Recommendations in LBSN Social Networks
Section 2
Concepts of LBSN Social Networks:
new social structure made up of individuals connected by the interdependency derived from their locations in the physical world
as well as location-tagged media content
here:
The physical locations mean the instant location of an individual at a given timestamp and the location history
that an individual has accumulated in a certain preriod
The interdependency mean includes not only the superficial message that two people show up at the same time, but also some other message
like the common interests and behaviors
LBSN consists of three graphs: location-location graph, user-user graph, user-location graph
-location-location graph: relations: physical distances or similarities or some users consecively visited
-user-user graph : 1.physical distances 2.friendship relationships 3. relationships derived from their check in data
-user-location graph: starting from a user and end at a location, with the weight of the rating or the times of visits
Unique Properties of Locations
1.Hierachical: Rank of the place, country->city->Venue... The hierachical level influences the connection of two users sharing same check in data
2.Measuable Distances: three kinds of distances locations, user-location,user-user, location-location
3.Sequential Odering: Not quite understand...
Existing Challenges:
1.Location Context Awareness:
a) current location,
{
1.different recommendations need different granularity,
2.distance influence user's dicision
3.current location influence next decision
}
b)The HIstorical Locations of the user,
{
data cannot be full and complete,
constantly changing
}
c)Location history of others: social opinion
{
1.how to weigh different uesers' data according to their knowlede and experience}
2.Heterogeneous Domain
3.Rate of growth: constant changing and evolve fast
4.Cold-start problem and data sparsity
Section 3
Location recommendations:
A) Stanalone location recommendation:
a) User-profile based (content based):
match user profile with loaction meta data, do not suffer cold-start problem but poor recommendation quality
b)user-loaction history :Collabrative Filtering, steps
1.calculating similarity betweeen users
2.selecting candidate location usig user;s current loaction(!!!this is differen from product rating)
3.scoring prediction
some papers suggest solely using friends's data will be more efficiant
some find that geograohical distance impact the most
some add a personalized travel distance model, which is te biggest ompact, extending by considering general popularity
form a category-regularized matrix constructed from the user location history,thus considering both user preferences and category similarity