Вы находитесь на странице: 1из 5

A Novel Social Influence Model based on

Multiple States and Negative Social Influences


Radhiyatammardhiyyah (1056037), Computer Science and Engineering
Yuan Ze University, Taoyuan City, Taiwan

AbstractNowadays, many social networks have said in their respective fields, this also can
been used by many citizens ranging from children influence them at the same time.
to adults. Since it gives a significant impact for In analyzing the influence of social
them, social influence becomes a hot research network, there are three highly representative
topic and the analytical research about it has been models. They are Linear Threshold (LT) Model,
studied extensively. There are many models to Independent Cascade (IC) Model and Heat
learn and predict the process of influence Diffusion (HD) Model. All those models classify
diffusion. This paper proposed a new propagation nodes into active and inactive states. Active node:
model that can improve from the previous This node has been inuenced, and has the power
method. Also, this work provides five states to to inuence others. Inactive node: This node has
represent the multiple states of influences where not received enough inuence to become an active
the impact of social influences will be reduced. node, and thus has no power to inuence other [1].
Each node has different level of influence and only
KeywordsSocial Influence analysis, Multiple the active nodes can give the influence to inactive
States, Social Network Influence, Negative social nodes. Those previous models face two problems.
Influences, Social Persuasion The first is they only contemplate the positive
influence possibilities. The second is the nodes that
have become active cannot return to the inactive
I. INTRODUCTION
nodes.
To solve those problems, this paper will
T he using of social media has been increasing
recently. The users come from different ages to
present a Multi-State Open Opinion Model (MOO
Model). It aims to stimulate several opinions which
use various social media. They used the following are held by users in the real world not in social
social websites and application such as Instagram, media. Moreover, this paper organizes the positive
Facebook, Twitter and LinkedIn. Because of this opinions into a Positive and Active (PA) state and
issue, many researchers tried to analyze the effect a Positive and Inactive (PI) state. Also, this work
of this kind of habit. Because social influence plays categorizes negative opinions into a Negative and
important role among citizens. Everything can be Active (NA) state and a Negative and Inactive (NI)
affected and changed through this social network state. Above and beyond positive and negative
such as habits, emotions and even beliefs. For states, there is also a Neutral (N) state which
Instance, by following the popular actors or represents the users with no specic opinion.
celebrity, users always follow up the information The rest of paper is organized as follows:
regarding them. The news about them were always Section II discuss the related works. Section III
shown on the news post because of his/her outward introduces the proposed method by which to
appearance. This situation will affect users in some propagate influence in MOO model. We propose
cases. Apart from those, we believe what our expert the experimental results in Section IV and the last
is Section V which concludes the paper.

1
Fig 1. Propagation of influence in LT model and IC model.

example is IC model which is described in the Fig.


II. RELATED WORK 1(c). it shows that each node affects neighbors
In this section, there are three models through inuential probability on the edge. In this
mentioned that will be discussed; Linear Threshold part, node C inuences node B and F
(LT) Model, Independent Cascade (IC) Model and successfully. Node B becomes active because of C
Heat Diffusion (HD) Model and their inuence but not A since the inuence is self-determining
propagation methods as well in order to provide between the two nodes.
deep insights. For the first example of LT model is Based on the previous work, Chen [2] projected
shown in the upper part of Fig. 1. It is assumed that an independent cascade model with negative
only node A is active and the other nodes are opinions, IC-N, which extends IC model. They
inactive. To illustrate the figure, node A begins to conclude an effective activation to two cases of
propagate inuence to nodes B and C through activation, positive activation and negative
edges, nodes B and C will get the inuence power activation. Finally, HD model characterizes
from node A as characterized by the blue inuences by heat features. Fig. 1(d) is an example
histograms. If a node receives a certain level of of HD model. In this figure, if we adopt the active
inuence greater than its personal threshold, the threshold is 3, node V1, V2, and V3 are active
node will be inuenced successfully, such as node nodes at the rst timestamp. The heat values of V1,
C in the step 1. In step 2, node C becomes active V2, and V3 are reduced since they share their heat
and starts to propagate its inuence to node B, D, values to other nodes. V4 and V5 receive the heat
E, and F. With the inuence power from node A values from their active neighbors V3 so that the
and C, node B becomes active too. In the last step, heat values of V4 and V5 increase. Therefore, V1,
nodes A, B, C, and F are active while node D and V3, V4 and V5 are active nodes after 10 iterations.
E, receiving inuence power less than the HD model also has some boundaries when negative
thresholds, remain inactive. For the second inuences try to be presented by heat resulting the
heat value is sharply decreased.
A. Multi-state Open Opinion Model
III. THE PROPOSED METHOD
There are two defined categories for social
This section explains about the proposed
networks; a weighted and directed graph. Those
method that will be used to define the attributes of
nodes have corresponded to all user media users.
MOO model.
Also, the engaged graph between two users can
direct to shows the connections for the target users.

2
To know further, following are some definitions The Diffusion Matrix (DM) above is as the
used in this model represent of the influence propagation. The
inuence propagation process can thus be
1. Opinion States (OS) formulated as the follow equation.
OS =[os1, os2, ..., osn], where osi is the opinion state
which is supported by user i. With respect to a where t is the current timestamp corresponding to
product or an event, people can have different kinds the current iteration, and t0 is the beginning time
of opinions. We categorize these diverse opinions at time zero. OV t is the normalized form of OVt,
into several opinion states. In this definition, there where ovi is from 0 to 1.
are five states;
Positive and Active (PA): Users hold positive
opinions and are willing to inuence others. C. Example of Multi-State Open Model
Positive and Inactive (PI): Users hold positive
opinions but do not inuence others
Neutral (N): Users without adequate opinion, and
thus do not inuence other users.
Negative and Inactive (NI): Users have negative
opinions but do not inuence others.
Negative and Active (NA): Users have negative
opinions and try to inuence other users.

2. Opinion V alue (OV)


OV=[ov1, ov2, ..., ovn], where ovi is the opinion
value of user i. An opinion can be quantied as a
value.

3. State Transition Table (STT)


The opinion value is mapped to opinion states Fig 2. Initial network and state threshold table
using a state transition table, which contains four
thresholds. STTi=[ PAi PIi NIi NAi].
The Fig 2 above is an example for MOO
model. To illustrate, there are four people as
B. Influence Propagation
variable V for each person. Every person adopts the
This part will describe a design of propagation same state transition table on the right for
method in estimating the opinion values of social simplicity.
media users and the updating of their current states Besides, the fig 3 below describes the
as well. alteration of OV and states from timestamp 0 to
stable timestamp 13. Until timestamp 4, OV4=
[1.00 0.96 0.53 0.39]T, and V4 receive enough
inuence to exceed the NI.

3
Fig 3. The Changing of opinion value and related states

kind of event. The upper-left pie chart illustrates


IV. EXPERIMENTS that 62% of people hold positive opinions for
This Part introduces the preliminary finding positive events. Additionally, there are 20% of
which is the dataset., users taking neutral opinions and 18% of users
having opposite feelings about these events.
A. The Preliminary Finding As mentioned before, the data was collected
This work used the dataset from eight events from the initial opinion states. Since then, this work
from Facebook websites including musical observes the inuence propagation and the change
performances, movies, public issues and even daily of states from users on the website used. On the
necessities. Before quantifying the STT from the other hand, MOO model gains the initial opinion
result of survey, a social influence website had values and predicts the nal opinion states with a
been built on Facebook. decay rate =0 .1 and relates with the correct nal
opinion states observed on the website used. The
All events mentioned above contain some precision measure is defined as:
selections with corresponding meanings with
respect to PA, PI, N, NI and NA. This work gives
the corresponding opinion values and opinion
states to each user depend on users selections. In where:
addition, this work assumes the opinion value of n is the number of nodes,
PA is 0.75, PI is 0.25, N is 0, NI is -0.25, and NA nc is the number of nodes whose opinions states are
is 0.75. The user not only can share and their predicted correctly.
opinions and comments with all their friends after
finishing the survey but also, they are allowed to n, nc, and precision of each state of each kind of
change their opinions as well after knowing their event are shown in the bottom half of Fig. 4.
friends opinion. The website investigated the Overall, the average precision value is 60.6%,
changes of these events in the period from May, including positive, negative and neutral events.
2014 to June, 2014.

B. Prediction Result Using MOO Model


Eight events have conducted in this paper. They
are divided into several parts. They are four
positive events, three negative events and one
neutral event.

Fig. 4 below is shown for the network


topology. In this figure, the upper half indicates the
average distribution of users initial states in each

4
REFERENCES

[1] Huang. Jen-Wei, Ma, Hao-Shang, Chen,


Yuan-Chang. Multi-state Open Opinion Model
based on Positive and Negative Social Influences.
2015 IEEE/ACM International Conference on
Advances in Social Networks Analysis and Mining

[2] W. Chen, A. Collins, R. Cummings, T. Ke, Z.


Liu, D. Rincon, X. Sun, Y. Wang, W. Wei, and Y.
Yuan. Inuence maximization in social networks
when negative opinions may emerge and
propagate. In SIAM International Conference on
Data Mining, pages 379390, 2011.

Fig 4. The network Topology

V. CONCLUSION
This paper proposed a new social network
influence model. The used method is based on
multiple state and negative influences. Regarding
the previous method, three related works; Linear
Threshold (LT) Model, Independent Cascade (IC)
Model and Heat Diffusion (HD) Model only
consider positive influence and only two states for
each user. It becomes the cons of those works
because there is no chance to return to inactive
states especially when some nodes are influenced
as active nodes.
After applying the proposed method (MOO
model), those limitations have been broken. MOO
model has five opinion states and by having this,
this model gives permissions for users to adjust
these states using state transition table. Moreover,
the function of diffusion matrix applied in this
paper shows the successful in predicting the
influence diffusion even after some iterations.
To conclude, this work has successfully applied
a social influence website in order to observe the
state transition habit. Those divided in three types
events; positive, negative and neutral.

Вам также может понравиться