C#代写|Gui代写|数据分析: 这是一个利用C#实现paper中的数据分析界面的代写
Abstract
Social Networks now become popular and powerful
platforms for people to share information. Everyone
may share their interested information with
their connections, or send messages to their friends.
However, sharing information costs both computational
and communicational resources, in addition
to personal time/attention of both the sender
and the receiver. Decision-making regarding which
piece of information should be shared with whom,
thus is important to individuals and the whole network.
In this work, we study the effects of different
information-sharing strategies using a social network
simulator framework. This paper describes
how social network is modeled, and the various factors
relevant to the sharing decisions. We propose
six information sharing strategies, and performed
simulation experiments to examine their influences
on individuals and the whole social network.
keywords: Social Networks, Information Sharing, Strategies,
Simulation.
1 Introduction
Social network here refers to a group of people connected
through Internet via emails, websites, and social media such
as Facebook, Twitter, blogs, and WeChat. Social Networks
have become increasingly popular and fast-growing platforms
for information sharing, job searching and product
marketing [Hogan, 2008; Ellison, 2007]. Information propagates
fast in social network through direct personal social
connections. There are various ways for sharing information,
including direct email message, forwarding a post to some or
all connections, recommending a post/blog/product to some
or all connections, posting/blogging on personal website allowing
a certain group of people to view, etc. Each method
works differently in term of the targeted audience and the implicitly
required degree of attention. A receiver of a direct
email feels more obliged to read the message and reply than
a receiver of an automatically-generated notification of a new
post by a friend.
∗
Given the six degree of separation, small world theory
[Travers and Milgram, 1969], and the fast communication
speed of internet, the information propagated in social network
may reach a very large population within a very short
period of time [Kim et al., 2010]. However, this powerful
and speedy circulation does not come free. The computational
and communicational costs are obvious, the sharing of
information increases the usage of computational resources
and generates large amount of internet traffics. Other lessobvious
costs include the personal time and attention of both
the sender and the receivers. Additionally, the increasing
amount of information reduces the receiver’s attention to any
particular piece of information, hence diminishes the influence
of the shared information or even damages the social
trust between the receiver and the sender. Therefore, it is important
to share the right information with the right person.
However, this is not an easy task if an individual has to
decide that with whom to share for each piece of received
information. The goal of this study is to develop automatic
decision-making mechanisms to help each individual in a social
network. The automatic decision-making is based on
the information relevance and the interest profile of each
connected individual. Each piece of information can be either
manually or more realistically, automatically classified
for its relevance to different categories or subjects, with the
help of context analysis and text mining tools [Aggarwal and
Zhai, 2013], or structured meta-level data such as semantic
web [Gruber, 2008]. On the other hand, each individual
may truthfully describe one’s interested subjects in one’s
published interest profile, given the motivation to reduce undesired
information. The focus of this work is to study different
information-sharing strategies, assuming both information
relevance and interest profile are available.
The uniqueness of this work lies in the aspect of viewing
each node as an intelligent individual and be able to make informed
decision. In traditional research on information propagation
/diffusion, some mathematic models such as Linear
Threshold Model and Independent Cascade Models [Kempe
et al., 2003; Guille et al., 2013] are used to describe the diffusion
process. With such models, each individual is simply
a data object [Smith et al., 2009] behaved according to
a fixed protocol and/or an pre-set attribute value, without
freedom to make its own decision on which information to
share with whom. In recent work on strategic networks with
self-interested agents [Zhang and van der Schaar, 2013], each
node is a strategic agent who benefits from producing and disseminating
information. This is similar to the model used in
our work: each individual may choose different informationsharing
strategy according to its own goal and preference.
This work studies the impact of different information-sharing
strategies not only on the whole network, but also on each
individual in the network. Such decision affects the amount
of information shared, the attention level of the receivers, the
propagation in the network and also the network connection
structure: too many un-appreciated information may cause
disconnection. Individual behavior also has impact on computational
and communication cost, which is a concern for
the society and also should be a concern for everyone, as the
environmental issue.
[Hudack et al., 2014] has studied information diffusion
with game theory. [Horel and Singer, 2015] describes an
adaptive information dissemination method that selects users
aiming to target their influential neighbors. However, our
study has different focus and model than these work.
To simplify the description, in the rest of the paper, we use
message to refer to all types of information one may share
in social network, including post, blog, video, pins, etc. To
send message or forward message means to share a piece of
information with the receiver. In current work, we assume
there is only one method of sharing information; in the future
we will study different sharing methods.
In Section 2 we describe the graph model of social network.
We then present the representation of information relevance
and interest profile of individuals in Section 3. Six
different information-sharing strategies are proposed in Section
4, and the network structure parameters and the experiments
are reported in Section 5. In Section 6 we define
four evaluation criteria for these strategies, and experimental
results are reported in Section 7. Finally we discuss our
conclusion and future work in Section 8.
2 Social Network Structure and Modeling
In this work, we model social network as a graph, with Node
representing individual person, and Edge (link) representing
connection between two persons. In this work, we model social
network as undirected graph, and the connections represent
symmetric relationships, such as friend relationship.
However, some connections in social network are asymmetric,
i.e. following relationship; these asymmetric relations
can be modeled as directed links in future study.
Degree of a node measures the number of connections of
this node. In a social network, nodes may have significantly
different degree: some persons have a lot of connections
while some may only have a few. A cluster, or community
is a group of nodes with many connecting edges between
them, and there are relatively few connecting edges between
nodes that belong to different clusters [Fortunato, 2010;
Golbeck, 2013]. Inside each cluster, a leader node is the one
who has the most number of connections within this cluster.
A leader node may serve as an information hub that connects
to other clusters.
Number of clusters and the size of each cluster have big
Figure 1: A network with 27 small clusters of size 10, 6
medium clusters of size 33, and 2 large clusters of size 57.
influence on the propagation of information in the social network
[Golbeck, 2013]. Hence we choose the following parameters
to characterize the network structure in this work:
1. Number of small clusters, and the size (number of
nodes) of a small cluster.
2. Number of medium clusters, and the size of a medium
cluster.
3. Number of large clusters, and the size of a large cluster.
We build a simulator that takes the above six parameter values
as input and create a network accordingly. Figure 1 shows
a network with 27 small clusters of size 10, 6 medium clusters
of size 33, and 2 large clusters of size 57. With the repelling
property of the nodes in D3 force directed graphs [Kobourov,
2012], nodes in the same cluster attract each other while those
in different clusters repel each other.
3 Information Relevance and Individual
Interest Model
To describe how much an individual in social network is interested
in a piece of information, we introduce the following
model.
Assume there are x categories (subjects) being modeled in
this framework: c1, c2, …, cx, each message m is associated
with a Category List CLm:
CLm = {(cm1, rm1), · · · ,(cmi, rmi), · · · ,(cmx, rmx)}
cmi is the ith category that message m is relevant to, and
Relevance Factor rmi ∈ [0, 1] describes how strong message
m is relevant to category cmi, where 1 stands for the strongest
relevance and 0 means no relevance at all.
For example, a message ma about how to choose running
shoes has a category list as CLma
shown in Table 1, and another
message mb on some health diet and exercise suggestions
for losing weight has a category list as CLmb
shown in
Table 1: Examples of Information Relevance and Individual Interest Model
Relevance of ma CLma {(F itness, 0.5),(Shoes, 0.8)}
Relevance of mb CLmb
{(F itness, 0.4),(Diet, 0.8),(W eightControl, 1.0),(Health, 0.7)}
Interest Profile of np F Lnp {(F itness, 1, 0.6),(Diet, 3, 0.9),(Shoes, 2, 0.9),(Health, 4, 0.8)}
Interest Set of CLma
and F Lnp
ISma,np {F itness, Shoes}
Interest Set of CLmb
and F Lnp
ISmb,np
{F itness, Diet, Health}
Size of Interest Set | ISma,np
| 2
Size of Interest Set | ISmb,np
| 3
Set of Interest Factors I(ma, np) {1, 2}
Set of Interest Factors I(mb, np) {2, 3, 4}
Set of Relevance Factors R(ma, np) {0.5, 0.8}
Set of Relevance Factors R(mb, np) {0.4, 0.8, 0.7}
Average Interest Factor Value Ia(ma, np) 1.5
Average Interest Factor Value Ia(mb, np) 3
Average Relevance Factor Value Ra(ma, np) 0.65
Average Relevance Factor Value Ra(mb, np) 0.63
Average Relevance Threshold Value RTa(ma, np) 0.75
Average Relevance Threshold Value RTa(mb, np) 0.77
Table 1 as well. Such category list can be automatically generated
with natural language processing tools [Aggarwal and
Zhai, 2013].
On the other hand, each individual n has a profile describing
one’s interest, represented as an Interest Factor List:
F Ln = {(cn1, fn1, fnr1), · · · ,(cnj , fnj , fnrj ), · · · ,
(cnx, fnx, fnrx)}
cnj is the jth category that node n is interested in, interest
factor fnj represents how interested node n is in category cnj .
fnj is an integer in the range of [0, 5]. Relevance threshold
fnrj is the minimum value of the relevance factor in category
cnj for a message to be considered as relevant by node
n. For example, an individual np with interest profile F Lnp
shown in Table 1, is interested in four categories: Fitness,
Diet, Shoes and Health. A message must has a relevance factor
value no less than 0.6 for Fitness, or 0.8 for Diet, 0.9 for
Shoes, or 0.8 for Health to be considered relevant by this individual,
to each category respectively.
Each individual in social network may set up his/her interest
profile to describe which categories one is interested and
how much interest one has. One may also adjust the relevance
threshold value dynamically based on the amount of
information one receives and one’s tolerance at that time.
Given a message m with category list CLm, and an individual
node n, Interest Set ISmn is the set of categories that
both message m is relevant to and also node n is interested
in. More formally stated:
ISmn = {c|∃i, cmi ≡ c ∧ rmi > 0 ∧ ∃j, cnj ≡ c ∧ fnj > 0}
Given a message m and an individual node n, the following
parameters are defined based on this intersection set ISmn:
1. size(ISmn) = | ISmn |, number of categories inside
ISmn.
2. I(m, n) = {fnj |cnj ∈ ISmn} the set of the interest
factors, each for one category cnj in ISmn.
3. R(m, n) = {rmi|cmi ∈ ISmn} the set of the relevance
factors, each for one category cmi in ISmn.
0
500
1000
1500
2000
2500
3000
Network1
Network2
Network3
270
1000
2000
198
500
500
114
500
500
#Nodes
Network Structures
Small Cluster Med. Cluster Large Cluster
Figure 2: Three Network Structures
4. Ia(m, n) =
X
cnj∈I(m,n)
fnj
size(ISmn)
, the average interest factor
value.
5. Ra(m, n) =
X
cmi∈I(m,n)
rmi
size(ISmn)
, the average relevance factor
value.
6. RTa(m, n) =
X
cmi∈I(m,n)
fnrj
size(ISmn)
, the average relevance
threshold value.
Table 1 shows the above parameter values given example
message ma , ma and individual np.
4 Information-Sharing Strategies
Based on the information relevance and individual interest
model described in Section 3, we propose six informationsharing
strategies, described below.
Table 2: Network Structure Information and Experimental Parameters
# Small Size # Med. Size # Large Size # Seed # Time
Network Clusters Small Clusters Med. Clusters Large #Nodes Messages Steps
#1 27 10 6 33 2 57 582 89 233
#2 100 10 20 25 10 50 2000 301 800
#3 200 10 20 25 5 100 3000 489 1300
• Strategy 1 Even Little Interested (ELI). Send message
m to node n if there exists at least one common category
in m’s category list and also in n’s interest profile, with
a relevance of factor value no less than 0.1. | ISmn |≥ 1
∧ min(R(m, n)) ≥ 0.1
Using Strategy 1 (ELI), both message ma and mb should
be sent to node np.
• Strategy 2 Average Interest in message (AI). Send
message m to node n if the average interest factor of
all categories in ISmn is no less than 3 and the average
relevance factor value of all categories in ISmn is no
less than 0.3.
Ia(m, n) ≥ 3 ∧ Ra(m, n) ≥ 0.3
Using Strategy 2 (AI), mb should be sent to node np but
message ma should not be sent to np.
• Strategy 3 High Interested and Relevance (HIR).
Send a message m to node n if n is very interested
in any category in ISmn, or the average interest factor
Ia(m, n) is no less than 3 and the average relevance factor
value Ra(m, n) is no less than the average relevance
threshold value RTa(m, n).
max(I(m, n)) = 5 or ( Ia(m, n) ≥ 3 ∧ Ra(m, n) ≥
RTa(m, n) )
Using Strategy 3 (HI), neither ma nor mb should be sent
to np.
• Strategy 4 Unless Not Interested (UNI). Send message
m to node n if there is at least one common category j
in m’s category list and also in n’s interest profile.
| ISmn |≥ 1
Using Strategy 4 (UNI), both message ma and mb
should be sent to node np.
• Strategy 5 Combined Interest and Relevance (CIR).
Send message m to node n if the combined average of
the average interest factor value (in percentage) and the
average relevance factor value (in percentage) is no less
than 50%.
Ia(m,n)
5 +Ra(m,n)
2 ≥ 50%
Using Strategy 5 (CIR), mb should be sent to node np
but message ma should not be sent to np.
• Strategy 6 Moderate Interest and Relevance (MIR).
Send message m to node n if if there is at least one common
category j in ISmn that node n’s interest factor for
j is no less than 3 and message m’s relevance factor for
j is no less than 0.5.
∃j ∈ ISmn, fnj ≥ 3 ∧ rmj ≥ 0.5
Using Strategy 6 (MIR), mb should be sent to node np
but message ma should not be sent to np.
5 Network Structure and Experiment Set up
Network structure may has significant influence on information
propagation in the network. In this study, we conducted
experiments with three networks using parameters shown in
Table 2. Network 1 has 27 small clusters with10 nodes each, 6
medium clusters with 33 nodes each, and 2 large clusters with
57 nodes each. In total, Network 1 has 582 nodes, Network
2 has 2000 nodes and Network 3 has 3000 nodes. Figure 2
illustrates the different structures of these three networks. We
created a pool of 10000 randomly-generated interest profiles,
each with random interest factor values and random relevance
threshold values. When a network is created, the specified
number of nodes are generated, each node is associated with
an interest profile drawn from this pool.
In each experiment with a given network, multiple (about
15% of the total number of nodes) seed messages are created,
each with a randomly-generated category list. The experiment
is running with synchronized simulation time steps. At
each time step, one new seed message is delivered to a randomly
selected node in the network. At each time step, each
node makes decision for each message received in the previous
time step. The decision includes whether to share this
message to its connected nodes, and which nodes to share
with, using its information-sharing strategy. This process
continues until a pre-set number of time steps is reached,
which is about 40% of total number of nodes. Table 2 also
reports these experiment parameter values. For example, 489
seed messages were created for Network 3 and the experiment
runs for 1300 time step.
6 Evaluation Criteria
To evaluate the influence of the information-sharing strategy
on individuals and also on the social network, we define the
following criteria.
• Interest Ratio, measures how many messages are interesting
to a node out of all its received contents.
Interest Ratio(n) for node n, is calculated as:
#Interesting(High,Med.,Low) Messages Received by n
#Received Messages of n
Depending on the degree of interest, three measures are
defined as:
– High Interest Ratio: the ratio of highly interesting
messages, with maximum interest factor value
max(I(m, n)) as 5.
– Medium Interest Ratio: the ratio of medium interesting
messages, with maximum interest factor
value max(I(m, n)) as 3 or 4.
0.21
0.13
1
0.21
0.3 0.35
0.4 0.61
0
0.39
0.46
0.64
0.38
0.25
0.38
0.23
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
Interest Ra?o: Network 1
High_Int/R Med_Int/R Low_Int/R
(a) Network 1
0.2
0.12
1
0.2
0.29 0.33
0.39
0.63
0
0.39
0.45
0.66
0.39
0.24
0.39
0.25
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
Interest Ra?o: Network 2
High_Int/R Med_Int/R Low_Int/R
(b) Network 2
0.2
0.12
1
0.2
0.29 0.33
0.4 0.62
0
0.39
0.45
0.66
0.39
0.24
0.38
0.24
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
Interest Ra?o: Network 3
High_Int/R Med_Int/R Low_Int/R
(c) Network 3
Figure 3: Interest Ratio: #Interesting(High,Med.,Low)Messages
#ReceivedMessages , Three Networks
0.93
0.12
1 1 1
0.59
0.95
0.32
0
1
0.81
0.63
0.93
0.15
0
1
0.42
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
Reachability RaBo : Network 1
Rcvd/High_Int Rcvd/Med_Int Rcvd/Low_Int
(a) Network 1
0.92
0.12
1 1 1
0.58
0.94
0.35
0
1
0.78
0.63
0.93
0.15
0
1
0.42
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
Reachability RaBo : Network 2
Rcvd/High_Int Rcvd/Med_Int Rcvd/Low_Int
(b) Network 2
0.91
0.12
1 1 1
0.59
0.92
0.34
0
1
0.79
0.64
0.92
0.15
0
1
0.42
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
Reachability RaBo : Network 3
Rcvd/High_Int Rcvd/Med_Int Rcvd/Low_Int
(c) Network 33
Figure 4: Reachability Ratio: #NodesReceivedMessages
#NodesInterested(High,Med.,Low)inMessages , Three Networks
– Low Interest Ratio: the ratio of low interesting
messages, with maximum interest factor value
max(I(m, n)) no more than 2.
The interest ratio for the whole social network is measured
as the average interest ratio of all nodes in this
network.
• Reachability Ratio, measures how many individuals actually
receive the message that they are interested in out
of all individuals who are interested in this message.
Readability Ratio(m) for a seed message m is calculated
as:
#Nodes Interested(High,Med.,Low) in m and Received m
#Nodes Interested(High,Med.,Low) in m
The reachability ratio for one experiment is the average
reachability ratio of all seed messages generated in this
experiment.
• Appreciation Ratio, measures how many messages
forwarded by a node s are appreciated by the receiver
node out of the total number of messages forwarded
by this sender node s. A message m is appreciated
by a receiver node n, if there exist a category j in
ISmn that the receiver’s interest factor fnj is no
less than 3 and the message’s relevance factor rmj
is no less than the receiver’s relevance tolerance fnj .
Appreciation Ratio(s, n) for a sender node s by
receiver node n is calculated as:
#Messages F orwarded by s and Appreciated by Receivern
#Messages F orwarded by s
Appreciation Ratio(s) for a sender node s is the average
appreciation ratio value by all its connected nodes.
In one experiment, the appreciation ratio of each node is
calculated, the average appreciation ratio of all nodes in
this network is measured too. In addition, all nodes in
this network are classified into three categories according
to their appreciation ratio values: above 0.6, between
0.3 and 0.5, below 0.3. Results are reported in Section
7.
• Message Node Ratio, measures the ratio of the total
number of messages to the total number of nodes in
the network, approximately the average number of messages
received by each node during the entire experiment
period. The Message Node Ratio relates to the cost
associated with each message forwarded in the network.
7 Simulation Results
Using the experimental set up described in Section 5, we
conducted 18 experiments with the six information-sharing
strategies proposed in Section 4 and three networks described
in Section 5. Each experiment is conducted with one of the
three networks, and one of the six strategies, which is used
by all nodes in the network. We collected all those measurements
defined in Section 6.
Figure 3 shows the comparisons of Interest Ratio for all six
strategies and for all three networks, Network 1, 2 and 3 from
0.00
0.73 0.63
0.07 0.06
0.95
0.66
0.18 0.36
0.62
0.88
0.05
0.34
0.09 0.01
0.31
0.05 0.00
0.32
0.45
0.6
0.32
0.46
0.75
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
Apprecia@on Ra@o and Classifica@on of Nodes,
Network 1
AI/FW >=0.6 AI/FW 0.3-0.5 AI/FW <0.3 AI/FW
(a) Network 1
0.00
0.52 0.60
0.00 0.07
0.95
0.73
0.43 0.37
0.64
0.89
0.05
0.27
0.04 0.03
0.36
0.04 0.00
0.32
0.46
0.62
0.32
0.46
0.76
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
Apprecia@on Ra@o and Classifica@on of Nodes,
Network 2
AI/FW >=0.6 AI/FW 0.3-0.5 AI/FW <0.3 AI/FW
(b) Network 2
0.02 0.09
0.63
0.02 0.08
0.72 0.80
0.85
0.35
0.68
0.89
0.26 0.15
0.06 0.02
0.30
0.03 0.05
0.33
0.46
0.62
0.32
0.47
0.75
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
Apprecia@on Ra@o and Classifica@on of Nodes,
Network 3
AI/FW >=0.6 AI/FW 0.3-0.5 AI/FW <0.3 AI/FW
(c) Network 3
Figure 5: Appreciation Ratio and Classification of Nodes Accordingly, Three Networks
70
22 30
73 60 36
227
77
106
234
191
124
398
142
186
414
342
226
0
100
200
300
400
500
600
700
800
S1_ELI S2_AI S3_HIR S4_UNI S5_CIR S6_MIR
#Messages/#Nodes
Network1 (582) Network2 (2000) Network3 (3000)
Figure 6: Message Node Ratio: #T otalMessages
#Nodes
left to right. Three measurements, High Int/R, Med Int/R,
Low Int/R, each represents the number of High, Medium and
Low interesting messages out of total number of received
messages, respectively. These three ratios add up to 1 by
definition. S3 HIR has High Int/R as 1, and S6 MIR has
Low Int/R as 0, both facts are consistent with how these
strategy work. Both S2 AI and S5 CIR have Low Int/R no
more than 0.25 (1/4), while both S1 ELI and S4 UNI have
Low Int/R greater than 0.33 (1/3). These facts are consistent
over all three networks.
Figure 4 shows the comparisons of Reachability Ratio.
Three measurements, Rcvd/High Int, Rcvd/Med Int, and
Rcvd/Low Int represent the ratio of the nodes who actually
received the message among all nodes that are High, Medium
or Low Interested in the message, respectively. S1 ELI,
S3 HIR, S4 UNI, and S5 CIR all have Rcvd/High Int close
to or equal to 1, meaning that they reach almost all nodes
that are highly interested in the message. In fact, S1 ELI and
S4 UNI reach almost all nodes that are interested in the message.
S5 CIR reaches about 80% medium interested nodes
and 40% medium interested nodes. S6 MIR has moderate
reachability in high and medium interested nodes as around
0.6, and it has 0 reachability among low interested nodes.
S2 AI has the lowest reachability, it only reaches about 12%
high interested nodes, 34% medium interested nodes and 15%
low interested nodes. Again, these observations are consistent
over all three networks.
Figure 5 shows the comparisons of Appreciation Ratio and
the classification of nodes according to their appreciation ratio.
S6 MIR has the highest average appreciation ratio as
≈ 0.75, followed by S3 HIR as ≈ 0.6, then S2 AI and
S5 CIR with ≈ 0.45. S1 ELI and S4 UNI have the lowest
appreciation ratio as ≈ 0.32. S6 MIR also has the highest
ratio of highly appreciated nodes, as 0.95 for Network 1 and
2, 0.80 for Network 3. S3 HIR has a consistent ≈ 0.60 ratio
of highly appreciated nodes over all three networks.
Some of the above facts are obvious given the definition
of the strategy, others are not. Most observations are consistent
over all three different network structures. Figure 6
presents the average number of messages received by each
node. Consider the communicational and computational cost
of information-sharing, S2 AI has the lowest cost, followed
by S3 HIR and then S6 MIR. The three high cost strategies
are S5 CIR, S1 ELI and S4 UNI being the most costly. Also
noted that the message/node ratio increases significantly as
the size of network increases, which can be explained by the
fact that the possible number of connections is the square of
the number of nodes in the network. Therefore, the choice of
information-sharing strategy becomes even more important
for large social networks.
Overall, S3 HIR has low cost, high interest ratio, perfect
reachability among highly interested node while zero among
other nodes, and moderate appreciation ratio ≈ 0.6. S6 MIR
has low cost, moderate interest ratio (no low interest message
delivered), and the highest appreciation ratio. S2 AI has
the lowest cost and moderate appreciation ratio, however, its
reachability is very low.
8 Conclusion and Future Work
In this paper we presented a graph model of social network
and a model of information relevance and node interest.
Based on these models, we proposed six information-sharing
strategies and defined a set of evaluation criteria including the
interest degrees, reachability, appreciation degrees and cost.
We conducted experiments to study the performance of each
strategy in three different networks. Some of the observations
are intuitive given how the strategies work, which in fact verify
that the simulation framework works correctly. In the future
we will study more realistic scenario, where each node
may choose different strategy and even dynamically change
its strategy responding to its environment , i.e. the number
of messages it receives. A node may also choose to different
response to received message depending on its source. We
also plan to model the real communication cost and computational
cost as a function of the number of messages in the
network, in order to study the performance of each strategy in
various settings. Intuitive conclusions rarely can be achieved
in such complicated setting, this experimental study framework
will be indeed appreciated. In the future, we will also
like to explore different methods of sharing information besides
sending messages, and study applying different strategies
with different sharing methods.
References
[Aggarwal and Zhai, 2013] Charu C Aggarwal and ChengXiang
Zhai. Mining text data. Springer Science & Business
Media, 2013.
[Ellison, 2007] Nicole B Ellison. Social network sites: Definition,
history, and scholarship. Journal of ComputerMediated
Communication, 13(1):210–230, 2007.
[Fortunato, 2010] Santo Fortunato. Community detection in
graphs. Physics reports, 486(3):75–174, 2010.
[Golbeck, 2013] Jennifer Golbeck. Analyzing the Social
Web. Elsevier Inc., 2013.
[Gruber, 2008] Tom Gruber. Collective knowledge systems:
Where the social web meets the semantic web. Web semantics:
science, services and agents on the World Wide
Web, 6(1):4–13, 2008.
[Guille et al., 2013] Adrien Guille, Hakim Hacid, Cecile ´
Favre, and Djamel A Zighed. Information diffusion in online
social networks: A survey. ACM SIGMOD Record,
42(2):17–28, 2013.
[Hogan, 2008] Bernie Hogan. Analyzing social networks.
The Sage handbook of online research methods, page 141,
2008.
[Horel and Singer, 2015] Thibaut Horel and Yaron Singer.
Scalable methods for adaptively seeding a social network.
In Proceedings of the 24th International Conference on
World Wide Web, pages 441–451. International World
Wide Web Conferences Steering Committee, 2015.
[Hudack et al., 2014] Jeffrey Hudack, Nathaniel Gemelli,
and Jae C. Oh. Modeling self-interested information diffusion
with game theory on graphs. In ICAART (2), 2014.
[Kempe et al., 2003] David Kempe, Jon Kleinberg, and Eva ´
Tardos. Maximizing the spread of influence through a social
network. In Proceedings of the ninth ACM SIGKDD
international conference on Knowledge discovery and
data mining, pages 137–146. ACM, 2003.
[Kim et al., 2010] Won Kim, Ok-Ran Jeong, and Sang-Won
Lee. On social web sites. Information Systems, 35(2):215–
236, 2010.
[Kobourov, 2012] Stephen G. Kobourov. Spring embedders
and force directed graph drawing algorithms. CoRR,
abs/1201.3011, 2012.
[Smith et al., 2009] Marc A Smith, Ben Shneiderman,
Natasa Milic-Frayling, Eduarda Mendes Rodrigues,
Vladimir Barash, Cody Dunne, Tony Capone, Adam Perer,
and Eric Gleave. Analyzing (social media) networks with
nodexl. In Proceedings of the fourth international conference
on Communities and technologies, pages 255–264.
ACM, 2009.
[Travers and Milgram, 1969] Jeffrey Travers and Stanley
Milgram. An experimental study of the small world problem.
Sociometry, 32(4):425–443, Dec. 1969.
[Zhang and van der Schaar, 2013] Y. Zhang and M. van der
Schaar. Strategic networks: Information dissemination
and link formation among self-interested agents.
IEEE Journal on Selected Areas in Communications,
31(6):1115–1123, June 2013.