method apply a statistical measure. The attribute are
ranked by the score and either selected to be kept or
removed from the dataset. Examples of some filter
method are: chi squared test, information gain and
correlation coefficient scores.
2. Wrapper methods consider the selection of set of
attributes as a search problem. Score is assigned
based on model accuracy where combination of
features get evaluated. The search process may use
different methods such as best-first search, random
hill-climbing algorithm or heuristic. Example of
wrapper method is recursive feature elimination
algorithm.
3. When model is created, embedded methods learn
which features best contribute to the accuracy of the
model. Regularization method is most common that
introduce additional constraints into optimization of a
predictive algorithm that prejudice the model towards
lower complexity. Example of this method are:
LASSO, Elastic Net and Ridge Regression [7].
IV. Time Series Analysis and Need of
IOT, challenges and conclusion.
Deep learning algorithms could apply to IOT and
Smart city domains in time series analysis. The new
model has been proposed with temporal patterns in
deep learning namely, the Recursive Convolutional
Bayesian
Model (RCBM), which is capable of addressing
2 tasks: identification of multi-scale signatures and
mining of compositional interactions [10].By
building a layered structure of signature detectors,
where each layer is responsible for a specific time
scale, is major idea behind RCBM.
The Trendalyze Decision is also dedicated in
developing and providing an analytic platform for
search, visual discovery, and operational monitoring
of frequently occurring patterns in time series data
streams generated by IOT.
Examples of time series applications include:
capacity planning, inventory replenishment, sales
forecasting and future staffing levels.
To address the need for connecting large number
of IOT to application infrastructure a new approach is
presented called SDP developed by cloud and its
public domain project available for free.
Further challenges in IOT: 1. The most
practical applications are happening in Industrial IOT
(IIOT) are nearly limitless such as : smarter and more
efficient factories, greener energy generation, self-
regulating buildings that optimize energy
consumption, cities that adjust traffic patterns to
respond to congestion, etc. implementation will be a
challenge. 2. Security is playing vital role at all
layers in IOT on the devices. There is no threat
detection can mitigate effectively. Major challenges
in security are: ubiquitous data collection, potential
for unexpected uses of consumer data and heightened
security risks. 3. IOT is not only who owns the data
but who controls and receive access to that data.
From a consumer perspective this will be a major
challenge. 4. Shared standards and infrastructure is
complex part of IOT. Structure of hardware, sensors,
applications and devices that need to be able to
communicate geographically and across verticals. A
largest players in the market is working on
developing such standards, AT&T, CISCO, IBM and
Intel and GE are on the way to improve integration of
physical and digital world.
V. Conclusion
This paper has focused a some algorithm on all
techniques of data mining that can be applied on IOT
with their advantages and disadvantages. We had also
covered some future challenges and the integrated
approach required to fulfil needs of IOT in present
era.
REFERENCES
[1] Shen Bin , Liu Yuan* , Wang Xiaoyi*,
Ningbo Institute of Technology, Zhejiang
University Ningbo, China, College of
Management, Zhejiang University
Hangzhou, China on “Research on Data

Mining Models for the Internet of Things”
in IEEE 2010.
[2] Joshua Cooper and Anne James on
“Challenges for Database Management in
the Internet of Things” in IETE
TECHNICAL REVIEW, Researchgate.net
SEPTEMBER 2009.
[3] Feng Chen , Pan Deng, Jiafu Wan, Daqiang
Zhang,Athanasios V. Vasilakos and Xiaohui
Rong on “Data Mining for the Internet of
Things: Literature Review and Challenges”.
[4] Amineh Amini, Hadi Saboohi, Teh Ying
Wah, and Tutut Herawan on “A Fast
Density-Based Clustering Algorithm for
Real-Time Internet of Things Stream” by
Hindawi in 2014.
[5] Mennatallah Amer and Markus Goldstein on
“Nearest-Neighbor and Clustering based
Anomaly Detection Algorithms for
RapidMiner” in Researchgate.net 2012
[6] Enzo Busseti, Ian Osband, Scott Wong on
“Deep Learning for Time Series Modeling
CS 229 Final Project Report” in 2012.
[7] Pawan Gupta, Susheel Jain, Anurag Jain on
“A Review Of Fast Clustering-Based
Feature Subset Selection Algorithm” in nov-
2014.
[8] Jay Cline on “Growing pressure for data
classification” in 2007.
[9] Numenta on “The Science of Anomaly
Detection” white paper 2015.
[10] Huan-Kai Peng*, Radu Marculescu on
“Multi-Scale Compositionality: Identifying
the Compositional Structures of Social
Dynamics Using Deep Learning” in April
2015.
[11] Rosaria Silipo on “Data mining and
Predictive Analysis” in October 2014.
[12] Jason Brownlee on “An introduction to
feature selection” in 2014.

Analysis on different Data mining Techniques and algorithms used in IOT

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Analysis on different Data mining Techniques and algorithms used in IOT (20)

Recently uploaded (20)

Analysis on different Data mining Techniques and algorithms used in IOT