GeoFeed: A Location-Aware News Feed System

GeoFeed: A Location-Aware News Feed System
GeoFeed: A Location-Aware News Feed System
Jie Bao1
1 Department
Mohamed F. Mokbel1
Chi-Yin Chow2
of Computer Science and Engineering, University of Minnesota, Minneapolis, MN, USA
of Computer Science, City University of Hong Kong, Kowloon, Hong Kong
{baojie,mokbel}@cs.umn.edu, [email protected]
2 Department
Abstract—This paper presents the GeoFeed system; a locationaware news feed system that provides a new platform for its
users to get spatially related message updates from either their
friends or favorite news sources. GeoFeed distinguishes itself from
all existing news feed systems in that it enables users to post
message with spatial extent rather than static point locations,
and takes into account their locations when computing news feed
for them. GeoFeed is equipped with three different approaches
for delivering the news feed to its users, namely, spatial pull,
spatial push, and shared push. We design a smart model for
GeoFeed to decide about using these approaches in a way that:
(a) minimizes the system overhead for delivering the locationaware news feed, and (b) guarantees a certain response time
for each user to obtain the requested location-aware news feed.
GeoFeed also supports location-aware news feed function for its
mobile users. Experimental results, based on real and synthetic
data, show that GeoFeed outperforms existing news feed systems
in terms of response time and maintenance cost.
I. I NTRODUCTION
Social networking systems, e.g., Facebook [6] and Twitter [27], and news aggregators, e.g., My Yahoo! [16] and
iGoogle [12], are among the most popular web services nowadays. A common functionality shared by such web services is
the news feed functionality, where users of social networks and
news aggregators receive a set of news from their friends and
favorite news sources, respectively. Due to the large volume of
related news for each user, existing news feed systems opt to
select a subset of k relevant news either based on the message
timestamp, i.e., most recent k messages, or based on some
diversity requirements. Unfortunately, such a selection ignores
the spatial aspect of related messages, and hence, users may
miss several important messages that are spatially related to
them either because they are not so recent or do not satisfy
diversity requirements. For example, when a traveling user
logs on to a social network site, the user would like to get the
news feed that match his/her new location, rather than sticking
to the most recent news feed. The same concept can also be
applied for users who keep logging on to the system from the
same location, yet, they have a large number of friends. It is
of essence for such users to limit their news feed to the ones
related to their locations.
In this paper, we present GeoFeed; a location-aware news
feed system that provides a new platform for its users to get
spatially related message updates from either their friends or
favorite news sources. GeoFeed complements the functionality
This work is supported in part by NSF under Grants IIS-0811998, IIS0811935, CNS-0708604, IIS-0952977, a Microsoft Research Gift, and grants
from the City University of Hong Kong (Project No. 7200216 and 7002686).
of existing social networks and news aggregators to make them
location-aware. Once a user u logs on to her favorite social
network site that is equipped with GeoFeed, u will find the
set of messages that are more relevant to her current location,
e.g., a message about local news, a comment about a local
store, or a status message targeting friends in a certain area.
For a user u that has a set Fu of N friends (in a social
network context) or follows a set Fu of N news sources (in
a news aggregator context), GeoFeed abstracts the locationaware news feed problem to evaluating a set Qu of N locationbased queries posed by u. Each query qi ∈ Qu is posed to
a friend fi ∈ Fu to retrieve the set of messages that are
issued by fi and overlap with u’s range of interest. u’s range
of interest could be the exact location of u, in which the
location-based queries are point queries, or a range around
u, in which location-based queries are range queries, e.g., get
all the messages posed by my friends within r miles form my
location. To limit the set of messages delivered to u, GeoFeed
gets only k messages from each friend fi ∈ Fu . In the mean
time, GeoFeed guarantees that each user u will get all the
requested news feed within a response time threshold Tu .
GeoFeed is equipped with three different approaches for
evaluating each query qi ∈ Qu , namely, (1) spatial pull approach, in which qi is answered through exploiting a spatial index over the messages of friend fi , (2) spatial push approach,
in which qi just retrieves the answer from a pre-computed
materialized view maintained by friend fi , and (3) shared push
approach, in which the pre-computation and the materialized
view maintenance at friend fi are shared among multiple users
that include u. Then, the main challenge of GeoFeed is to
decide on when to use each of these three approaches to
which queries. GeoFeed is equipped with an elegant decision
model that decides about using these approaches in a way that:
(a) minimizes the system overhead for delivering the locationaware news feed, and (b) guarantees a certain response time
Tu for each user u to obtain the requested location-aware
news feed. A better response time calls for using the spatial
push approach for all queries, where all news feed are precomputed. However, this results in a huge system overhead
to maintain a massive number of materialized views, hence
limit the scalability of the system to support more users. In
contrast, favoring system overhead may result in evaluating
more queries using the spatial pull approach as less views are
maintained. However, users with large numbers of friends will
suffer a significantly long delay when retrieving their news
feed. GeoFeed takes these factors into account when deciding
on which approach to use to evaluate each query qi in a way
that minimizes the system overhead, i.e., supports more users,
and guarantees a response time threshold.
A distinguishing characteristic in GeoFeed is that it builds
its decision model for each single query qi instead of the whole
system or the set of queries Qu for a given user. This means
that for a certain user u that has two friends fi and fj , GeoFeed
may opt to retrieve the messages from fi through the spatial
push approach while retrieving the messages from fj through
the spatial pull approach. Similarly, for a certain user f that
feeds two users ui and uj , GeoFeed may opt to have ui retrieve
her messages from f with the spatial push approach while uj
retrieve her messages from f with the spatial pull approach.
Extensive experimental results, based on real and synthetic
data, show that (a) GeoFeed is favorable over existing news
feed systems, and (b) the accuracy of the GeoFeed decision
model in guaranteeing user response time while minimizing
the total system overhead.
The closest work to ours is the feeding frenzy approach [23],
which is a news feed system equipped with pull and push
approaches to retrieve the most recent news feed items.
Unfortunately, the feeding frenzy system cannot be directly
applied to the location-aware news feed problem as it does
not consider the message location aspect at any of its stages.
The only way to turn feeding frenzy to be location-aware is to
attach a wrapper around it in a form of a spatial filter, which is
extremely inefficient as the spatial filter is applied afterthought.
Our proposed location-aware news feed system, GeoFeed,
distinguishes itself from feeding frenzy [23] and other systems
in having all the following aspects: (1) GeoFeed is designed
while having the location-awareness in mind, and thus makes
use of the spatial extents of each message to early prune nonrelevant messages, (2) GeoFeed modifies the traditional pull
and push approaches to support spatial filters, (3) GeoFeed
goes beyond the traditional pull and push approaches, and
introduces the shared push approach as a third alternative
that reduces the system overhead while not sacrificing the
user response time, (4) GeoFeed builds its decision model to
minimize the system overhead to support more users, while
guaranteeing a certain user response time threshold for each
user, (5) GeoFeed gives the message issuer the ability to
determine the spatial validity of the each posted message, e.g.,
the weather service provider may decide a tornado warning is
relevant only to followers residing in a certain area.
The rest of the paper is organized as follows: Sections II
and III give related work and overview of GeoFeed. Section IV
discusses the spatial pull, spatial push, and shared push approaches. The cost and decision models of GeoFeed are given
in Sections V and VI. Section VII discusses GeoFeed with
mobile users. Experimental results are given in Section VIII.
II. R ELATED W ORK
This section highlights related work to GeoFeed in two main
areas; news feed systems and location-aware social networks.
News feed systems. Most of existing news feed systems work
in a similar way to publich/subscribe services, e.g., [3], [4],
Location-based
Social Network
Facebook Places [6]
Renren [20]
Sina Weibo [24]
Loopt [14]
Google Buzz [10]
Foursquare [8]
Twinkle [26]
GeoFeed
Location
Tag
√
√
√
√
√
√
√
√
Range
Queries
Spatial
Messages
√
√
√
√
√
√
TABLE I
A TAXONOMY OF L OCATION - BASED S OCIAL N ETWORKS
[32], which use a push approach to fan out the message notices
to all their users. However, such systems are not applicable
to address the location-aware news feed, as (a) they do not
consider the spatial relevance of each message, and (b) using
the push approach does not scale up for large number of
publishers and subscribers as it is the case for social networks.
For commercial systems, our only knowledge is about the
Feeding Frenzy system [23] from Yahoo!, which we consider
as our closest work and compare with in Section VIII. The
main idea of Feeding Frenzy is to build a cost model for
deciding upon using the pull or push approaches as means
for retrieving the news feed for a registered user. The only
way to use the Feeding Frenzy system for the location-aware
news feed problem is to attach a wrapper around it to filter
any spatially irrelevant message from the users’ news feed.
However, that would be very inefficient as the spatial filter
is applied as an afterthought solution. Our proposed system,
GeoFeed, distinguishes itself from Feeding Frenzy in that it is
built with the location-awareness functionality in mind. Thus,
the query evaluation methods, the cost model, and decision
model take into account the spatial aspect of each posted
message along with the location of each user.
Location-aware social networks. Existing commercial
location-based social networks fall in two categories, as summarized in Table I. The first category includes Facebook
Places [6], Renren [20], and Sina Weibo [24], where they
consider the location information of the message issuer as just
an additional tag attached with the message. Then, a system
user will get the same news feed (associated with location
tags) regardless of the user location. The second category
includes Loopt [14], Google Buzz Mobile [10], Foursquare [8],
and Twinkle [26] where, in addition to having the location
tags, they also give their users the ability to issue range
queries to view the whereabouts of their friends. GeoFeed
distinguishes itself from all these commercial products in two
main aspects: (1) GeoFeed gives its users the ability to set
the spatial validity range of each posted message., and hence
give control to the message issuer to decide who should get the
posted information. For example, the weather service provider
may decide a tornado warning is relevant to followers located
only in a certain area. (2) Unlike all existing systems that are
built mainly to be used by mobile devices, GeoFeed offers
a more flexible way for the users to share their geo-tagged
messages. Users of GeoFeed can access their account in the
same way they use Facebook, yet, they will retrieve more
relevant location-aware news feed than that of Facebook users.
0HVVDJH 7LPHVWDPS 6SDWLDO
&RQWHQW
0
6
%DFNWRKRWHO
0
6
$QLFHEDU
0
6
(DWLQJDWDEDU
6
6
6
D/RFDWLRQEDVHGPHVVDJHVSRVWHGE\XVHU$OLFH
0HVVDJH 7LPHVWDPS 6SDWLDO
&RQWHQW
0
6
+DYLQJ FRIIHH
0
6
$QDFFLGHQW
0
6
:RUN ILQLVKHG
E/RFDWLRQEDVHGPHVVDJHVSRVWHGE\XVHU%RE
Fig. 1.
6
&DURO 6
6
F6SDWLDODUHDVRIWKHORFDWLRQEDVHGPHVVDJHV
Location-based messages and news feed.
On the other side, related research prototypes in locationbased social networks have either focused on: (a) message
sharing [1], [2], where users can broadcast or receive public
location-based messages, yet with no social awareness, i.e.,
there is no concept of friendship. Applying techniques form
message sharing to the news feed problem is equivalent to
having all queries evaluated with the (spatial) pull approach,
which is very inefficient, (b) Privacy-aware search queries [9],
[13], [21], [22], which enables private query search over
users’ friends, with no interest of the location-aware news feed
functionality, or (c) location-aware recommender systems [17],
[29], [30], [31], which suggests new places for their users.
The functionality of recommender systems is fundamentally
different from news feed systems, where the main goal is
predict what the user would like rather than delivering the
news feed from posted messages of users’ friends.
III. G EO F EED : S YSTEM OVERVIEW
This section gives an overview of GeoFeed as follows:
Location-based messages. GeoFeed users can either send or
receive location-based messages to or from their friends with
a spatial extent. A location-based message is represented by
the tuple: (MessageID, Content, Timestamp, Spatial), where
MessageID and Content represent the message identifier and
contents, respectively, Timestamp is the message generation
time, and Spatial indicates the spatial extent of the message.
Figures 1a and 1b give examples of six messages, as messages
M2 , M3 , and M5 issued by user Alice and M1 , M4 , and M6
issued by user Bob . The spatial extents of the six messages
are represented by the areas S1 to S6 in Figure 1c.
System users. A user u, located in u.location with range
of interest u.range has a list of friends Fu , who are also
considered to be system users, can log on to GeoFeed, and:
(a) read all the messages posted from any friend fi ∈ Fu
where u.range overlaps with the message spatial extents,
and/or (b) post a new message M that should be broadcasted
to all friends in Fu that have range of interest overlap with the
spatial extents of M . Each user u has a response time threshold
Tu where GeoFeed guarantees to provide the location-aware
news feed for u within Tu . The value of Tu for each user could
be set by the system as either a default value for all users, or
a value that reflects how valued and appreciated is the user.
Location-based news feed queries. GeoFeed abstracts the
location-aware news feed functionality for a user u to a set
Qu of location-based queries posed to the user’s friends Fu .
Each query qi ∈ Qu is posed to a friend fi ∈ Fu to retrieve
the k most recent spatially relevant messages to u’s location.
Then, u may opt to get all the received messages, or select
only the k most recent and spatially relevant messages from all
users. Similar to the feeding frenzy system [23], we will only
focus on retrieving k message from each user, as an additional
filter will be a trivial step. As an example, consider user Carol,
depicted by a triangle in Figure 1c, which is a friend of Alice
and Bob. As the spatial range of interest of Carol includes only
her location, she issues two location-based point queries, with
k = 2, one for Alice (returns M3 and M5 ) and one for Bob
(returns M4 and M6 ). Carol can add an additional filter to get
the most recent two messages, i.e., M5 and M6 . The locationbased news feed query can also use spatial range instead of
the point location to retrieve the most recent k messages that
overlap with the querying range.
Query evaluation methods. GeoFeed is equipped with three
different approaches for evaluating each query qi ∈ Qu ,
namely, spatial pull, spatial push, and shared push approaches.
Details of these approaches will be discussed in Section IV.
Problem formulation. The decision model of GeoFeed can
be formulated as follows: For each location-based news feed
query q posed by a user u, find out the best approach among
spatial pull, spatial push, and shared push approaches, to
evaluate q once u logs on to the system in a future time, such
that: (a) the GeoFeed computational overhead for all system
queries is minimized, and (b) the response time that u will
encounter to get all the requested location-aware news feed is
within the required threshold Tu .
IV. G EO F EED Q UERY E VALUATION
In this section, we present three different query evaluation
approaches, spatial pull, spatial push, and shared push, to
evaluate a location-based news feed query posed from user
u, located at u.location, to a friend fi ∈ Fu , where Fu is the
list of u’s friends. GeoFeed employs these three approaches,
monitors their cost (Section V), and then uses its decision
model (Section VI), to decide on which approach should be
used for which queries. The section starts by discussing the
underlying data structure, then it goes on to the details of
each approach. For simplicity, and without loss of generality,
we use location-based point queries where the user range of
interest is limited to the user location rather than a spatial area
around the user location.
A. Data Structure
To support its three query evaluation approaches, GeoFeed
maintains typical information for each user u that includes
ID, name, and location. In addition, GeoFeed maintains the
following two data structures for each user u.
List of friends. Each user u maintains a list of friends Fu ,
where each friend fi ∈ Fu is just another system user that
has ID, name, location/querying spatial ranges, list of friends,
and gird structure.
Grid index structure. Each user u maintains a grid index
structure Gu that consists of n×n equal area grid cells, as it is
efficient with frequent message updates and works seamlessly
for both messages and friends’ querying spatial range. Each
ORFDWLRQEDVHG
QHZVIHHGTXHU\
$OLFH
ORFDWLRQEDVHG
QHZVIHHGTXHU\
*ULGLQGH[
$OLFH
JHW
FHOO
$OLFHಬV
ORFDWLRQ
ORFDWLRQ
EDVHG
0HVVDJHV
QHZVIHHG 6SDWLDO
ILOWHU
PHVVDJHVLQWKHFHOO
D6SDWLDOSXOODSSURDFK
Fig. 2.
0DWHULDOL]HG
YLHZ
LQVHUW
ORFDWLRQEDVHG
QHZVIHHG
$OLFH
2WKHUXVHUಬV
LQVHUW
PDWHULDOL]HG
YLHZV
ORFDWLRQEDVHG
QHZVIHHGTXHU\
ORFDWLRQ
EDVHG
1HZ
PHVVDJH
NPHVVDJHV
6
2WKHU
XVHUV
UDQJH
TXHU\
QHZVIHHG
2WKHU
XVHUV
)LOWHU
UDQJH
TXHU\
6KDUHG
PDWHULDOL]HG
YLHZ
ORFDWLRQEDVHG
QHZVIHHGTXHULHV
E6SDWLDOSXVKDSSURDFK
1HZ
PHVVDJH
XSGDWH
F6KDUHGSXVKDSSURDFK
Three location-based news feed evaluation approaches in GeoFeed.
grid cell C ∈ Gu maintains: (1) The set of IDs and spatial
ranges of u’s friends, whose querying ranges overlap with C,
and (2) The set of message IDs produced from u with spatial
extents that overlap C. If a message overlaps with several grid
cells, its ID will be stored in every grid cell it overlaps.
B. Spatial Pull Approach in GeoFeed
The spatial pull approach takes advantage of the grid index
structure Gfi maintained at friend fi for early spatial pruning
at the friend’s side. This is the main distinguishing difference
between the spatial pull approach employed in GeoFeed and
the traditional pull approach employed in feeding frenzy [23].
With the early pruning pushed inside the spatial pull approach,
GeoFeed avoids retrieving unnecessary messages as will be
encountered in the traditional pull approach.
Figure 2a gives an example of the main idea of the spatial
pull approach where user Alice needs to get her k spatially
relevant messages from her friend Bob. The execution flow
goes as follows: (1) Alice submits her location and the news
feed query to Bob. (2) Bob exploits his grid index structure
GBob to find out cell C that includes Alice’s location, and
retrieves all the messages stored in C. (3) Bob applies a
spatial filter over all the messages returned from GBob to only
report those messages that include Alice’s location (depicted
by a black triangle), as there could be messages in C that
do not overlap Alice location. (4) If more than k messages
are returned from the spatial filter, Bob forwards the k most
recent ones to Alice as part of her news feed. Alice will get
the rest news feed from the other friends.
C. Spatial Push Approach in GeoFeed
The spatial push approach in GeoFeed pre-computes and
stores the answer of the location-based news feed query in
a materialized view maintained by the friend fi . Then, once
the user u logs on, u only retrieves the news feed from the
materialized view. Although the spatial push approach is very
appealing to the user, it poses a large overhead over the
system resources to continuously maintain the materialized
view while the user is offline. Same as in the case of the
spatial pull approach, the spatial push approach is distinct
from the traditional push approach, used in the feeding frenzy
system [23], in that it exploits the grid index structure Gfi
maintained at friend fi to significantly reduce the overhead of
pushing irrelevant messages to the materialized view.
Figure 2b gives an example of the main idea of the spatial
push approach in GeoFeed where user Alice needs to get
her k spatially relevant messages from Bob. The spatial push
approach consists of two orthogonal parts, namely, query processing and view maintenance, detailed as follows: (1) Query
processing. Alice registers her location with the grid index
structure maintained by Bob, GBob . In turn, Bob deals with
Alice location as a continuous location-based point query [5],
[15], [19] that needs to be maintained, in a materialized view,
even if Alice goes offline. Once Alice logs on to GeoFeed,
she just probes her materialized view, maintained by Bob, to
retrieve her news feed. (2) View maintenance. The friend Bob
uses his grid index structure, GBob , to maintain a materialized
view for each user employing the spatial push approach to
retrieve his/her messages from him. Any new message M
produced from Bob generates a range query over the grid
index, GBob , to retrieve the locations of all friends of Bob
whose querying ranges overlap the spatial range of M and use
the spatial push approach to retrieve their news feed from Bob.
For each of these friends, M is forwarded to the designated
materialized view. Figure 2b shows Alice’s location in GBob
as a black triangle. The message M produced from Bob is
depicted as a shaded circle over GBob , which updates the
materialized view of Alice among other views of those users
who retrieve their news feed from Bob with spatial push.
D. Shared Push Approach in GeoFeed
The shared push approach in GeoFeed is designed to take
advantage of the users locality to reduce the system overhead
of the spatial push approach while only slightly increasing
the response time of location-based news feed queries. The
main idea of the shared push approach is to maintain one
materialized view shared among different spatially co-located
users within one grid cell. The benefit is that GeoFeed will
maintain much less views than the spatial push approach. On
the negative side, the messages returned by the shared view
need an additional filter to split the answer among the shared
views, imposing little overhead over the spatial push approach.
Figure 2c gives an example of how Alice gets her k relevant
messages from her friend Bob, with the shared push approach.
The scenario is very similar to the spatial push approach with
the following two differences in query processing and view
maintenance: (1) Query processing. As several users share
the same materialized view with Alice, Bob needs to add an
additional filter to filter out those news items that are not
relevant to Alice. (2) View maintenance. Instead of maintaining
one materialized view for each friend, Bob maintains one
shared materialized view for all friends with the spatial push
approach in each cell. Thus, a new message coming out from
Bob will be inserted in much less materialized views.
The choice of which users to share views together will be
discussed in Section VI as part of the GeoFeed decision model.
V. G EO F EED C OST M ODEL
This section builds cost models for the spatial pull, spatial
push, and shared push approaches in GeoFeed. The cost
models are designed to measure: (a) The system overhead
encountered by GeoFeed to process a location-based news feed
query from a user u to a friend fi , and (b) The user response
time taken by GeoFeed to prepare the location-aware news
feed for a user u from a friend fi . These cost models will be
used in the GeoFeed decision model (Section VI) to decide on
what would be the best approach to evaluate each news feed
query in order to: (a) minimize the GeoFeed system overhead,
i.e., providing the ability to scale up GeoFeed to support more
users, and (b) ensure that each user u will obtain the news feed
within a threshold time Tu . It is important to note that both
the GeoFeed cost and decision models for any query issued
by user u are designed when u is offline. Once u becomes an
online user, GeoFeed sends the initial news feed to u based
on the selected approach. As u remains online, new relevant
messages to u are just pushed to u as in the spatial push
approach. So, we ignore any maintenance cost for online users
as it will be the same for all approaches. We will first discuss
the required data structures to monitor the cost models. Then,
we present the cost model.
A. Data Structure
In order to monitor the cost models, GeoFeed maintains a
new data structure, termed Stat Table, as follows:
Stat Table. This is a global statistics table that includes the
following four statistics for each user u: (1) Total number of
messages (Nu ). This is the total number of messages produced
from u since joining the GeoFeed system; updated with
every message posted from u. (2) Response time requirement
(Tu ). This is the time threshold set for each user as a hard
requirement for GeoFeed to produce the location-aware news
feed for u within Tu . The value of Tu could be set by the
system as either a default value for all users, or a value that
reflects how valued and appreciated is the user. (3) Update
frequency (UF u ). This is the average number of messages
produced from u per hour; can be either computed from the
time that u has joined GeoFeed, over the last hour, or as a
weighted average over a certain time window. (4) Predicated
offline time (OT u ). This is set once u logs off the system, and
it reflects the predicted offline time (in hours) before u logs
on again to the system. OT u can be either computed as an
average offline time based on user history or similar to last
observed offline time.
In addition, GeoFeed adds additional fields to the already
maintained data structures, list of fiends and grid structure:
List of friends (Fu ). For each friend fi ∈ Fu , we maintain
the total number of messages produced from u to fi , Nu→fi ,
since u has joined GeoFeed.
Grid index structure (Gu ). With each grid cell C ∈ Gu , we
maintain the number of messages Nu→C that are produced
from user u and overlap with cell C since u, has joined
GeoFeed. This value is updated with each message produced
from u.
B. The Spatial Pull Approach
The cost model for the spatial pull approach is developed
for each news feed query issued from user u to her friends.
System overhead. As the spatial pull approach simply executes a query from u to exploit a grid index structure Gfi , we
assume that the system overhead for the spatial pull approach
is a constant cost, CostP ull , regardless of the user u and the
friend fi . This is mainly because every spatial pull query in
GeoFeed will go through the same procedure, and hence they
all have the same cost.
User response time. As everything in the spatial pull approach is done only when the user logs on to the system, the
user response time will be the same as the system overhead to
evaluate a location-based news feed query, which is CostP ull .
C. The Spatial Push Approach
Similar to the case of the spatial pull approach, the cost
model for the spatial push approach is developed for each
news feed query issued from user u to a friend fi ∈ Fu .
Since, as we will see, this cost model depends on the number
of messages produced from fi while u is offline, the cost of
the spatial push approach needs to be reevaluated every time
u logs off the system. This is mainly because the GeoFeed
decision model is all about what would be the best approach
to evaluate u’s query the next time u will log on to the system.
Based on the decision, GeoFeed will decide whether to start
maintaining a materialized view for u at fi till u logs on the
next time (i.e., use the spatial push approach), or simply do
nothing for u till the next log on time (i.e., use the spatial pull
approach).
System overhead. The total cost encountered by GeoFeed
to support the spatial push approach, CostP ush , is the summation of two parts: (1) CostQuery : The cost to answer the
user query issued from u to fi , and (2) CostV iew : The cost
of maintaining the materialized view at fi to serve u’s query,
while u is offline. The details of these two costs are as follows:
1) CostQuery . This cost only includes retrieving the news
items from the maintained materialized view. We denote
this part of the cost as CostRV iew , which is a constant
value as it is a simple selection operation, regardless of
u and fi .
2) CostV iew . This is the cost of the background processing,
done by GeoFeed, to maintain a materialized view for
the messages coming from fi and including u’s location.
CostV iew can be computed as the multiplication of
the following two terms: (a) the cost of inserting a
message posted by fi in the materialized view for u,
which is a constant cost, CostIV iew , regardless of fi
and u, and (b) the number of such messages posted
from fi while u is offline. This can be computed as
Nfi →u
Nfi →u
Nfi ×UF fi ×OT u , where Nfi represents the ratio
of messages produced from fi to u to the total number
of messages produced from fi , UF fi is the frequency
of messages coming from fi (the number of messages
per hour), and OT u is the predicted offline time for the
user u (in hours). Notice that UF fi , Nfi , and OT u are
stored in the stat table while Nfi →u is stored in the
friend list of fi .
Then, the cost of the spatial push approach, CostP ush , is:
Nf →u
CostP ush = CostRV iew +CostIV iew × i ×UF fi ×OT u
Nfi
User response time. Once u logs on to GeoFeed to retrieve
the news feed from fi through the spatial push approach,
GeoFeed simply returns the already maintained materialized
view, which has the same cost as the first part of the system
overhead, CostRV iew .
D. The Shared Push Approach
Unlike the cost models for the spatial pull and spatial push
approaches that were developed for a particular query issued
from user u to a friend fi , and reevaluated every time u logs
off GeoFeed, the cost model for the shared push approach is:
(a) developed for a set of queries QC posed from different
users in the same grid cell C to the same friend fi , and (b) as
the queries that share the view are from multiple users, the
shared push approach is reevaluated every time any user of
this shared view logs off.
System overhead. Similar to the spatial push approach, the
total cost of the shared push approach, CostShared , is the summation of two main parts: (1) CostSQuery : The cost to answer
the user queries using the shared view, and (2) CostSV iew :
The cost of maintaining the shared materialized view at fi .
The details of these two costs are as follows:
1) CostSQuery . This is similar to the case of the spatial
push approach with two differences: (1) The cost for
retrieving the query answer from the shared materialized
view, denoted as CostRSV iew , which is slightly higher
than the spatial push approach, due to an additional filter
over the results returned from the shared materialized
view. (2) As the cost of the shared push approach is
computed for a set of queries QC , the query cost needs
to be multiplied by the number of queries that share the
view |QC |. Thus, CostSQuery = |QC | ×CostRSV iew .
2) CostSV iew . In a very similar way to the case of
the spatial push approach, this cost can be comNfi →C
C)
× MBR(Q
puted as: CostISV iew × N
Area(C) × UF fi ×
fi
min∀u∈QC (OT u ). This includes the following three
differences from the case of the spatial push approach:
(1) We use CostISV iew instead of CostIV iew to reflect
the cost to update one message to the shared view.
Nfi →C
Nfi →u
C)
(2) We use N
× MBR(Q
as
Area(C) instead of
Nfi
fi
the ratio of the messages produced from fi in cell C to
the total number of massages produced from fi , multiplied by the ratio of the covered area of the minimum
bounding rectangle (MBR) of the point queries sharing
the same materialized view to the area of the grid cell C.
(3) We use min∀u∈QC (OT u ) instead of OT u to reflect
the time when the first user of the shared view in cell
C logs off the system, in which the cost model needs
to be reevaluated.
Then, the cost of the shared push approach, CostShared , is:
CostShared = |QC | × CostRSV iew + CostISV iew
M BR(QC )
Nfi →C
× UF fi × min (OT u )
×
×
∀u∈QC
Nfi
Area(C)
User response time. Similar to the case of the spatial push
approach, the cost here will be to only read from the shared
materialized view, which is CostRSV iew .
VI. G EO F EED D ECISION M ODEL
In this section, we discuss the GeoFeed decision model that
is applied for each query posed by a user u and is triggered
every time u logs off from GeoFeed. The goal of this decision
model is to find out the best approach, among the spatial pull,
spatial push, and shared push approaches, to evaluate each
news feed query issued from user u to a friend fi ∈ Fu , for
the next time that u will log on to the system, i.e., after OT u
time units. The objective is to minimize the system overhead
encountered by GeoFeed while ensuring that user u will get all
the requested news feed within the time threshold Tu . Once u
logs on to the system, GeoFeed issues |Fu | news feed queries
as one for each friend fi ∈ Fu . Each query will be evaluated
using the approach that was selected by the GeoFeed decision
model. As u remains online, GeoFeed keeps pushing new
messages from any of u’s friends to u as news feed. Once
u logs off the system, the decision model will be reevaluated
again based on the new expected OT u .
The GeoFeed decision model consists of three main steps,
detailed in the rest of this section: (1) Step 1, Response time
guarantee, finds out the maximum number of spatial pull
queries, N QP ullu, that user u can afford while having the
news feed within Tu , (2) Step 2, Pull vs. Push selection,
decides on what are these N QP ullu , out of all the queries
posed by u that will be assigned to the spatial pull approach,
other queries will be assigned to the spatial push approach,
and (3) Step 3, Refinement with shared push, finds out if using
the shared push approach for any grid cell C maintained at
friend fi can reduce GeoFeed system overhead.
A. Step 1: Response Time Guarantee
Objective. Since the spatial pull queries are mostly favorable
to the system due to their lower overhead, yet, they result in
high response time, this step finds out the number N QP ullu
as the maximum number of spatial pull queries that u can
afford to guarantee that the news feed will be delivered to u
within time threshold Tu .
Main idea. In terms of user response time, the spatial push
approach gives much lower response time than that of the
spatial pull approach, i.e., CostRV iew CostP ull , regardless
of the user u and the friend fi . In the mean time, the user may
not actually feel the difference between the spatial push and
the shared push approaches as their user response times are
very close to each other, though the latter is slightly higher.
Thus, from the user perspective, the user would always like
to avoid using the spatial pull approach and have all the
queries evaluated with either the spatial push or shared push
approaches. However, a large number of views introduces
significant system overhead to maintain them. To balance
between these two contradicting factors, GeoFeed uses the
response time requirement Tu for each user u to decide that
u can tolerate having some queries evaluated with the spatial
pull approach while the overall response time for all queries
is less than Tu .
Algorithm. Since user u has to issue |Fu | location-based news
feed queries, we want to find the maximum number of queries
N QP ullu among |Fu | that can be evaluated with the spatial
pull approach while ensuring that u will get the required news
feed within the time threshold Tu . If N QP ullu queries will
be evaluated using the spatial pull approach with response
time CostP ull , then the rest of queries posed by user u, i.e.,
|Fu | − N QP ullu, will be evaluated through either the spatial
push or shared push approach. Given that these two latter costs
are very close, with CostRSV iew is slightly higher, then the
following inequality should hold for any user u:
N QP ullu×CostP ull +(|Fu |−N QP ullu)×CostRSV iew < Tu
Thus, we can get N QP ullu using the following equation:
Tu − |Fu | × CostRSV iew
N QP ullu =
CostP ull − CostRSV iew
B. Step 2: Pull vs. Push Selection
Objective. Now that GeoFeed finds out that u can afford
having N QP ullu queries with the spatial pull approach, it is
the objective of this step to decide which N QP ullu queries
out of the total of |Fu | queries that will be evaluated using
the spatial pull approach. The decision is taken to minimize
GeoFeed system overhead.
Main idea. The main idea of this step is to employ the
following two concepts: (1) Two queries from the same user u
to friends fi and fj may have different system overhead costs
fi .P ush and fj .P ush for the spatial push approach based
on the update frequency and spatial distribution of messages
coming from fi and fj . Assuming that fi .P ush > fj .P ush,
and that u can afford having only one query with the spatial
pull approach, then, GeoFeed will select the spatial push
approach for fj , as it has lower system overhead, leaving fi
to be evaluated with spatial pull. Thus, if u would accept
having N QP ullu queries with the spatial pull approach, then
the main idea here is to select the set of N QP ullu queries that
are the worst in terms of system overhead in spatial push to be
assigned to the spatial pull approach. All other queries will be
assigned to the spatial push approach. The main reason here is
that from the user perspective, it does not really matter which
queries will be selected as spatial pull, while this decision has
significant impact on the system overhead. (2) In some cases,
using the spatial push approach for a certain query may have
less system overhead than using the spatial pull approach.
Consider, for example, a user u with an expected very short
offline time OT u . In this case it is better for GeoFeed to
incrementally maintain a view of the last reported answer for
u for a short time rather than reevaluating u’s query with
Algorithm 1 Step 2: Pull vs. Push Selection
Input: (1) N QP ullu ; the number of queries with the spatial pull approach that u can
afford, and (2) The data structure used for computing the cost model.
Output: Set the decision fi .Decision for each query posed from u to a friend fi ∈
Fu to either spatial pull or spatial push.
1: for each friend fi ∈ Fu do
Nf →u
2:
fi .P ush ← CostRV iew + CostIV iew × Ni
× U F fi × OT u
fi
3:
Insert fi in a Max Heap MH based on fi .P ush
4: end for
5: PullCount ← 0
6: fm ← The top element from the heap MH
7: while PullCount < N QP ullu AND fm .P ush > CostP ull do
8:
Remove fm from the Max Heap MH
9:
fm .Decision ← spatial pull
10:
fm ← The top element from the heap MH
11:
PullCount ← PullCount + 1
12: end while
13: for each remaining fh in the heap MH do
14:
fh .Decision ← spatial push
15: end for
the spatial pull approach. This will be also favorable to the
user for the much lower user response time. Thus, it may be
desirable to have less than N QP ullu queries using the spatial
pull approach, if there are queries among the worst N QP ullu
that still have lower system overhead when using the spatial
push approach.
Algorithm. Algorithm 1 gives the pseudo code for Step 2 in
the GeoFeed decision model. The input to the algorithm is
the number of queries N QP ullu that u can afford having
in the spatial pull approach, along with the data structure
used to compute the cost model. The algorithm starts by
calculating the system overhead cost for using the spatial
push approach for each query posed by user u to a friend fi .
While calculating the cost, we insert u’s friends in a maximum
heap structure ordered by the calculated cost. Then, we keep
removing friends from the maximum heap one by one, and
assign them to the spatial pull approach, till any one of these
two stopping conditions takes place: (a) the number of queries
with the spatial pull approach has reached its maximum, which
is N QP ullu . In this case, assigning more spatial pull queries
will increase the user response time for u to be more than the
threshold Tu , or (b) we find a query that has lower system
overhead cost with the spatial push approach than the cost
of the spatial pull approach. In this case, it is better for
both the system and the user to have this query, and all
subsequent queries, with the spatial push approach. Note that
all subsequent queries will have a lower spatial push overhead
cost than the current query, and hence will also favor the
spatial push approach. Once any of these two conditions is
satisfied, we assign all the remaining queries in the maximum
heap data structure to the spatial push approach.
C. Step 3: Refinement with Shared Push
Objective. Up to now, we have an initial decision for each
query posed from u that satisfies the requirement OT u . In
this step, we look further for all queries with the spatial push
approach, and find out if some of them can be grouped together
to use the shared push approach, and hence amortize the cost
of maintaining a materialized view among several queries.
As the response times of the spatial push and shared push
approaches are similar, and we have already used the shared
push cost in Step 1, the refinement step will never break the
response time requirement for any user.
Main idea. The main idea here is to go through all the grid
cells of each user in the system, and check if the shared push
approach will be useful. Notice that u is located in |Fu | grid
cells as one per each friend fi ∈ Fu . In any of these grid
cells, using the shared push approach may be clearly favorable
if most of the users in this cell are using the spatial push
approach. In this case, the system overhead to maintain the
shared view is amortized over these users. Unfortunately, there
is no magic number for the number of spatial push queries
in a cell that would call for going towards the shared push
approach, as this depends on several factors that include the
first offline time for any user in the cell, the frequency of users’
updates, and the size of the minimum bounding rectangle of
the shared view. So, a full computation of the gain/loss from
using the shared push approach for all the users in the cell
needs to be done before a decision is taken.
Algorithm. This step can be evaluated by having a loop over
all the |Fu | grid cells that include u. For each such cell C
at friend fi , we will calculate two costs: (1) Costshared , as
the cost of using the shared push approach for all the users in
this cell. This is computed as described in Section V-D, and
(2) CostC , as the current cost encountered for the users in C
that currently use the spatial push approach. To compute this
cost, we will need to go through all the users with the spatial
push approaches and compute their current cost. At the end if
Costshared < CostC , we set the decision to apply the shared
push approach for all users in C with spatial push approach,
otherwise, we do not change the decisions taken so far.
VII. G EO F EED
FOR
M OBILE U SERS
All our previous discussions consider static registered locations for users logging on to GeoFeed. So, once a user logs
off GeoFeed, we compute the cost and decision models for
that user based on its registered location. User locations can
be registered with GeoFeed either explicitly from the user or
implicitly by detecting that a user is frequently logging on
from a certain location. If a user has multiple registered static
locations, e.g., home and work, GeoFeed treats each location
separately, where the cost and decision models can be different
for each registered location. As mentioned earlier, once a user
is logged on, incoming news feed will just be pushed to the
user based on the first location he/she logged on from.
In the case of mobile users, i.e., users keep moving during
a login session, just pushing the incoming news feed based
on the initial logging location is not applicable as users keep
changing their locations. On the other hand, pulling the news
feed for each new location has a prohibitive cost due to the
high frequency of location changes. To this end, we modified
the spatial push approach to support mobile users. The basic
idea is to extend a user u’s point location to be the whole
grid cell C in which u is located in. Now, C is considered as
the region of interest of user u where the news feed will be
retrieved as those messages from u’s friends and overlap with
C. Then, a filter will be added to show only those messages
that overlap with u’ location. As long as u moves within its
cell C, incoming messages that overlap with C are pushed
to u as in the spatial push approach. Once u moves out of
C to another grid cell C , GeoFeed employs the spatial pull
approach to retrieves news feed overlap with C .
VIII. E XPERIMENTS
This section gives experimental evaluation of GeoFeed
based on an actual system implementation in PostgreSQL
database management system [18]. We compare GeoFeed
against an adaptation of the feeding frenzy approach [23], an
industrial solution from Yahoo!, through an additional spatial
filter as well as variations of the query evaluation approaches
within GeoFeed. All experiments are based on a mixture of
real and synthetic data. The real part comes from Twitter messages collected by our web crawler via Twitter Search API [28]
by continuously issuing a query to the API with a spatial
range of 150 × 150 miles space (approximately the size of the
state of Minnesota). We query the API continuously for one
week, and got 646,697 geo-tagged distinct tweets. The Twitter
search API returns the location information with each tweet as
either a longitude/lattitude coordinate or a semantic location,
e.g., a city name, in which we use TinyGeoCoder [25] and
Google GeoCoding [11] APIs to map it to longitude/latitude
coordinates. Then, we randomly generate message spatial
extents as circular areas centered at each message issuing
location with a radius generated uniformly from 5 to 20
miles. We use a synthetic data set of 10,000 users randomly
distributed over the map of Minnesota, where we randomly
associate the real tweets to the synthetic users. Mostly taken
from Facebook statistics [7], and unless mentioned otherwise,
each user has an average of 150 friends, eight hours offline
time before logging on, 500ms response time requirement,
range of interest includes the user location only, and generates
90 messages per month. All average values are generated using
Zipf distribution [33] (with skewness 0.5) that gives higher
probabilities for smaller values. Experiments were evaluated
on a server computer with Intel Core 2 Quad CPU 2.83GHz
processor and 4 GB RAM with Ubuntu Linux 9.04.
To build the cost model, we ran 1,000 queries
using each approach. We find that CostP ull =7.8ms,
CostRV iew =0.5ms, CostRSV iew =0.8ms, CostIV iew =21.2ms
and CostISV iew =22.5ms. This confirms our earlier assumption that CostRV iew CostP ull , and CostRV iew is relatively
close to CostRSV iew . We use these constant values in our
GeoFeed decisions. The rest of this section is as follows:
Section VIII-A gives an inside look on how GeoFeed takes
its decisions. Comparison with other approaches is in Section VIII-B. Section VIII-C tests GeoFeed performance with
respect to various parameters by simulating the workload for
one day with 300 users.
A. Inside GeoFeed
Figure 3 gives an inside view of how GeoFeed smartly takes
its decisions in selecting the right query evaluation approach
among spatial pull, spatial push, and shared push approaches
Spatial pull
Spatial push
Shared push
∞
00
0
10
0
80
0
50
0
30
00
10
0
∞
0
10
0
80
0
50
0
30
10
0
Response Time Requirements (ms)
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
Response Time Requirements (ms)
7000
6000
5000
GeoFeed
All spatial push
All spatial pull
Spatial feed frenzy
3000
Maintenance Cost (s)
Spatial pull
Spatial push
Shared push
Average Response Time (ms)
Percentages of Each Appraoch
Percentages of Each Appraoch
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
4000
3000
2000
1000
0
100
150
200
250
Average Number of Friends
300
2500
GeoFeed
All spatial push
All spatial pull
Spatial feed frenzy
2000
1500
1000
500
0
100
150
200
250
Average Number of Friends
300
(a) Offline Time = 1 hour.
(b) Offline Time = 8 hours.
Fig. 3. Inside GeoFeed Decision Model.
(a) Average Response Time.
(b) Total System Overhead.
Fig. 4. Different Friends Numbers.
for each user query issued from a user u to a friend fi ∈ Fu .
We vary the user response time requirements Tu from zero to
infinity for a user offline time of one hour (Figure 3a) and
eight hours (Figure 3b). For each value of Tu , we plot the
ratio of queries evaluated with spatial pull, spatial push, and
shared push approaches.
This experiment gives the following four very interesting
insights: (1) With the increase of Tu , more spatial pull
approaches are selected. This is mainly because GeoFeed feels
that it has a relaxed Tu , so, it goes for minimizing the system
overhead by having more spatial pull approaches that does
not cost much of system overhead. In the mean time, we
have less shared push approaches with the increase of Tu ,
which is natural as we have less spatial push approaches,
which decreases the probability that GeoFeed finds queries
that can be shared together. (2) When Tu =0ms, (i.e., the user
needs the news feed as fast as possible), no spatial pull
approaches are applied, as they definitely pose more query
response time for system users. However, it is interesting to
notice that not all the queries are evaluated by the spatial
push approach where a significant portion of the queries use
the shared push approach. This is mainly due to our valid
assumption that the query costs for both the spatial push
and shared push approaches are similar. So, GeoFeed aims
to reduce the system cost through having more of the shared
push approach. (3) When Tu =∞ (i.e., u does not have any
response time requirement), GeoFeed aims to only minimize
the system overhead through employing much of the spatial
pull approach. However, it is interesting to notice that some
queries are still evaluated with other approaches, especially
with smaller offline time (Figure 3a). This is mainly because
the GeoFeed decision model takes into account the case that
the spatial push approach may be cheaper than the spatial
pull approach, which takes place with low update frequency
and/or short user offline time. (4) Comparing Figures 3a and 3b
together shows that with a smaller offline time, more spatial
push approaches are applied. As a smaller offline time means
that the user will log soon again to GeoFeed, so, it is better to
maintain the materialized view in the spatial push approach
rather than executing the query from scratch upon the next log
on time as in the spatial pull approach.
B. Comparison with Other Approaches
solution optimization in location-based systems, that pushes
the spatial pruning to the bottom and (c) having all the queries
evaluated with the spatial push approach, which is the stateof-the-art optimization in the publish-subscribe systems, that
prunes out the unnecessary update notices early. We vary the
average number of friends for each user from 100 to 300, and
measure the average response time (Figure 4a) and system
overhead cost (Figure 4b). It is clear that feeding frenzy has
a very bad user response time as the spatial filter is applied
afterthought, and it does not take the spatial aspects of the
messages in its decisions. Though the all spatial pull approach
gives better response time than feeding frenzy, but it is still
unacceptable as it evaluates the query from scratch for each
friend. The performance of GeoFeed and all spatial push are
very similar in terms of response time, and almost have a
steady performance with the increase of the number of friends,
which is orders of magnitude better than that of both the
feeding frenzy and all spatial pull approaches. With respect
to system overhead, the all spatial push approach has a much
higher cost than that of GeoFeed (about double cost for 300
friends). This is mainly due to the large number of materialized
views maintained for the all spatial push approach.
From this experiment, we conclude that both feeding frenzy
and all spatial pull approaches are completely impractical due
to their unacceptable user response time, and hence we will
not consider them later. Similarly, we will not consider the all
spatial push approach later due to its clearly higher overhead
cost than GeoFeed.
C. GeoFeed Performance
Figure 4 compares GeoFeed performance against: (a) an
adaptation of feeding frenzy [23] to handle spatial data through
an additional spatial filter, (b) having all the queries evaluated
with the spatial pull approach, which is the state-of-the-art
In this section, we compare the performance of two variations of GeoFeed; one with only the spatial pull and spatial
push approaches (termed GF-PP or GeoFeed-PP), and the
other one is our complete GeoFeed system (termed GF-PPS
or GeoFeed-PPS) with different parameters: (1) user response
time requirement Tu , (2) offline time, (3) news update rate,
(4) user spatial distribution, (5) grid cell granularity, (6) user
querying range, and (7) GeoFeed with mobility. For a fair
comparison, we use the same fixed scale in x-axis for the
most experiments when comparing the average response time
and total update cost between GF-PP or GeoFeed-PP.
Response Time Requirements. Figure 5 gives the impact
of user response time requirements on the performance of
GeoFeed-PPS and GeoFeed-PP, where the required response
time varies from 100 to 1,000 ms. When the response time
requirement becomes less strict, the system can use more
spatial pull approaches to process more location-based queries.
500
400
300
200
1200
1100
1000
900
800
1,400
1,200
1,000
800
600
400
700
1,000ms
0
PS
P
P
F−
P
F−
S
PP
F−
P
P
F−
300ms
500ms
800ms
Response Time Requirements
G
G
G
G
S
S
PP
F−
PP
F−
PP
F−
P
P
F−
G
G
G
G
PS
0
100
300
500
800 1000
Response Time Requirements (ms)
1,600
1100
1,400
1000
900
800
1,000
800
600
200
600
S
PP
F−
G P
P
F−
G
S
PP
F−
G P
P
F−
G
S
PP
F−
G P
P
F−
G
S
PP
F−
G
PP
F−
PS
P
4
G
P
F−
20
P
F−
8
12
16
Averge Offline Time (hours)
G
0
4
G
20
1,200
400
700
500
Maintenance Cost
Query Processing Cost
8
12
16
Average Offline Time
(b) Maintenance Cost.
(c) Total Cost.
Fig. 6. User offline time periods.
Since the response time of the spatial pull approach is higher
than that of the spatial push and shared push approaches,
the response time of both GeoFeed-PPS and GeoFeed-PP
gets worse with longer required response times (Figure 5a).
When there are more location-based queries selected to use
the spatial pull approach, i.e., a smaller number of queries
is processed by the spatial push approach, the chance of
grouping queries to use the shared push approach becomes
slimmer. Thus, the improvement of GeoFeed-PPS on the view
maintenance cost of materialized views reduces (Figure 5b).
From Figure 5c, we can observe that with tight requirements,
there is significantly large maintenance cost than query processing cost, as GeoFeed tends to avoid using spatial pull in
order to be able to satisfy the user requirements, and thus has
to pay maintenance cost for using spatial push and shared
push. With relaxed requirements (increase of Tu ), GeoFeed
tends to optimize its maintenance cost in favor of having
more query processing cost, such that the overall total cost
is decreased. Since maintenance cost optimization reduces the
use of spatial push, it also reduces the possibility of having
shared push refinement, and hence the performance of both
GeoFeed variations becomes similar. This shows that the use
of shared push refinement is more clear when the system is
overloaded either with tight requirements, short offline times,
or high update frequencies.
Figure 5d gives the ratio of user queries that exceed the
user response requirements for GeoFeed-PPS with average
offline time of one hour and eight hours. We vary Tu from
zero to ∞. What is interesting here is the failure ratio is
exponentially decreasing with relaxed time constraints. Also,
we have less failure with smaller offline time, as more queries
use the spatial push approach which has more deterministic
user response time.
User Offline Time Periods. Figure 6 compares the performance of GeoFeed-PPS and GeoFeed-PP with respect to
various user offline time periods from 4 to 20 hours. Increasing
the user offline time period also increases the cost of the
20
Percentages of Each Appraoch
1,800
∞
(d) Fail Ratio.
2,000
1200
Total Cost (s)
Maintenance Cost (s)
300
(a) Average Response Time.
P
350
8
12
16
Averge Offline Time (hours)
100ms
GeoFeed-PPS
GeoFeed-PP
1300
400
4
P
F−
500
100
300
500
800
1000
Average Response Time Requirement (ms)
1400
450
250
40
(b) Maintenance Cost.
(c) Total Cost.
Fig. 5. Response time requirements.
GeoFeed-PPS
GeoFeed-PP
500
P
F−
550
60
0
G
100
100
300
500
800
1000
Average Response Time Requirement (ms)
GeoFeed-PPS-1
GeoFeed-PPS-8
80
20
200
600
(a) Average Response Time.
Average Response Time (ms)
1,600
Failure Ratio(%)
600
Maintenance Cost
Query Processing Cost
1,800
Total Cost (s)
Maintenance Cost (s)
700
100
2,000
GeoFeed-PPS
GeoFeed-PP
1300
G
Average Response Time (ms)
1400
GeoFeed-PPS
GeoFeed-PP
800
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
Spatial pull
Spatial push
Shared push
4
8
12
16
20
Average Offline Time (hours)
(d) Inside GeoFeed Decisions.
spatial push approach because the cost of maintaining materialized views during a user offline time period is higher. As
a result, the system selects more queries to use the spatial pull
approach. Since the response time of the spatial pull approach
is higher than that of the spatial push approach, the response
time of both GeoFeed-PPS and GeoFeed-PP gets higher when
the user offline time period increases, as depicted in Figure 6a.
Since the number of location-based queries using the spatial
push approach decreases as the user offline time period gets
longer, smaller numbers of queries can be grouped to use
the shared push approach; and hence, the improvement of
GeoFeed-PPS on the total view maintenance cost is smaller.
From Figure 6c, we can observe that with the increase of
offline time, both the query processing and maintenance cost
decrease as less queries are posed to the system, and most
of them are evaluated with spatial pull. Overall, with the
increase of offline time, we have a similar performance of
GeoFeed-PPS and GeoFeed-PP, because there are less spatial
push queries, which is illustrated in Figure 6d, and hence the
impact of the shared push refinement is reduced.
News Update Rates. Figure 7 compares the performance between GeoFeed-PPS and GeoFeed-PP by varying the average
number of news update rates per month from 30 to 150. Since
a higher news update rate results in more messages generated
in the system, both GeoFeed-PPS and GeoFeed-PP need to
use the grid index to prune more irrelevant messages; and
thus, their response time increases (Figure 7a). When the
news update rate increases, both algorithms have to update
materialized views more frequently. Since the materialized
view update cost of the shared push approach is shared by
a group of users, it is more efficient than updating each
materialized view of the spatial push approach individually.
Therefore, GeoFeed-PPS performs much better than GeoFeedPP in terms of the view maintenance cost of maintaining
materialized views when the news update rate gets higher
(Figure 7b). In Figure 7c, the average update frequency is
varied from 30 to 150 per month. With the increase of the
400
350
300
1500
1000
2,000
1,500
1,000
500
500
0
PS
P
PS
P
PS
rm
1%
%
25
0.
Percentages of Each Appraoch
Total Cost (s)
ifo
un
00
%
%
25
11
0.
%
06
0.
PS
P
PS
P
F−
P
P
F−
1%
Spatial pull
Spatial push
Shared push
04
G
G
PS
P
F−
P
P
F−
PS
P
F−
P
P
F−
PS
P
F−
P
P
F−
P
F−
(b) Maintenance Cost.
(c) Total Cost.
Fig. 9. Grid cell granularities.
update frequency, the maintenance cost increases significantly
for both approaches due to the need of more updates to the
materialized views. However, GeoFeed-PPS has less total cost,
as the shared push approach reduces the number of insertions
to the materialized views. It shows the efficiency gained from
sharing multiple views together to significantly reduce the
maintenance cost. With respect to query processing cost, there
is a slight increase for both approaches with the growth of
the update frequency, which is illustrated by the decisions
inside GeoFeed in Figure 7d. Overall, GeoFeed-PPS gives
much lower total cost than GeoFeed-PP.
User Spatial Distributions. Figure 8 gives the experimental
results of GeoFeed-PPS and GeoFeed-PP with respect to two
different user distributions: (1) Uniform distribution where
users are randomly distributed in the map and (2) Gaussian
distribution where users are distributed in the map according
to different standard deviations SD that indicate different user
densities in a certain range distance of a given location. The
Gaussian distribution is mainly used to simulate a hot spot
area in the map, i.e., the downtown area in the city, to further
explore the impact of user spatial locality and the advantage
of applying the shared push approach. A smaller value of SD
indicates that users are concentrated in a denser area. In this
experiment, SD varies from 1 to 1,000.
Since the user density does not affect the selection between
the spatial pull and spatial push approaches, varying the user
density only slightly affects the performance of GeoFeed-PP
(Figures 8a to 8c). However, the response time of GeoFeed-
G
G
G
G
G
G
P
F−
0.0,625%
0.11%
0.25%
Grid Cell Granularities
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
0.
0
0.04%
0
=1
600
200
1%
0
800
G
0.0625% 0.11%
0.25%
Different Grid Cell Granularities
SD
1,000
G
500
0.04%
0
=1
1,200
400
GeoFeed-PPS
GeoFeed-PP
SD
1,400
700
0
=1
Maintenance Cost
Query Processing Cost
800
SD
PS
P
PS
P
PS
Different User Spatial Distributions
(d) Inside GeoFeed Decisions.
1,600
900
Spatial pull
Spatial push
Shared push
=1
P
F−
P
F−
P
F−
P
F−
P
uniform
1,800
1000
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
G
G
G
G
P
F−
PS
SD=10
SD=100
SD=1,000
User Spatial Distributions
(c) Total Cost.
User distributions.
1100
30
60
90
120
150
Average Number of Messages Per Month
SD
Percentages of Each Appraoch
Total Cost (s)
Maintenance Cost (s)
P
F−
P
F−
P
F−
P
F−
G
P
F−
P
Maintenance Cost(s)
G
G
G
G
P
G
P
F−
P
F−
PS
P
SD=1
G
G
P
F−
1%
P
F−
0
SD=10 SD=100 SD=1000 uniform
Different User Spatial Distributions
1200
600
(a) Average Response Time.
200
GeoFeed-PPS
GeoFeed-PP
P
F−
300
0.0625%
0.11%
0.25%
Different Grid Cell Granularities
600
G
350
250
0.04%
800
400
700
1300
400
G
800
1,000
1400
450
PS
900
(b) Maintenance Cost.
Fig. 8.
GeoFeed-PPS
GeoFeed-PP
P
F−
1000
G
500
1,200
1100
(a) Average Response Time.
550
1,400
1200
Spatial pull
Spatial push
Shared push
(d) Inside GeoFeed Decisions.
Maintenance Cost
Query Processing Cost
500
SD=1
SD=10 SD=100 SD=1000 uniform
Different User Spatial Distributions
150
1,600
600
250
SD=1
G
300
P
350
P
F−
400
60
90
120
Average Number of Messages Per Month
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
(c) Total Cost.
News update rates.
1300
450
G
500
30
1400
GeoFeed-PPS
GeoFeed-PP
P
F−
(b) Maintenance Cost.
Fig. 7.
G
550
60
90
120
150
Averge Number of Updates Per Month
PS
30
P
0
60
90
120
150
Averge Number of Updates Per Month
P
F−
30
P
F−
250
G
Average Response Time (ms)
Maintenance Cost
Query Processing Cost
2,500
2000
(a) Average Response Time.
Average Response Time (ms)
3,000
Total Cost (s)
Maintenance Cost (s)
450
GeoFeed-PPS
GeoFeed-PP
2500
Percentages of Each Appraoch
3000
GeoFeed-PPS
GeoFeed-PP
500
G
Average Response Time (ms)
550
Different Grid Cell Granularities
(d) Inside GeoFeed Decisions.
PPS increases as the users are distributed more densely. This
is because the system with a higher user density has a higher
probability to use the shared push approach to process user
queries, as given in Figure 8d, which increases the query
response time. On the positive side, using the shared push
approach to process more user queries leads to a lower update
cost of maintaining shared materialized views (Figure 8b).
As a result, the view maintenance cost of GeoFeed-PPS
performs much better than that of GeoFeed-PP when the user
distribution becomes denser.
Grid Cell Granularities. Figure 9 depicts the performance
of GeoFeed-PPS and GeoFeed-PP with respect to various cell
granularities of the grid index which is defined as a ratio of the
cell area to the total system area. This experiment considers
five different grid cell granularities 0.04%, 0.0625%, 0.11%,
0.25% and 1%. The results give that the grid cell size only
slightly affects the performance of GeoFeed-PP (Figures 9a
to 9c). On the other hand, when the grid cell size gets larger,
there is a higher chance for GeoFeed-PPS to use the shared
push approach to share a materialized view with a group of
user queries, as confirmed by Figure 9d. Due to the fact that the
response time of the shared push approach is slightly higher
than the spatial push approach, the response time of GeoFeedPPS increases as the grid cell size gets larger.
User Querying Ranges. Figure 10a depicts the performance
of GeoFeed-PPS and GeoFeed-PP with different user querying
ranges with respect to varying from 1 mile to 5 miles.
Figure 10b demonstrates the decisions inside GeoFeed system.
1,000
800
600
400
200
S
PP
F−
G P
P
F−
G
S
PP
F−
G P
P
F−
G
S
PP
F−
G P
P
F−
G
S
PP
F−
G P
P
F−
G
S
PP
F−
PP
F−
G
G
0
1
2
3
4
Average Querying Ranges
(a) Total Cost.
Fig. 10.
5
Spatial pull
Spatial push
Shared push
1
2
3
4
5
Average Querying Ranges (miles)
(b) Inside GeoFeed Decisions
User Querying Ranges.
With the increase of the user querying ranges, more queries
are executed with spatial pull approach, because the larger
querying range has more chance to get the message updates,
which increases spatial push cost to maintain the materialized
view. As a result, the total querying process cost in Figure 10a increases with the growth of user querying ranges.
Moreover, more shared push approaches are adapted in the
system, because the higher cost of spatial push makes shared
push approach more efficient. As a result, our GeoFeed-PPS
becomes more efficient with a lower overall system overhead
than GeoFeed-PP with the increase of user querying ranges.
GeoFeed Mobility. Figure 11a depicts the performance of
GeoFeed and GeoFeed-Mobility for mobile users with respect
to varying their query frequencies from 5 to 60 queries per
hour with the user’s traveling speed at 5 miles per hour,
and the grid cell granularity as 0.25%. The results show that
GeoFeed-Mobility is much more scalable than the original
GeoFeed. In fact, the system overhead of GeoFeed-Mobility is
not dependent of the user query frequency because GeoFeedMobility only needs to push new messages to a user when the
user moves across grid cells.
Figure 11b gives the results of GeoFeed-Mobility for mobile
users moving at two mobility speeds (i.e., 5 and 20 miles per
hour) in different grid cell granularities, increasing from 0.04%
to 1%, with the default query frequency at 10 queries per hour.
The results show that the response time of GeoFeed-Mobility
improves as the grid cell size gets larger because the users
take longer time to move outside a grid cell; thus, the users
can compute their query answers locally without enlisting the
server for help. Similarly, GeoFeed-Mobility with the lower
mobility speed provides the better response time, because the
users can stay in a grid cell for longer time.
IX. C ONCLUSION
We have presented the GeoFeed system; a location-aware
news feed system that takes into account the spatial extents
of messages and user locations when deciding upon the
selected news feed. GeoFeed is equipped with three different
approaches, namely spatial pull, spatial push and shared push
for delivering the news feed to its users. Based on an accurate
developed cost model for each approach, GeoFeed employs
a smart decision algorithm that decides about using these
approaches in a way that: (a) minimizes the system overhead
for delivering the location-aware news feed, and (b) guarantees
a certain response time for each user to obtain her locationaware news feed. GeoFeed further extends the spatial push
approach to support the moving users. Experimental results,
5500
5000
4500
4000
3500
3000
2500
2000
1500
1000
500
0
GeoFeed
GeoFeed-mobility
5
10
20
30
60
User Query Frequencies (query/hour)
Average Response Time (ms)
Total Cost (s)
1,200
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
Total System Overhead (ms)
1,400
Percentages of Each Appraoch
Maintenance Cost
Query Processing Cost
1,600
1100
1000
900
800
700
600
500
400
300
200
100
0
0.04%
GeoFeed-mobility-5
GeoFeed-mobility-20
0.0625% 0.11%
0.25%
Grid Cell Granularities
1%
(a) Query Frequencies.
(b) Grid Cell Granularities.
Fig. 11. GeoFeed with Moving Users.
based on real and synthetic data, show that GeoFeed is
favorable over existing news feed systems, with a minimal
system overhead.
R EFERENCES
[1] L. Aalto, N. Göthlin, J. Korhonen, and T. Ojala. Bluetooth and wap push
based location-aware mobile advertising system. In MobiSys, 2004.
[2] Y. Cai and T. Xu. Design, analysis, and implementation of a large-scale
real-time location-based information sharing system. In MobiSys, 2008.
[3] B. Chandramouli and J. Yang. End-to-end support for joins in large-scale
publish/subscribe systems. PVLDB, 1(1), 2008.
[4] B. Chandramouli, J. Yang, P. K. Agarwal, A. Yu, and Y. Zheng. ProSem:
scalable wide-area publish/subscribe. In SIGMOD, 2008.
[5] R. Cheng, Y. Xia, S. Prabhakar, and R. Shah. Change Tolerant Indexing
for Constantly Evolving Data. In ICDE, 2005.
[6] Facebook. http://www.facebook.com/.
[7] Facebook Statistics. http://www.facebook.com/press/info.php?statistics,
2010.
[8] FourSquare. http://www.foursquare.com/.
[9] D. Freni, C. R. Vicente, S. Mascetti, C. Bettini, and C. S. Jensen.
Preserving location and absence privacy in geo-social networks. In
CIKM, 2010.
[10] Google Buzz Moblie. http://www.google.com/buzz.
[11] Google GeoCoding. http://code.google.com/apis/maps/documentation/geocoding/.
[12] iGoogle. http://www.google.com/ig.
[13] A. Khoshgozaran and C. Shahabi. Private Buddy Search: Enabling
Private Spatial Queries in Social Networks. In SIN, 2009.
[14] Loopt. http://www.loopt.com.
[15] M. F. Mokbel, X. Xiong, and W. G. Aref. SINA: Scalable Incremental
Processing of Continuous Queries in Spatio-temporal Databases. In
SIGMOD, 2004.
[16] MyYahoo! http://my.yahoo.com/.
[17] M. Park, J. Hong, and S. Cho. Location-based recommendation system
using bayesian users preference model in mobile devices. UIC, 2007.
[18] PostgreSQL . www.postgresql.org.
[19] S. Prabhakar, Y. Xia, D. V. Kalashnikov, W. G. Aref, and S. E. Hambrusch. Query Indexing and Velocity Constrained Indexing: Scalable
Techniques for Continuous Queries on Moving Objects. IEEE Trans.
Computers, 51(10), 2002.
[20] Renren. http://www.renren.com/.
[21] L. Siksnys, J. Thomsen, S. Saltenis, and M. L. Yiu. Private and Flexible
Proximity Detection In Mobile Social Networks. In MDM, 2010.
[22] L. Siksnys, J. R. Thomsen, S. Saltenis, M. L. Yiu, and O. Andersen. A
Location Privacy Aware Friend Locator. In SSTD, 2009.
[23] A. Silberstein, J. Terrace, B. F. Cooper, and R. Ramakrishnan. Feeding
Frenzy: Selectively Materializing User’s Event Feed. In SIGMOD, 2010.
[24] Sina Weibo. http://www.weibo.com/.
[25] TinyGeoCoder. http://tinygeocoder.com/.
[26] Twinkle. http://tapulous.com/twinkle/.
[27] Twitter. http://www.twitter.com/.
[28] Twitter Search API. http://search.twitter.com.
[29] M. Ye, P. Yin, and W. Lee. Location recommendation for location-based
social networks. In ACM SIGSPATAIL GIS, 2010.
[30] V. Zheng, Y. Zheng, X. Xie, and Q. Yang. Collaborative Location and
Activity Recommendations with GPS History Data. In WWW, 2010.
[31] Y. Zheng, Y. Chen, X. Xie, and W.-Y. Ma. GeoLife2.0: A LocationBased Social Networking Service. In MDM, 2009.
[32] Y. Zhou, A. Salehi, and K. Aberer. Scalable delivery of stream query
results. PVLDB, 2(1), 2009.
[33] G. K. Zipf. Human Behaviour and the Principle of Least Effort: an
Introduction to Human Ecology. Addison-Wesley, 1949.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement