CHAPTER 5 SOCIAL INTELLIGENCE MECHANISM
5.1 Scenario of Social Intelligence Mechanism
5.2.4 Alternative Synthesizing
In the decision criteria discovering module, the decision criteria conforming to a decision problem has been identified. Hsu and Chen [34] proposed a fuzzy similarity aggregation method (SAM), in which similarities between decision group members were collated and fuzzy numbers assigned directly to each member to determine the agreement degree between them. Taking the degree of importance of each member into consideration, the original weighting method was modified as below. In Hsu and Chen [34], the average agreement degree of member i i, 1, 2,...,n is given by:
60
However, the relative importance of experts varies. Some are more important than the others, and some are more experienced than others. Therefore, the relative importance weight of each expert was considered. The most important person among experts was selected and assign him weight one. The relative importance of expert i is formulated as [34]:
For RI i , in the original definition the weight of the most important member is 1, that ( ) is, ( ) 1r i . Then the kth member is compared with the most important one, and a relative weight r k( ) is assigned. This work improved the calculation of relative importance of decision group member and consensus degree coefficient was improved to capture the spirit of social network. Since the decision group was selected based on friendship, the member with highest friendship index is considered to be the most important one with r i( ) 1 , for all other members, r k( )FE i kt( , ) FE i jt( , ) . Therefore, the relative importance of members was reformulated as:
1
Finally, the aggregation result R can be defined as [34]:
1
61
where R i i( )( 1, 2,..., )n is the estimated ratings of decision group member E i( ) on decision criteria.
Having decision criteria in hand, fuzzy AHP was used to construct the alternatives of decision problem. In the proposed system, the same fuzzy AHP process from prior works [13, 14] was adopted, and the important data required to complete this process is describe as follows:
1. the problem hierarchical structure: the decision problem is structured hierarchically at different levels, each level consisting of decision criteria. The top level of the hierarchy represents the overall goal of decision problem, while the lowest level is composed of all possible alternatives. One or more intermediate levels embody the decision criteria and sub-criteria. In the proposed system, the decision criteria and the possible alternatives are obtained from decision criteria discovering module.
For the simplicity of system implementation, if a decision group member proposed possible alternatives in his comments, the alternatives are tagged with predefined syntax so that they can be easily identified and avoid confusion with decision criteria.
2. the pair-wise comparison matrix: the relative importance of the decision criteria is assessed by using equation (5.15).
At the end of this process global priorities are used for final ranking of the alternatives.
5.3 Experiment
5.3.1 Experiment Process
To further prove the feasibility, an empirical study alone with system development was conducted. The procedures of this experiment are depicted in Figure 5.7. To implement proposed mechanism, one of the most popular social network sites Facebook was selected to be experiment platform to collect required data. First, a linear regression model is required to measure friendship index between friends. For the purpose of collecting basic data required by building regression model, a group of social network users were invited to be participants. By using 3 (S) stages 3 (K) names snowball sampling, 96 participants for each social network (group) were invited. After filtering the people not interested in the experiment, finally 156 participants, 52 for each group, were included in the experiment. Of all the 156 participants, 50 users were randomly
62
selected to collect data needed for constructing friendship regression model. The average year of Facebook usage is 1.6 years and the average number of friends is 205.
The characteristics of these social networks are summarized in Table 5.2.
Start Snowball sampling to
Figure 5.7 Experiment process for social intelligence mechanism
Table 5.2 Characteristics of the three networks
ATTRIBUTES
SOCIAL NETWORKS
STUDENT OFFICE WORKER RANDOM MEMBER
Number of participants 60 60 60
Average betweenness centrality 37.901 40.103 38.221
Average closeness centrality 56.919 56.956 57.440
Average distance 1.758 1.757 1.742
Since the closeness centrality, betweenness centrality and social similarity can be calculated without participants’ help, only the data required to build up friendship estimation regression model has to be collected. The social relationships index (SRI) was developed as a self-report version of the social support interview [11, 12, 66], and this scale has demonstrated good test–retest reliability and internal consistency [79].
An online survey package, including a cover letter explaining the research objectives
63
and the questionnaire was distributed to the 50 users to survey their friendship with those friends who also participated in the experiment. These friends were then rated in terms of how helpful and upsetting they were (1 = not at all, 10 = very much) when the participant needed support such as advice, understanding, or a favor. Using responses on this measure, friendship quality was classified as “supportive” or “ambivalent” as described by Uno, Uchino, and Smith [81]. For example, if user A was selected to provide friendship data for the purpose of regression model building, and his friends B,C and D also participated in the experiment, then he will receive an online SRI questionnaire survey to rate his friendship with B,C and D. After collecting the friendship dataset, SPSS was used to build the regression model by entering friendship from SRI, social similarity, closeness and betweenness centrality as base data. This model was then used to predict the friendship index at the time when a decision problem is issued.
After the regression model was built, 18 randomly selected users (6 from each group) were invited to be decision makers. In the experiment, they can issue a decision problem and evaluate the effectiveness of decision criteria and alternatives. The decision makers were asked to issue 2 decision problems during the experiment, and these problems were delivered to the decision group members selected by the system.
Users with top-5 friendship index were selected as decision group members, and the processes repeated every time when a problem was issued. When a decision problem was presented to decision group, members can make their comments online. They were asked to post their comments together with alternatives about the decision problems.
Every alternative was asked to provide a hyperlink containing related information, so that the click stream data can be collected to compare with other methods. Follow the method proposed in this research, decision criteria and alternatives are collected and presented to decision makers.
To maintain basic requirement of Delphi method, during the process, individual comments are unknown to others. The alternatives collected from both this work and benchmark methods (described in next section) are presented to decision makers in the same page, and the links are listed randomly to minimize interfere of presentation order.
To avoid information overloading, the first decision alternative of each method were selected, and total 4 alternatives were presented to decision makers for each problem.
Finally, the click count and stay time of each page linked to alternatives were recorded,
64
and satisfaction was rated on a 5-point Likert Scale for each alternatives presented to decision makers: Very Useful, Useful, Neither Useful nor Useless, Useless and Very Useless by a rating score of 5, 4, 3, 2 and 1. The experiment was executed in each of the three group 3 times, and each experiment lasted for one week. The parameters of this experiment are shown in Table 5.3.
Table 5.3 Experiment setting of social intelligence mechanism
ITEM SETTING
Type of support Product selection criteria and candidates listing Participant sampling Snowball sampling
No. of participants Office worker:6 (out of 52) Random member: 6 (out of 52)
Regression model Social relationships index survey: 50 participants (randomly selected)
Benchmark method
Random: rank products in candidate list randomly
Social network analysis: select support group members by SNA Group centrality: select support group by group centrality Evaluation method Clickstream: browsing time on the product pages provided
Perceived helpfulness: questionnaire survey by 5-point Likert scale
5.3.2 Benchmark Methods
To compare proposed mechanism with others, three methods were selected as benchmark.
1. Random: this method was used as baseline benchmark. All the experiment process was the same as proposed mechanism except the decision group members were randomly selected from the friends of decision makers.
2. SNA: this method used social relation data to be selection rule. The friendship estimation was derived purely from participants’ profile. This is a baseline to test if it is useful to include time factor when estimating friendship to select decision group members. The regression model used in this method can be rewritten based on equation (5.7) as:
0 1 2
( , ) ( ) ( , ).
SSC i jt C j SS i j (5.16)
65
3. SNA with single point friendship estimation: to see if the proposed friendship estimator (equation (5.5)) is better than using single point friendship estimation (equation (5.3)), a single-point friendship estimator regression model was also included in the benchmark methods. The regression model can be formulated based on equation (5.7) as:
1
0 1 2 3 4
( , ) ( ) ( , ) t ( ).
t ij P
SSC i j C j SS i j PK i (5.17)
5.4 Result and Discussion
To evaluate the result from selection criteria discovering module in the proposed mechanism, the precision and similarity between keywords from Epinions.com and extracted selection criteria were calculated based on the product information page of Epinions.com. For example, resolution, camera type, optical zoom and LCD screen size were listed in the Epinions.com as product selection condition for digital camera.
These keywords were used to expand their synonym set, and this set was used to compare with the possible selection criteria extracted from proposed mechanism.
Illustrations of these processes are depicted in Figure 5.8 and Figure 5.9. The average precision rate is 0.63, and the average word similarity is 0.61.
Table 5.4 The coefficients of regression models
Factor
Proposed Mechanism SNA SNA + Single Point
Coefficient( )
Constant 0.375 5.128 5.509
Similarity 0.361 0.777 0.596
Friendship 6.385 - 0.827
Centrality 0.607 0.841 0.802
Expertise 2.356 - 2.431
66
WordNet Similarity WordNet
Synsets
Resolution Optical Zoom
Features Screen Size
Weight ...
Epinions
MegaPixel Size Weight Zoom-in/Zoom-out
...
Social Intelligence
Figure 5.8 Precision and similarity calculation process of selection criteria
67
Figure 5.9 Similarity comparison of Epinions keyword and extracted criteria
During the experiment, clickstream data of every alternative presented was collected, and the average stay time of different methods and groups on every alternative was plotted. As shown in Figure 5.10 , the proposed mechanism attracted decision makers to spend more time on the alternatives than other methods. More, as shown in Figure 5.11, the average usefulness level of alternatives generated by the proposed mechanism is also higher than other methods. To further examine if there are significant differences in average stay time and average usefulness level, a statistical method is required.
68
Figure 5.10 Average stay time for different groups and methods
Figure 5.11 Average usefulness level for different groups and methods
69
Two-way analysis of variance (ANOVA) is a statistical analysis in which two independent factors are examined with regard to their impact on a dependent variable and on one another. To test the impact of method used and user group on average stay time and average usefulness level, in this work two-way ANOVA was used. Asseen from Table 5.5, the method used in the experiment has impact on the average stay time as the test result is significant at 0.05 (as 0.00<0.05). In contract, the user group has no impact as 0.51>0.05. For the same reason, based on Table 5.6 the satisfaction can only be influenced by method used during the experiment.
Table 5.5 Tests of between-subjects effects for average stay time
Dependent Variable: Average Stay Time
SOURCE TYPE III SUM OF SQUARES DEGREE OF FREEDOM MEAN SQUARE F SIG.
User Group 0.52 2 0.26 0.66 0.51
Method 520.38 3 173.46 444.41 0.00
Table 5.6 Tests of between-subjects effects for average usefulness level
Dependent Variable: Average Usefulness Level
SOURCE TYPE III SUM OF SQUARES DEGREE OF FREEDOM MEAN SQUARE F SIG.
User Group 0.40 2 0.20 0.61 0.54
Method 657.94 3 219.31 673.96 0.00
Post hoc tests such as Tukey's test most commonly compare every group mean with every other group mean. Knowing that the methods used in the experiments could affect stay time and usefulness level, Tukey’s test was used to see if there is a significant difference between different methods. Table 5.7 and Table 5.8 show that there are significant differences between the proposed mechanism and other benchmark methods; also the average stay time and average usefulness level are higher than other methods. Since there is a strong tendency for users to spend a greater length of time reading articles of interest to them [32, 56], it is likely that the proposed approach is more effective when compared with other methods.
70
Table 5.7 Multiple comparisons of stay time
(I) METHOD (J) METHOD MEAN DIFFERENCE (I-J)
Random
SNA -0.4815*
SNA + Single Point Estimation -1.4417*
Proposed Mechanism -1.9736*
SNA
Random 0.4815*
SNA + Single Point Estimation -0.9602*
Proposed Mechanism -1.4921*
SNA + Single Point Estimation
Random 1.4417*
SNA + Single Point Estimation 0.5319*
*. The mean difference is significant at the .05 level.
Table 5.8 Multiple comparisons of usefulness
(I) METHOD (J) METHOD MEAN DIFFERENCE (I-J)
Random
SNA -.7037*
SNA + Single Point Estimation -1.8796*
Proposed Mechanism -2.1481*
SNA
Random .7037*
SNA + Single Point Estimation -1.1759*
Proposed Mechanism -1.4444*
SNA + Single Point Estimation
Random 1.8796*
SNA + Single Point Estimation .2685*
*. The mean difference is significant at the .05 level.
To test the influence of friendship index in the selection of decision group, all the decision problems were issued at different time period. Since the selection rule of decision group was based on friendship index which is influenced partially by time, this study investigated if the members change across different time. For example, when
71
decision maker A issued his first problem, a decision group consisting of 5 members was built, say(M M1, 2,M3,M4,M5). At the time when second problem was issued, another decision group, say (M M1, 2,M3,M7,M8), was built. In these two decision process, there were 10 users selected as decision group, but only 7 unique users since
1 2 3
(M M, ,M )was overlapped. The detail numbers are listed in Table 5.9. Further analysis found that decision makers with a large number of unique users are more active than those who with small number. However, no evidence showed that there was significant difference of time spent on the alternatives suggested by different decision groups even if there exists different number of unique users. This observation showed that the proposed mechanism was stable enough and won’t be influenced by the uniqueness of decision group.
5.5 Chapter Summary
In this chapter, time factor was introduced into social network analysis. From the viewpoint of academic contribution, by using regression a friendship estimation model was proposed to predict friendship between two users in specific time period. By equipping FDM with online decision criteria filtering mechanism, time consuming problem of conventional Delphi method was solved. Furthermore, an adoptive SAM was also suggested to further improve the application of FAHP on social network related research. An empirical study further proved the feasibility and effectiveness of this work. This research successfully introduced the decision process theory and social psychology into the development of social network-based application. Besides, the concept of decision support system development was extended to utilize social network platform. From the viewpoint of practice, a feasible way to develop a social network-based decision support system together with the related techniques was demonstrated. By dividing the system framework into modules, those who are interested in developing such kind of applications can further improve the system by plugging in new modules as needed.
72
Table 5.9 The number of unique decision group members
DECISION MAKER
TOTAL NUMBER OF USERS SELECTED AS DECISION GROUP
NUMBER OF UNIQUE USERS Student 1
30
( 5 decision group members for each problem, 2 problems for each experiment, 3 experiments, 5*2*3=30)
23
Student 2 20
Student 3 24
Student 4 19
Student 5 15
Student 6 19
Office Worker 1 16
Office Worker 2 25
Office Worker 3 19
Office Worker 4 19
Office Worker 5 22
Office Worker 6 16
Random Member 1 19
Random Member 2 17
Random Member 3 23
Random Member 4 16
Random Member 5 22
Random Member 6 15
73