ido guy, naama zwerdling, inbal ronen, david carmel, erel uziel sigir ’ 10

18
Ido Guy, Naama Zwerdling, Inbal Ronen, David Carmel, Erel Uziel SIGIR’10

Upload: doris-washington

Post on 29-Dec-2015

219 views

Category:

Documents


1 download

TRANSCRIPT

Ido Guy, Naama Zwerdling, Inbal Ronen, David Carmel, Erel Uziel

SIGIR’10

OutlineIntroductionRecommender system

Recommender WidgetSocial Media PlatformRelationship AggregationUser ProfileRecommendation Algorithm

ExperimentsConclusion

2

Introduction

3

IntroductionUsers are flooded with contentHow to judge the validity of so much content?As social media grows larger everyday, these

web sites are increasingly challenged to attract new users and retain existing ones.

Contribution: Study personalized recommendation of social media items

4

Recommender systemRecommender Widget

5

Recommender SystemLotus Connections:

A social software application suiteprofiles, activities, bookmarks, blogs,

communities, files, and wikis.Recommendation platform for the system

6

Recommender systemRelationship Aggregation

SaND Models relationships through data collected across

all LC applications. Aggregates any kind of relationships between

people, items, and tags. For each user, weighted lists of PEOPLE, ITEMS

and TAGS are extracted

7

Recommender systemRelationship Aggregation

SaND builds an entity-entity relationship matrix

direct relations indirect relations

8

Recommender systemUser Profile

P(u): an input to the recommender engine once the user u logs into the system.

N(u): 30 related peopleT(u): 30 related tags

9

Recommender systemUser Profile

Person-person relations Aggregate direct and indirect people-people

relations into a single person-person relationship. Each direct relation adds a sore of 1. Each indirect relation adds a score in the range of

(0,1].

10

Recommender systemUser Profile

User-tag relations used tags

direct relation based on tags the user has used incoming tags

direct relation based on tags applied on the user indirect tags

indirect relation based on tags applied on items related on the user

11

Recommender systemTag Profile Survey – participants are asked to

evaluate tags as indicators of topic of interest

Combination of used and incoming tags is the best indicator to generate T(U) from SaND system

12

Recommender systemRecommendation Algorithm

d(i): number of days since the creation date of iw(u,v) and w(u,t): relationship strengths of u to

user v and tag tw(v,i) and w(t,i): relationship strengths

between v and t, respectively, to item i

13

Recommender systemRecommendation Algorithm

User-item relation: authorship (0.6), membership (0.4), commenting (0.3), and tagging (0.3)

Tag-item relation: number of users who applied the tag on the item, normalized by the overall popularity of the tag.

14

Evaluation5 recommenders

PBR: β=1TBR: β=0or-PTBR: β=0.5and-PTBR: β=0.5POPBR: popular item recommendation.

Each participant is assigned to one recommender

15

EvaluationRecommended Items Survey

16

EvaluationRecommended Items Survey

17

ConclusionThe combination of directly used tags and

incoming tags produces an effective tag-based user profile.

Using tags for social media recommendation can be highly beneficial.

Combining tag and person based recommendations perform better.

Future Work:Large scale evaluationComputationally intensive algorithm may be used.

18