[PPT] - Two-way Recommendation Methods for Social Networks Richi Nayak PowerPoint Presentation

SLIDE 1

Queensland University of Technology

CRICOS No. 00213J

Two-way Recommendation Methods for Social Networks

Richi Nayak Data Science Discipline Queensland University of Technology Brisbane, Australia

SLIDE 2

CRICOS No. 00213J

a university for the world

real

R

Introduction

Number of users and information

they provide in social networks are rapidly increasing.

Eg, more than 500 million active

users in Facebook, with 50% of the active users logging on in any given day.

Beside the explicit data, social

matching systems keep implicit data which is a rich data source, however it usually ignored.

SLIDE 3

CRICOS No. 00213J

a university for the world

real

R

The Overload and mismatch problem

3

SLIDE 4

CRICOS No. 00213J

a university for the world

real

R

Too specific – No results

4

SLIDE 5

CRICOS No. 00213J

a university for the world

real

R

Solution: Recommender Systems

Traditional recommendation systems

Provide one-way recommendation
Unable to handle sparse datasets effectively
Have scalability issues

5

SLIDE 6

CRICOS No. 00213J

a university for the world

real

R

Social Recommendation – 2 way Matching

6

SLIDE 7

CRICOS No. 00213J

a university for the world

real

R

Rich Source of Information

7

SLIDE 8

CRICOS No. 00213J

a university for the world

real

R

Underlying Social Network Data: A Sample

3 Months Data Value # of distinct active users 163,050 # of female users 82,500 # of male users 80,550 # unique kiss senders 122,396 # unique successful senders 91,487 # unique kiss recipients in the network 198,293 # unique kiss recipients who are active during the chosen period 83,865 # unique kisses 886,396 # unique successful kisses 171,158 # unique negative kisses 346,193 # unique null kisses 369,045

8

SLIDE 9

CRICOS No. 00213J

a university for the world

real

R

Evaluation Metrics

9 u by initated kisses unique

f

Number u by initiated kisses successful unique

f

Number u SR = ) (

) ( ) Re ( ) ( Re Partners Kissed

f

Number partners commended partners Kissed

f

Number u call ∩ =

users

f

tions recommenda received who users

f

no Total No Coverage =

links new

f

Number links missed

f

Number FN =

links unconnect

f

Number links predicted false

f

Number FP =

SLIDE 10

CRICOS No. 00213J

a university for the world

real

R

Bow Tie Structure

10

SLIDE 11

CRICOS No. 00213J

a university for the world

real

R

Indegree & Outdegree

11

SLIDE 12

CRICOS No. 00213J

a university for the world

real

R

Centrality Analysis

SLIDE 13

CRICOS No. 00213J

a university for the world

real

R

All communication in SN Successful communication in SN

SLIDE 14

CRICOS No. 00213J

a university for the world

real

R

Reachability

SLIDE 15

CRICOS No. 00213J

a university for the world

real

R

Co-clustering Methods

15

SLIDE 16

CRICOS No. 00213J

a university for the world

real

R

Experimental Results

16

SLIDE 17

CRICOS No. 00213J

a university for the world

real

R

Experimental Results

17

SLIDE 18

CRICOS No. 00213J

a university for the world

real

R

Segmentation for Recommendation Method

18

SLIDE 19

CRICOS No. 00213J

a university for the world

real

R

Recommendation Strategies for Different Segment

Active User Segment

– Motivation: Users send messages frequently but with 10.7% of SR. – Task: Provide advice to user, a classification task which is cheaper to implement than recommendation – Gradient Boosting is performed for the classification

Moderate User Segment

– Moderate users send a moderate number of messages, the users may like to receive several more options before starting long term relations. – Task: Interaction based Co-clustering is a quick recommendation process and can achieve high success rate.

Quiet User Segment

– Motivation: TSM captures the latent relations between users and features and therefore is able to provide quality and quantity recommendation. – Apply tensor decomposition algorithm, calculate core tensor, reconstruct

riginal tensor in order to make recommendation.

19

SLIDE 20

CRICOS No. 00213J

a university for the world

real

R

Experimental Results

Active User Segment
Moderate User Segment

20

Classification Method Precisio n Recall Naïve Bayes 0.75 0.79 Neural Network 0.77 0.80 Linear Regression 0.79 0.82 Decision Tree 0.77 0.81 Gradient Boosting 0.87 0.89

SLIDE 21

CRICOS No. 00213J

a university for the world

real

R

Experimental Results

Method BSR Success Rate SRI Gradient Boosting method for active user segment 0.107 0.39 3.6 Co-clustering for moderate user segment 0.133 0.58 4.4 TSM for quiet user segment 0.169 0.72 4.3 Average SR of three methods 0.136 0.56 4.1 SocialCollab 0.156 0.35 2.2 CollabNet 0.156 0.54 3.5 Adapted SimRank (CDAS) 0.156 0.36 1.8 CF 0.156 0.168 1.1 CF+ 0.156 0.30 1.9

21

SLIDE 22

CRICOS No. 00213J

a university for the world

real

R

Discussion

Analysing the structure and features of a social

network helps to understand how the recommendations in the network should be carried out.

– indegree & outdegree, – bow tie structure, – reachability, – user behaviour.

22

SLIDE 23

CRICOS No. 00213J

a university for the world

real

R

Discussion

Memory-based methods vs. Model-based methods

– Memory-based methods achieve lower recall than model- based methods – Most of times model-based methods outperform memory- based methods in terms of SR

Effect of input data

– Interaction only vs. combined static and interaction data, using combined data achieves higher SR.

Effect of one-way vs. two-way recommendation

– Two-way recommendation methods generate high success rate or precision than one-way methods, but one- way methods outperform two-way methods in recall.

23

SLIDE 24

CRICOS No. 00213J

a university for the world

real

R

Discussion

Co-clustering methods

– Co-clustering presents higher quality recommendation than traditional clustering – Inclusion of both interaction data and static data enhances the results performance – Inclusion of learning algorithm enhances performance in SR.

Segmentation methods

– Segmentation method helps to save computation costs – Tensor based method achieves high SR and recall. However it is time-consuming to implement.

24

SLIDE 25

CRICOS No. 00213J

a university for the world

real

R

Potential Future Work

Cold-start problem
Methods that can achieve higher recall
Scalable methods

25

SLIDE 26

CRICOS No. 00213J

a university for the world

real

R

Publication

Kutty S, Chen L, Nayak R, (2013) A people-to-people matching system using graph mining techniques. World Wide Web Journal, pp. 1-

39, online available on http://link.springer.com/article/10.1007%2Fs11280-013-0202-z

Chen L, Nayak, R (2012) Leveraging the network information for evaluating answer quality in a collaborative question answering portal.

Social Network Analysis and Mining 2 (3), pp. 197-215.

Noor I, & Nayak R, (2014) Tensor-based Item Recommendation using Probabilistic Ranking in Social Tagging Systems, Proceedings of

the WWW Companion 2014.

Chen L, Nayak R, (2013) A recommendation approach dealing with multiple market segments, Proceedings of the IEEE/WIC/AMC

International Conference on Web Intelligence (WI) and Intelligent Agent Technology (IAT), pp. 89-94.

Chen L, Nayak R, (2013) A reciprocal collaborative method using relevance feedback and feature importance. Proceedings of the

IEEE/WIC/AMC International Conference on Web Intelligence (WI) and Intelligent Agent Technology (IAT), pp. 133-138.

Chen L, Nayak R, Kutty S, Xu Y, (2013) Users segmentations for recommendation. Proceedings of the 28th Annual ACM Symposium on

Applied Computing, ACM, pp. 279-280

Kutty S, Chen L, Nayak R, (2012) A people-to-people recommendation system using tensor space models, Proceedings of the 27th

Annual ACM Symposium on Applied Computing, pp. 187-192.

Chen L, Nayak R, Xu Y, (2012) A common neighbour based two-way collaborative recommendation method, Proceedings of the 27th

Annual ACM Symposium on Applied Computing, pp. 214-215.

Alsaleh, S., Nayak, R., Xu, Y.(2012) Grouping people in social networks using a weighted multi-constraints clustering method, IEEE

International Conference on Fuzzy Systems, http://dx.doi.org/10.1109/FUZZ-IEEE.2012.6250799

Chen L, Nayak R, Xu Y, (2011) A recommendation method for online dating networks based on social relations and demographic

information, Proceedings of the 2011 International Conference on Advances in Social Networks Analysis and Mining, ASONAM 2011, pp. 407-411.

Chen, L., Nayak, R. (2011) Social network analysis of an online dating network, Proceedings of C and T 2011 - 5th International

Conference on Communities and Technologies, pp. 41-49.

Alsaleh S, Nayak R, Xu Y, (2011) Finding and matching communities in social networks using data mining, Proceedings of the 2011

International Conference on Advances in Social Networks Analysis and Mining, ASONAM 2011, pp. 389-393.

SLIDE 27

CRICOS No. 00213J

a university for the world

real

R

Acknowledgment

My post docs, Phd students, Research

assistants

– Dr Lin Chen, Dr Slah Alsaleh, Dr Sangeetha Kutty and many others.

Industry Partner to provide us the dataset
CRC-Smart Services to provide funding