Mao Ye, Peifeng Yin, Wang-Chien Lee, Dik-Lun Lee Pennsylvania State Univ. and HKUST SIGIR 11.

Post on 26-Dec-2015

215 views 0 download

Tags:

Transcript of Mao Ye, Peifeng Yin, Wang-Chien Lee, Dik-Lun Lee Pennsylvania State Univ. and HKUST SIGIR 11.

Mao Ye , Peifeng Yin, Wang-Chien Lee, Dik-Lun Lee

Pennsylvania State Univ. and HKUST

SIGIR 11

OutlineIntroduction and MotivationModel Experiments & EvaluationConclusions

IntroductionLocation Based Social Network(LBSNs):

Foursquare, Gowalla, Brightkite, Loopt etc. Allow share tips or experience of Point-of-

Interest(POIs) e.g. restaurants, stores, cinema through check-in behaviors

Main Elements in LBSNs

MotivationRecommend new POIs to users can help them

explore new places and know their cities better In LBSNs, different from other systems, “cyber”

connections among users as well as “physical” interactions between users and locations captured in the systems, thus POIs recommendation in LBSNs is promising and interesting

The idea of incorporating the geographical influence between POIs has not been investigated previously

ModelThree important factors:

Geographical influenceUser preference of POIsSocial Influence

A fusion framework combine all three

Geographical Influencemeasures how likely two of a

user’s check-in POIs within a given distance

User power law distribution to model the check-in probability to the distance between two POIs visited by the same users:

Given user i and his check-in history Li, then

baxy

Geographical InfluenceThen for a new location lj , we have the

probability for user I to check in lj as follows:

User-based CFBased on user similarity

is the predicted check-in probability. is the similarity of user i and user k, and

computed as follows:

Friend Based CFBased on recommendation from friends

Friends have closer social tie Friends show more similar check-in bahavior

Fusion FrameworkCombine all of the three factors

Data Set

Performance MetricsMark off some POIs and the systems return

top-N recommended POIs Mainly examine below two metrics

The ratio of recovered POIs to N, precision@NThe ration of recovered POIs to the total POIs

which are marked off , recall@N

ExperimentsModel in this paper denotes as USG

U for user preference S for social influence G for geographical influence

Compared Methods User-based CF (U) : set α=β=0 Friend-based CF (S): set α = 1, β=0 GI-based (G): set α = 0, β=1 Random Walk with Restart(RWR) User preference/social influence based (US): set β=0 User preference/geographical influence based(UG): set α = 0

Tuning ParametersUser preference plays a dominate role in

contributing to the optimal recommendation Both social and geographical influence are

innegligible

Performance Comparison ResultUSG always the bestRWR may not be suitable for POI

recommendationSocial influence and geographical influence

can be utilized to perform POI recommendation

Study on Item-based CFRegard POIs as “items” and denotes as L ,

and combine it with user preference(U) and geographical influence(G)

L brings no advantage at all in enhancing U or L in POI recommendationPOIs in LBSNs not have been visited by

sufficient users

Study on Social InfluenceUser check-in behaviors and the user

similarity calculated based on RWRCheck-in behaviors and social tie strengthThe similarity in friends’ check-in behaviors

not necessarily be reflected through social tie strength

Impact of Data SparsityThe larger the mark-off ratio x is , the sparser

the user-Check-in matrix is Geographical plays an extremely important

role when data is very sparse.

Test for Cold Start UsersConsider users who have less than 5 check-

ins after mark off 30%For cold start users, user preference is hard

to capture, thus U performs bad , and as few check-ins, G also affects, and S is more useful in this situation

ConclusionsFirst incorporate geographical influence into

POI recommendationIncorporate U,S,G into a fusion frameworkExperiments conclusions

Geographical influence shows a more significant impact than social influence

RWR may be not suitable for POI recommendation, friends’ taste is different( friends have low common check-in ratio)

Item-based CF is not effective

Future WorkCombine semantic tags , e.g. location

categories such as Store, RestaurantsCombine geographical influence into Matrix

Factorization MethodTake location transition sequence into

consideration

Thanks Q&A