A Privacy Preserving Efficient
Protocol for Semantic Similarity Join
Using Long String Attributes
Bilal Hawashin, Farshad Fotouhi
Traian Marius Truta
Department of Computer Science
Wayne State University
Northern Kentucky University
Outlines







What is Similarity Join
Long String Values
Our Contribution
Privacy Preserving Protocol For Long String Values
Experiments and Results
Conclusions/Future Work
Contact Information
Motivation
Is Natural Join always suitable?
Name
John Smith
Mary Jones
Address
4115 Main St.
2619 Ford Rd.
Major
Biology
Chemical Eng.
…
Name
Address
Smith, John 4115 Main
Street
Mary Jons
2619 Ford Rd.
Monthly Sal.
1645
…
2100
Similarity Join



Joining a pair of records if they have SIMILAR
values in the join attribute.
Formally, similarity join consists of grouping
pairs of records whose similarity is greater than a
threshold, T.
Studied widely in the literature, and referred to
as record linkage, entity matching, duplicate
detection, citation resolution, …
Our Previous Contribution: Long
String Values (ICDM MMIS10)






The term long string refers to the data type representing any
string value with unlimited length.
The term long attribute refers to any attribute of long string data
type.
Most tables contain at least one attribute with long string
values.
Examples are Paper Abstract, Product Description, Movie
Summary, User Comment, …
Most of the previous work studied similarity join on short
fields.
In our previous work, we showed that using long attributes as
join attributes under supervised learning can enhance the
similarity join performance.
Example
P1 Title
P2 Title
P3 Title
…
P1 Kwds
P2 Kwds
P3 Kwds
P1 Authrs P1 Abstract
P2 Authrs P2 Abstract
P3 Authrs P3 Abstract
…
…
…
P10 Title
P11 Title
…
P10 Kwds P10 Authrs P10 Abstract …
P11 Kwds P11 Authrs P11 Abstract …
Our Paper (Motivation)

Some sources may not allow sharing its whole data in
the similarity join process.




Solution: Privacy Preserved Similarity Join.
Using long attributes as join attributes can increase the
similarity join accuracy.
Up to our knowledge, all the current Privacy Preserved
SJ algorithms use short attributes.
Most of the current privacy preserved SJ algorithms
ignore the semantic similarities among the values.
Problem Formulation

Our goal is to find a Privacy Preserved Similarity
Join Algorithm when the join attribute is a long
attribute and consider the semantic similarities
among such long values.
Our Work Plan


Phase1: Compare multiple similarity methods
for long attributes when similarity thresholds are
used.
Phase2: Use the best method as part in the
privacy preserved SJ protocol.
Phase1: Finding Best SJ Method
for Long Strings with Threshold

Candidate Methods:
Diffusion Maps.
 Latent Semantic Indexing.
 Locality Preserving Projection.

Performance Measurements

F1 Measurement: the harmonic mean between
recall R and precision P.
Where recall is the ratio of the relevant data
among the retrieved data, and precision is the
ratio of the accurate data among the retrieved
data.
Performance Measurements(Cont.)



Preprocessing time is the time needed to read the
dataset and generate matrices that could be used later as
an input to the semantic operation.
Operation time is the time needed to apply the
semantic method.
Matching time is the time required by the third party, C,
to find the cosine similarity among the records
provided by both A and B in the reduced space and
compare the similarities with the predefined similarity
threshold.
Datasets

IMDB Internet Movies Dataset:


Movie Summary Field
Amazon Dataset:
Product Title
 Product Description

Phase1 Results

Finding best dimensionality reduction method
using Movie Summary from IMDB Dataset
(Left) and Product Descriptions from Amazon
(Right).
Phase2 Results

Preprocessing Time:
Read Dataset (1000
Movie Summaries)
TF.IDF Weighting
12 Sec.
Reduce Dimensionality
using Mean TF.IDF
Find Shared Features
0.5 Sec.
1 Sec.
Negligible
Phase2 Results

Operation Time for the best performing
methods from phase 1.

Matching Time is negligible.
Our Protocol

Both sources A and B share the Threshold value
T to decide similar pairs later.
Our Protocol
Source A
P1 Title
P1 Authors
P1 Abstract
…
P2 Title
P2 Authors
P2 Abstract
…
P3 Title
P3 Authors
P3 Abstract
Px Authors
Px Abstract
Source B
Px Title
…
Ma
Find Term_LSV Frequency
Matrix for Each Source
LSV1
LSV2
LSV3
Image
4
0
0
Classify
5
0
0
Similarity
0
6
5
Join
0
6
4
Find TD_Weighted Matrix Using
TF.IDF Weighting
WeightedMa
LSV1
LSV2
LSV3
Image
0.9
0
0
Classify
0.7
0
0
Similarity
0
0.85
0.9
Join
0
0.7
0.85
TF.IDF Weighting
TF.IDF weighting of a term W in a long string
value x is given as:
where tfw,x is the frequency of the term w in the
long string value x, and idfw is , where N is the
number of long string values in the relation, and
nw is the number of long string values in the
relation that contains the term w.
MeanTF.IDF Feature Selection



MeanTF.IDF is an unsupervised feature selection
method.
Every feature (term) is assigned a value according to its
importance.
The Value of a term feature w is given as
Where TF.IDF(w, x) is the weighting of feature w in
long string value x, and N is the total number of long
string values.
Apply MeanTF.IDF on WeightedMa and Get
Important Features to Imp_Fea.
Add Random features to Imp_Fea to get
Rand_ Imp_Fea.
Rand_ Imp_Fea and Rand_ Imp_Feb are
returned to C.
C Finds the intersection and return the
shared important features SF to both A
and B.
Reduced WeightedM Dimensions in
Both Sources using SF.
SFa
LSV1
LSV2
LSV3
Image
0.9
0
0
Similarity
0
0.85
0.9
…
Add Random Vectors to SF
Rand_Weighted_a
LSV1
LSV2
LSV3
Random
Cols
Image
0.9
0
0
0.6
Similarity
0
0.85
0.9
0.2
…
Find Wa (The Kernel)
…
1-Cos_Sim(LSV1,LSV1)=0
1-Cos_Sim(LSV1,LSV2)=0.2
1-Cos_Sim(LSV1,LSV3)=0.3
…
1-Cos_Sim(LSV2,LSV1)=0.2
1-Cos_Sim(LSV2,LSV2)=0
1-Cos_Sim(LSV2,LSV3)=0.87
…
1-Cos_Sim(LSV3,LSV1)=0.3
1-Cos_Sim(LSV3,LSV2)=0.87
1-Cos_Sim(LSV3,LSV3)=0
…
…
…
…
|Wa| = D x D, where D is total number of
columns in Rand_Weighteda
Use Diffusion Maps to Find
Red_Rand_Weighted_a

[Red_Rand_Weighted_a,Sa,Va,Aa] = Diffusion_Map(Wa , 10, 1, red_dim),
red_dim < D

Red_Rand_Weighted_a=Diffusion Map Representation of first row of W
a
Diffusion Map Representation of second row of Wa
Diffusion Map Representation of third row of Wa
Col1
Col2
0.4
0.1
0.8
0.6
0.75
0.5
…
…
Colred_dim
C Finds Pairwise Similarity Between
Red_Rand_Weighted_a and
Red_Rand_Weighted_b
Red_Rand_
Weighted_a
1
Red_Rand_ Cos_Sim
Weighted_b
1
0.77
1
2
0.3
…
…
…
2
1
0.9
If Cos_Sim>T, Insert the tuple in
Matched
Matched
Red_Rand_
Weighted_a
1
Red_Rand_ Cos_Sim
Weighted_b
1
0.77
2
1
0.9
2
7
0.85
…
…
…
Matched is returned to both A and B.
A and B remove random vectors from
Matched and share their matrices.
Our Protocol (Part1)
Our Protocol (Part2)
Phase2 Results

Effect of adding random columns on the
accuracy.
Phase2 Results

Effect of adding random columns on the
number of suggested matches.
Conclusions



Efficient secure SJ semantic protocol for long string
attributes is proposed.
Diffusion maps is the best method (among compared)
to semantically join long string attributes when
threshold values are used.
Mapping into diffusion maps space and adding random
records can hide the original data without affecting the
accuracy.
Future Work
Potential further works:
 Compare diffusion maps with more candidate
semantic methods for joining long string
attributes.
 Study the performance of the protocol on huge
databases.
Thank You …


Dr. Farshad Fotouhi.
Dr. Traian Marius Truta.