IRJET- Sampling Selection Strategy for Large Scale Deduplication of Synthetic and Real Datasets using Apache Spark

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 04 | Apr-2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 485
Sampling Selection Strategy for Large Scale Deduplication of
synthetic and real datasets using Apache Spark
Gaurav Kumar1, Kharwad Rupesh2, Md.Shahid Equabal3, N Rajesh4
1,2,3,4 Dept. of Information Science and Engineering, The National Institute of Engineering, Mysuru,
Karnataka, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract -Due to the enormous increaseinthegenerationof
information by a number of sources, the requirement of
several new applications has become mandatory. These
applications may be media streaming, digital libraries etc.
Data quality is degraded due to the presence of duplicate
pairs. This is a very serious issue regarding the quality and
authenticity of data.
Therefore, data deduplication task is necessary to be
performed with the datsets. It detects and removes duplicates
and provide efficient solutions to this problem. In very large
datasets, it is very difficult to produce the labeled set from the
information provided by the user as compared to the small
datasets.
Guilherme Dal Bianco, Renata Galante, Marcos Andre
Goncalves, Sergio Canuto, and Carlos A. Heuser proposed a
Two-stage sampling selection strategy (T3S) [17]thatselects
a reduced set of pairs to tune the deduplication process in
large datasets. T3S follows two stages to select the most
representative pairs. The first stage contains a strategy to
produce balanced subsets of candidate pairs for labeling. The
second stage proposes a rule-based active selective sampling
incrementally invoked to remove the redundant pairs in the
subsets created in the first stage in order to produce an even
smaller and more informative training set. This training set
can be furthur utilized. Active fuzzy region selectionalgorithm
is proposed to detect the fuzzy region boundaries by using the
training set. Thus, T3S reduces the labeling effortsubstantially
while achieving superior matching quality when compared
with state-of-the-art deduplication methods in large datasets.
But, performing the deduplication in a distributed
environment offers a better performance over the centralized
system in terms of speed and flexibility. So, in this work, a
distributed approach is implemented for the above method
using Apache Spark. Also, a comparison is done betweenTwo-
stage sampling selection strategy and FSDedup. It showsthat
T3S reduces the training set size by redundancy removal and
hence offers better performance than FSDedup. A graph is
plotted for the same.
Index Terms— Scala, Apache spark, Deduplication,
Sampling Selection Strategy, T3S
1 INTRODUCTION
Data deduplication is a specialized data compression
technique for eliminating duplicate copiesof repeating data.
This technique is used to improve storage utilizationandcan
also be applied to network data transfers to reduce the
number of bytes that must be sent. In the deduplication
process, unique chunks of data, or byte patterns, are
identified and stored during a process of analysis. As the
analysis continues, other chunksare compared to thestored
copy and whenever a match occurs, the redundant chunk is
replaced with a small reference that points to the stored
chunk. Deduplication may occur in-line, asdata isflowing,or
post-process, after it has been written. With post-process
deduplication, new data is first stored on the storage device
and then a process at a later time will analyze the data
looking for duplication. This is the process where the
deduplication hash calculations are created on the target
device asthe data entersthe device in real time. If the device
spots a block that it already stored on the system it does not
store the new block, just references to the existing block.
There has been a dramatic growth in the generation of
information from a wide range of sources such as mobile
devices, streaming media, and social networks. Data quality
is also degraded due to the presence of duplicate pairs with
misspellings, abbreviations, conflicting data, and redundant
entities, among other problems. Record deduplication aims
at identifying which objects are potentially the same in the
data repository. In the context of large datasets, it is a
difficult task to produce a replica-free repository. A typical
deduplication method is divided into three main phases:
1.1 Blocking:
The Blocking phase aims at reducing the number of
comparisons by grouping together pairs that share common
features. A simplistic blocking approach, for example, puts
together all the recordswith the same firstletterof the name
and surname attributes in the same block, thus avoiding a
quadratic generation of pairs (i.e., a situation where the
records are matched all-against-all).
1.2 Comparison:
The Comparison phase quantifies the degree of similarity
between pairs belonging to the same block, by applying
some type of similarity function (e.g. Jaccard, Levenshtein,
Jaro).
1.3 Classification:
Finally, the Classification phase identifieswhich pairsare
matching or non-matching. This phase can be carried out by
selecting the most similar pairs by means of global

thresholds, usually manually defined [1], [2], [3], [4] or
learnt by using a classification model based on a trainingset.
2 RELATED WORKS
Record deduplication studies have offered a wide range of
solutions exploiting supervised, semi-supervised, and
unsupervised strategies. Supervised and unsupervised
strategies rely on expert usersto configurethededuplication
process. The former assumesthe presence of alargetraining
set consisting of the most important patterns present in the
dataset (e.g., [8], [9]). The latter relies on threshold values
that are manually tuned to configure the deduplication
process (e.g., [1], [2], [10], [4]).Committee-based strategies
for deduplication, called ALIAS and Active Atlasrespectively,
are outlined in [7] and [11]. The committee identifies the
most informative pairs to be labeled by the user as the
unlabeled pairs that most classifiersdisagreeregardingtheir
prediction. Active Atlas employs a committee composed by
decision trees, while ALIAS uses randomized decision trees,
a Naive Bayes and/or a SVM classifier. An alternative active
learning method for deduplication was proposed in [5],
where the objective is to maximize the recall under a
precision constraint. The approachcreatesanN-dimensional
feature space composed of a set of similarity functions that
are manually defined, and actively selects the pairs by
carrying out a binary search over the space. However, theN-
dimensional binary search may lead to a large number of
pairsbeen queried, increasing the manual effort [6]. In [6], a
strategy, referred as ALD, is proposed to map any active
learning approach based on accuracy to an appropriate
deduplication metric under precision constraints. This kind
of approach projects a quality estimation ofeachclassifierby
means of points in a two-dimensional space. ALD conductsa
binary search in this space to select the optimal classifier
that respects the precision constraint. The spacedimensions
correspond to the classifiers’ effectiveness, estimated by
means of an ―oracle‖. The pairs used for training are
selected by the IWAL active learning method [12].
3 PROPOSED MODEL AND SYSTEM DESIGN
3.1 Terminologies
Sig-Dedup has been used to efficiently handle large
deduplication tasks. It maps the dataset strings into a set of
signatures to ensure that similar substringsresult in similar
signatures. The signatures are computed by means of the
inverted index method. To overcome the drawback of
quadratic candidate generation [15] prefix filtering [16] is
used. The prefix filtering is formally defined below:
Definition 1: Assume that all the tokens in each record are
ordered by a global ordering ϑ. Let p-prefix of a record be
the first p tokens of the record. If Jaccard(x,y) ≥ t, then the
(p)-prefix of x and (p)-prefix of y must share at least one
token. where, Jaccard(x,y) is defined as: J(x,y) = _____ Prefix
length of each record u is calculated as |u| − t · ⌈ ⌉ +1 , where
t= Jaccard similarity threshold.
3.2. Framework
The framework for large scale deduplication using the two
stage sampling selection strategy is illustrated in Figure 4.1.
First, the candidate pairs are produced after identifying the
blocking threshold. Next, T3S strategy is employed. In its
first stage, T3S produces small balanced subsamples of
candidate pairs. In the second stage, the redundant
information that is selected in the subsamplesisremovedby
means of a rule-based active sampling.Thesetwostepswork
together to detect the boundaries of the fuzzy. Finally, the
classification approach is introduced which is configured by
using the pairs manually labeled in the two stages.
All these steps are implemented in the distributed
environment using Apache Spark.
Figure-3.1 Framework for Large Scale Deduplication
using T3S
3.2.1. Determining Blocking Threshold
In large datasets, it is not feasible to run the Sig-Dedupfilters
with different thresholds due to the high computational
costs. So, a stopping criterion is introduced. The method
employed is defined as:
Definition 2: Consider a subset S, created from a randomly
sampled dataset D and a range of thresholds with fixed step
thj = 0.2, 0.3,. . ., and 0.9. The subset S is matched using each
threshold value thj. The initial threshold will be the first thj
that results in a number of candidate pairs smaller than the
number of records in S.
After finding the global initial threshold value for the
blocking process, the entire dataset is matched to create the
set of candidate pairs.
3.2.2. First stage of T3S: Sample Selection Strategy
The first stage of T3S adopts the concept of levels to allow
each sample to have a similar diversity to that of the full set
of pairs. The ranking, created by the blocking step, is
fragmented into 10 levels (0.0-0.1,0.1-0.2, 0.2-0.3,. . ., and
0.9-1.0), by using the similarity value of each candidate pair.

The similarity value of each candidate pair is found using
Jaccard similarity. This fragmentation produces levels
composed of different matching patterns to prevent non-
matching pairs dominating the sample.
3.2.3. Second Stage of T3S: Redundancy Removal
Several pairs selected inside each level are composed of
redundant which does not help to increase the training set
diversity. Selective Sampling using Association Rulesisused
to remove redundancy in the information randomlyselected
as shown in Figure 3.2.
Figure-3.2 Illustration of SSAR
3.2.3.1. SSAR Method
The second stage of T3S aims at incrementally removing the
non-informative or redundant pairsinside eachsamplelevel
by using the SSAR (Selective Sampling using Association
Rules) active learning method [14]. In the beginning, when
the training set D is empty, SSAR selects the pair that shares
most feature values with all other unlabeled pairsto initially
compose the training set. SSAR selects an unlabeled pair ui
for labeling by using inferences about the number of
association rules produced within a projected training set
specific for ui. The projected training set is produced by
removing from the current training set D instances and
features that do not share features values with ui. When
compared with the current training set, the unlabeled pair
with less classification rules over the projected training set
represents the most informative pair. If this pair is not
already present in the training set, it is labeled by the user
and inserted into the training set. After this, a new round is
performed and the training set must be re-projectedforeach
remaining unlabeled pair to determine which one is most
dissimilar when compared to the current training set. If the
selected pair is already present in the training set, the
algorithm converges.
3.2.3.2. Computational Complexity
The computational complexity of SSAR is O(S * |U| * 2m),
where ―S‖ is the number of pairs selected to be labeled,
―|U|‖ represents the total number of candidate pairs and
―m‖ is the number of features. ―|U|‖ pairs must be re-
projected each time that a labeled pair is attached to the
current training set, producing a computationallyunfeasible
time to process large datasets.
3.2.4. Fuzzy Region Detection
Definition 3: Let Minimum True Pair-(MTP) represent the
matching pair with the lowest similarity value amongtheset
of candidate pairs.
Definition 4: Let Maximum False Pair-(MFP) represent the
non-matching pair with the highest similarity value among
the set of non-matching pairs.
The fuzzy region is detected by using manually labeledpairs.
The user is requested to manually label pairs that are
selected incrementally by the SSAR from each level. First,
SSAR is invoked to identify the informative pairs
incrementally inside each level toproduceareducedtraining
set. The pairs labeled within each level are used to identify
the MFP and MTP pairs.
MTP and MFP pairs define the fuzzy region boundaries the
similarity value of the MTP and MFP pairs identifies α and β
values. The pairs belonging to the fuzzy region are sent to
the Classification Step.
3.2.5. Classification
The Classification step aims at categorizing the candidate
pairs belonging to the fuzzy region as matching or non-
matching.
The classifier, T3S-NGram maps each record to a global
sorted token set and then applies both the Sig-Dedup
filtering and a defined similarity function(suchasJaccard)to
the sets. The NGram Threshold is required to identify the
matching pairs inside the fuzzy region using the NGram
tokenization.First, the similarity of each labeled pair is
recomputed by means of a similarity function along withthe
NGram tokenization. After this, the labeled pairs are sorted
incrementally by the similarity value and a sliding window
with fixed-size N is applied to the sorted pairs. The sliding
window is relocated in one position until it detects the last
windows with only non-matching pairs.
Finally, the similarity value of the first matching pair
encountered after the last windowswith only non-matching
pairs, defines the NGram threshold value. The candidate
pairs that survive the filtering phase and meet the Ngram
threshold value are considered as matching ones.
4 CONCLUSIONS
In this project, we have proposed a distributedalgorithmfor
large scale deduplication using sampling selection strategy
which produces the same result as the centralized system
but speeds up the process by a considerable amount. As
followed from our experiments, our distributed approach

performs the same processes in a lesser time with a greater
flexibility and scalability. We have also compared the T3S
approach with the FSDedup. T3S reduces the user effort by
reducing the training set size and results in a lesser number
of matching pairs.
ACKNOWLEDGEMENT
The authors can acknowledge any person/authoritiesinthis
section. This is not mandatory.
REFERENCES
[1] R. J. Bayardo, Y. Ma, and R. Srikant, ―Scaling up all pairs
similarity search,‖ in Proc. 16th Int. Conf. World Wide Web,
pp. 131–140, 2007.
[2] S. Chaudhuri, V. Ganti, and R. Kaushik, ―A primitive
operator for similarity joins in data cleaning,‖ in Proc. 22nd
Int.Conf. Data Eng., p. 5, Apr. 2006.
[3] J. Wang, G. Li, and J. Fe, ―Fast-join: An efficient method
for fuzzy token matching based string similarity join,‖ in
Proc. IEEE 27th Int. Conf. Data Eng., 2011, pp. 458–469.
[4]C. Xiao, W. Wang, X. Lin, J. X. Yu, and G. Wang,―Efficient
similarity joins for near-duplicate detection,‖ ACM Trans.
Database Syst., vol. 36, no. 3, pp. 15:1–15:41, 2011.
[5]A. Arasu, M. Gotz, and R. Kaushik, ―On active learning of
record matching packages,‖ in
[6]K. Bellare, S. Iyengar, A. G. Parameswaran, and V.
Rastogi,―Active sampling for entity matching,‖ inProc.18th
ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2012,
pp. 1131–1139.
[7]S. Sarawagi and A. Bhamidipaty, ―Interactive
deduplication using active learning,‖ in Proc. 8th ACM
SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2002, pp.
269–278.
[8] M. G. de Carvalho, A. H. Laender, M. A. Goncalves, andA.S.
da Silva, ―A genetic programming approach to record
deduplication,‖ IEEE Trans. Knowl. Data Eng., vol. 24, no. 3,
pp. 399–412, Mar. 2012.
[9]J. Wang, G. Li, J. X. Yu, and J. Feng, ―Entity matching: How
similar is similar,‖ Proc.
[10]R. Vernica, M. J. Carey, and C. Li, ―Efficient parallel set-
similarity joins using mapreduce,‖ in Proc. ACMSIGMODInt.
Conf. Manage. Data, 2010, pp. 495–506.
[11]S. Tejada, C. A. Knoblock, and S. Minton, ―Learning
Domain-independent string transformationweightsforhigh
accuracy object identification,‖ in Proc. 8th ACMSIGKDDInt.
Conf. Knowl. Discovery Data Mining, 2002, pp. 350–359.
[12]A. Beygelzimer, S. Dasgupta, and J. Langford,
―Importance weighted active learning,‖ in Proc. 26th Annu.
Int. Conf. Mach. Learn., pp. 49–56, 2009.
[13] G. Dal Bianco, R. Galante, C. A. Heuser, and M. A.
Gonalves, ―Tuning large scale deduplication with reduced
effort,‖ in Proc. 25th Int. Conf. Scientific Statist. Database
Manage. 2013, pp. 1–12.
[14] R. M. Silva, M. A. Gonc¸alves, and A. Veloso,―Atwo-stage
active learning method for learning to rank,‖ J. Assoc.
Inform. Sci. Technol., vol. 65, no. 1, pp. 109–128, 2014.
[15]M. Bilenko and R. J. Mooney, ―On evaluation and
Training-set construction for duplicate detection,‖ in Proc.
Workshop KDD, 2003, pp. 7–12.
[16]A. Arasu, C. R_e, and D. Suciu,―Large-scalededuplication
with constraints using dedupalog,‖ in Proc. IEEE Int. Conf.
Data Eng., 2009, pp. 952–963.
[17] Guilherme Dal Bianco, Renata Galante, Marcos Andr_e
Gonc¸alves, Sergio Canuto, and CarlosA.Heuser, ―APractical
and Effective Sampling Selection Strategy for Large Scale
Deduplication,‖ IEEE Transactions On Knowledge And Data
Engineering, Vol. 27, No. 9, September 2015.

IRJET- Sampling Selection Strategy for Large Scale Deduplication of Synthetic and Real Datasets using Apache Spark

More Related Content

What's hot (20)

Similar to IRJET- Sampling Selection Strategy for Large Scale Deduplication of Synthetic and Real Datasets using Apache Spark (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Sampling Selection Strategy for Large Scale Deduplication of Synthetic and Real Datasets using Apache Spark