+ All Categories
Home > Documents > Socially-aware Query Routing in Mobile Social Networks Andreas Konstantinidis, Demetrios...

Socially-aware Query Routing in Mobile Social Networks Andreas Konstantinidis, Demetrios...

Date post: 02-Jan-2016
Category:
Upload: caren-burns
View: 216 times
Download: 1 times
Share this document with a friend
Popular Tags:
29
Socially-aware Query Routing in Socially-aware Query Routing in Mobile Social Networks Mobile Social Networks Andreas Konstantinidis , Demetrios Zeinalipour-Yazti Department of Computer Science, University of Cyprus, Cyprus and Kun Yang School of Computer Science and Electronic Engineering, University of Essex, UK Hellenic Data Management Symposium, Hellenic Data Management Symposium, 2010 2010
Transcript

Socially-aware Query Routing in Mobile Socially-aware Query Routing in Mobile Social NetworksSocial Networks

Andreas Konstantinidis, Demetrios Zeinalipour-YaztiDepartment of Computer Science, University of Cyprus, Cyprus

and Kun YangSchool of Computer Science and Electronic Engineering, University of Essex, UK

Hellenic Data Management Symposium, 2010Hellenic Data Management Symposium, 2010

Speaker: Andreas Konstantinidis – University of Cyprus

Social Networks (on the Web)

Social Network: a set of people or groups of people with some pattern of contact or interaction among them– Attracted billions of active users under major online social

network systems – Examples: MySpace, Facebook, Twitter

Speaker: Andreas Konstantinidis – University of Cyprus

Mobile Social Networks (MoSoNets)

Mobile Social Network: Social Network applications for smartphone devices.– Examples: Google Latitude and Google Buzz, Foursquare,

Gowalla and Loopt.

• Smartphone: offers more advanced computing and connectivity than a basic 'feature phone'.

• E.g., OS: Android, Nokia’s Maemo, Apple X

Speaker: Andreas Konstantinidis – University of Cyprus

Mobile Social Networks (MoSoNets)

• Mobile Social Network applications are projected to grow in the future.

• Google Latitude already reports over 3 Million Users with more than 1 Million Users available online concurrently.

Speaker: Andreas Konstantinidis – University of Cyprus

Motivation

• Numerous research challenges arise in the context of Mobile Social Networks– Data Management Challenges: Query Processing and

Retrieval, Storage (Cloud vs. Local), Access Methods, etc.– Mobility Challenges: Context Awareness, etc.– Social Challenges: Privacy, etc.– System Challenges: Architectures, Platforms etc.

In this work we attempt to exploit knowledge about the underlying social network in order to improve query routing in Mobile Social Networks.

Speaker: Andreas Konstantinidis – University of Cyprus

Example Scenario

Scenario: Five (5) users moving in Lower Manhattan collecting data (video, photos, sound, rss, …)

U1

U2

U4

U5

U3

Speaker: Andreas Konstantinidis – University of Cyprus

Example Scenario: Assumptions

Assumptions• Users feature “long-range” connectivity (e.g., WiFi | 3G) and “short-range” connectivity (e.g., Bluetooth)• Communication Links are Expensive (i.e., due to energy and bandwidth constraints)

>> Bandwidth Constraints• 4G nets in the US (Sprint, AT&T) promise 3-10MBps but offer as low as 0,6MBps.>> Power Constraints :• 0.40W – No connections• 0.52W – Bluetooth Connection Established• 1.73W – Download 120KBps via 3G

Speaker: Andreas Konstantinidis – University of Cyprus

Example Scenario

Find Video of street artists performing right now?

U1 U2 U3 U4 U5

{(X,Y,T,obj) | X,Y: spatial, T: temporal, Obj: object}

Fact: Content is Distributed and there is no Global Index!Problem: How to find the answer without flooding the SmartNet

Mobile Social Networking Service

Speaker: Andreas Konstantinidis – University of Cyprus

Example Scenario

U1 U2 U3 U4 U5

Interest Matrix (Profile)Arts Food Cinema

U1 XU2 X XU3 XU4 X X

Query Routing Tree

(T) Disseminate Query using T

MoSoNet Service

Query Processor

Social Graph (G)

(WiFi| 3G) Bluetooth (cheaper)

Bluetooth(cheaper)

Speaker: Andreas Konstantinidis – University of Cyprus

Example Scenario

U1 U2 U3 U4 U5

Download Photo\Video(via WiFi|3G|Bluetooth)

MSN Service

Query Processor

We do not consider this phase in greater detail

Speaker: Andreas Konstantinidis – University of Cyprus

Overview

• Introduction and Motivation

• Problem Formulation

• Multi-objective Optimization of Query Routing Trees

• Experimental Setup & Evaluation

• Current/Future work

Speaker: Andreas Konstantinidis – University of Cyprus

• Why Use Query Routing Trees (T)?– Avoid Flooding the Network w/ Queries (Scalable)

• More Efficient in terms of Energy, Communication, etc.

– Better Query response quality • An out-of-sync centralized data repository performs worse

than a “live” decentralized data repository.

– Optimally exploit short vs. long range communication links (i.e., Bluetooth vs. WiFi|3G)

– Finally, it offers more Privacy (No single authority has a global view of all data).

Query Routing Trees (T)

Speaker: Andreas Konstantinidis – University of Cyprus

Problem: Construct a Query Routing Tree (T), for a mobile social network, that optimizes the following three (3) conflicting objectives, concurrently:– Α) Minimize Overhead, in conducting the

query– B) Maximize (Query Result) Quality.– C) Maximize Social Interaction (i.e., exploit

interactions in the physical space)

More formal measures defined next…

Query Routing Tree Problem (QRTP)

Speaker: Andreas Konstantinidis – University of Cyprus

QRTP: Objective 1

• A) Minimum Overhead: a lower number of answers, assures lower traffic load and lower bandwidth consumption.

||)(min XXOv Smaller Tree, Less Answers Lower Quality! Lower Overhead Neutral Interactions

Speaker: Andreas Konstantinidis – University of Cyprus

QRTP: Objective 2

• B) Maximize Quality: higher number of relevant answers based on interests matrix.

n

iijj IntsjNnXivQXQI

1

),,|()|(max

Larger Tree, More Answers Higher Quality! Higher Overhead Neutral Interactions

Speaker: Andreas Konstantinidis – University of Cyprus

QRTP: Objective 3• C) Maximize Social Interaction: Frequency of user interaction in physical space.

– How this can be determined? Based on Bluetooth interactions of users in physical space

– Solution 1Solution 1: few users with HIGH SI Lower Quality! Lower Overhead

– Solution 2Solution 2: many users with HIGH SIHigh Quality! High Overhead

i

iz XzimXSI ))(,|()(max

Speaker: Andreas Konstantinidis – University of Cyprus

Overview

• Introduction and Motivation

• Problem Formulation

• Multi-objective Optimization of Query Routing Trees

• Experimental Setup & Evaluation

• Current/Future work

Speaker: Andreas Konstantinidis – University of Cyprus

Multi-Objective Optimization (MOO)• Classical single objective optimization has the form:

– where x is a discrete vector representing a solution (e.g. a network design, a route)– y is a real value representing the solution quality– f is the objective function

• Multi-Objective Optimization

– No single solution is optimal under all objectives– Improve one deteriorates the others– Partial ordering of solutions (“y dominates z“)

– Pareto optimal set (maps to the Pareto Front (PF) )

)(max xfy

),...,,(

))(),...,(),(()(max

21

21

n

m

xxxxwhere

xfxfxfxfZ

jiii zymjzymiZ ,..1,..1, zyzy,

)()(,|* ** xxxx ffXX

non-dominated solutions in PF

dominated solution

PF

f2

f1

y

z

x

Speaker: Andreas Konstantinidis – University of Cyprus

MOO Approaches: MOEAsEAs to MOEAs, good in obtaining a set of non-dominated solutions in a single run:

– Deal with a population

of solutions.– Converge towards near-

optimal solutions fast.

Main steps of EAs:– Objective functions– Encoding Representation– Initialization– Genetic components

• Selection• Crossover• Mutation

– Update (elitism: use of archive)

Initialization

Selection

Reproduction: Crossover Mutation

Survival

Update

… …

… …

Speaker: Andreas Konstantinidis – University of Cyprus

KEY CHRACTERISTICS• Decomposes a MOP into a set of SOPs

using any technique for aggregating functions:

– e.g. weighted sum, Tchebycheff:

• Tackles them simultaneously, using neighbourhood information and SOO techniques.

• Hybridize with local-search based techniques.

• Incorporate problem-specific knowledge.

• Andreas Konstantinidis, Kun Yang, Qingfu Zhang and Demetrios Zeinalipour-Yazti, "A Multi-Objective Evolutionary Algorithm for the Deployment and Power Assignment Problem in Wireless Sensor Networks", SI-New Network Paradigms, Computer Networks, vol. 54, pp. 960-976, 2010.

MOEA/D framework

Speaker: Andreas Konstantinidis – University of Cyprus

QRTP Operation Summary

Speaker: Andreas Konstantinidis – University of Cyprus

Overview

• Introduction and Motivation

• Problem Formulation

• Multi-objective Optimization of Query Routing Trees

• Experimental Setup & Evaluation

• Current/Future work

Speaker: Andreas Konstantinidis – University of Cyprus

Experimental Setup– Simulator: We have implemented a trace-

driven simulator in Java (a good starting point for evaluating ideas at a preliminary stage)

– Datasets: Synthetic based on Random Distributions (for Social Interaction and Interest Matrix)

– Query-By-Example:• SELECT IP, Filename• FROM MobileSocialNetwork• WHERE similar(multimedia-object)

– Evaluation Metrics: Next Slide

Speaker: Andreas Konstantinidis – University of Cyprus

Performance MetricsEvaluation metrics

– Quality & diversity of solutions (using five metrics).– Bandwidth cost BW(X): the product of n ≤ N in

tree X and the number of fragmented packets f of size MTU for data of a particular type (e.g. video, image, email) and size l:

– Latency L(X): the sum of the information of size f×MTU, transferred per node over a specific wireless network (e.g. WiFi) with a data rate DR:

where f = l/(MTU −hd) and hd is the TCP/IP header size.

Speaker: Andreas Konstantinidis – University of Cyprus

Results & Discussion

• MOEA/D vs NSGA-II

• NSGA-II the state-of-the-art in MOEAs based on Pareto dominance.• Pairs of two objective are used.• Similar conclusions for the third objective.

Higher Quality of QRTsHigher Quality of QRTsHigher number of Non-dominated SolutionsHigher number of Non-dominated Solutions

Better DiversityBetter Diversity

Speaker: Andreas Konstantinidis – University of Cyprus

• Bandwidth Consumed during Searches

A) Agnostic Approach: Search by flooding.

B) Informed Approach: Search over Optimal QRT.

Results & Discussion

50GB

7GB20MB

180MB

Standard Deviation

is low

Speaker: Andreas Konstantinidis – University of Cyprus

Overview

• Introduction and Motivation

• Problem Formulation

• Multi-objective Optimization of Query Routing Trees

• Experimental Setup & Evaluation

• Conclusions and Future work

Speaker: Andreas Konstantinidis – University of Cyprus

Conclusions and Future Work• Mobile Social Networks are a new area with many new opportunities.• In the future we aim to:

– Deploy more realistic mobility models (GEOLife GPS Trajectories by Microsoft Asia).

– Real implementation using Android technology.

– Use realistic data sets for generating the interests matrix (currently working on DBLP dataset).

– Evaluate the time cost for solving the QRT problem on larger-scale information spaces.

• Future:– Hybridization of MOEA/D with local-search heuristics.

Speaker: Andreas Konstantinidis – University of Cyprus

Socially-aware Query Routing in Mobile Networks

Thank you!Questions?

Andreas KonstantinidisUniversity of Cyprus

[email protected]


Recommended