Fitness Aggregation

Although the user preference could be observed from the posts, the importance levels of these posts should differ when they are used for evaluating the user preference fitness to the product. For example, two articles that are highly correlated with the product were posted yesterday and three months ago. The former means the user is

-53-

focusing on the related product information now, so that the user would be more willing to adopt and share the product information. However, the latter might reflect that the user had once surveyed the related information, but might not still be interested in the related product information if his/her focus has changed recently. A preference weighting function for article i, which is decreasing by time, is defined as:

( )_i 1

i

pw t

 ,

t

(4.6)

where

t is denotes the time periods since article i was posted. For example,

_i

t

_i  1 indicates that the article was posted within one recent month. Finally, the preference fitness of user

u to the product is formulated as: 4.1.2 Transition Flow Inference Module

The basic concept of the Markov model is to determine the transition probability of transitions from one state to another. In the context of a social network, a state stands for a user and the transition between two states is interpreted as interaction between two users. Specifically, the transition probabilities between possible states are estimated according to social interactions.

4.1.2.1 Interaction Network Construction.

We leverage the social interaction data from online social networks to obtain the set of active social nodes with respect to a specific user and use the identified nodes as the possible transition states from the current state (the specific user). When the circle of people’s friendship grows, there is an increasing need for friend management. Research by Dunbar [30] indicates that there is an approximate natural group size in which everyone can really know each other. Though one can have hundreds of online friends, most of them are just a name in one’s friend list and do not incur any social interaction.

A recent study also shows that social media users have a very small number of offline

-54-

friends compared with the number of online friends they declare [48]. We construct a network of social interactions to filter out the active friends of a user and use these nodes as the possible information transition states. Specifically, the directed interaction network of a specific user is constructed by analyzing the social interaction data collected from his/her micro-blogosphere. The edge direction of the interaction network represents the direction of the interaction flow. When a user posts a micro-blogging message, he/she is likely to expect some responses. In the current paper, we define a micro-blogging message poster and replier as “interaction requester” and

“interaction provider,” respectively. For example, as illustrated in Figure 4.3,

u ,

_A

u

_B and

u post message in the micro-blogosphere, which means

_C

u ,

_A

u and

_B

u are

_C interaction requesters.

u replies to all of them, implying that

_D

u is an interaction

_D provider. Consequently, there would be “interaction” flowing from

u to

_D

u ,

_A

u and

_B

u .

C

Figure 4.3 Directed interaction network

4.1.2.2 Transition Probability Inference

After obtaining the set of active social nodes (possible transition states), the following formulation is used to determine the transition probability between states.

 

^{i j}

interaction transition probability from

u to

_i u_j , and

i j

ju u^{}

 denotes the total  

-55- interaction network is represented as a transition matrix (TM ):

 

, where

m denotes the total number of active social nodes.

4.1.3 Customer Value Evaluation Module

The purpose of this module is to evaluate the network-structure-based measurements:

influenceability and reachability. In this module, the friendship network constructed by the friend list in the micro-blogosphere is used to obtain the eigenvector centrality and reach centrality to evaluate the influenceability and reachability, respectively.

First, the friend network is represented as a bipartite graph G( , )V E , where V

denotes the vertices in the network and E denotes the edges between V . Next, for the influenceability and reachability analysis, G is transformed to an adjacency matrix

( _{v t},)

A



a

, if vertex v and vertex t are connected,

a

_{v t,}=1, otherwise

a

_{v t,}=0. In this research, we use UCINET to compute the following two measurements of centrality.

4.1.3.1 Influenceability Analysis

For business, the greatest interest of the marketers is to know how many purchase intentions of potential consumers could be stimulated by the marketing information they receive. In this respect, the influence of a node plays an important role in enhancing the diffusion effectiveness of marketing information for the purpose of seeking business opportunities. Kiss and Bichler [62] compare plenty of measures of influence including different centrality measures in customer networks and suggest that the eigenvector centrality is one of the effective measures for estimating the influence of a node in a network. In the current research, the eigenvector centrality is used to compute the influenceability of the users. Conceptually, different neighbors may have different values contributing to the eigenvector centrality. That is, the eigenvector centrality of user

u is contributed to by the eigenvector centrality of the connected

_i neighbors of

u . The eigenvector centrality of

_i

u is determined as:

_i

-56-

For the purpose of comparisons within a graph, it is appropriate to use the eigenvector centrality with maximum normalization [104], which is derived as:

( ) ( )

From the network structure, a person with higher centrality could influence more other nodes in a social network. Besides, a person is influenced by another through the social interactions between them. Therefore, the influenceability of

u is measured as:

_i

( )_{i norm}( )_i ( )_i

IA u



ec u



asn u

, (4.11)

where ( )

asn u is the total number of active social nodes with respect to

_i

u

_i

4.1.3.2 Reachability Analysis

In relation to establishing brand expression, determining how many potential consumers can be reached during the marketing information diffusion process is what marketers care about most. Hanneman [42] suggest the m-step reach centrality [13] to measure the reach efficiency (e.g. the portion of all others whom one can reach in a network). In the current research, the m-step reach centrality is used to evaluate the reachability of

u . The m-step reach centrality measures the number of reachable

_i nodes within m steps from a given social node. That is, reachability indicates how many users

u could reach on average per step. The reachability of

_i

u is measured as:

_i

-57-

marketers. According to the small-world effect [92], the value of

n is no need to be

greater than 6.

4.1.4 Diffusion Path Planning Module

4.1.4.1 Sharing Behavior Analysis.

The expected value of the diffusion reward is impacted on by the willingness-to-share of social nodes. Despite a node obtaining higher influenceability and greater reachability than others, he/she might just like to engage in daily chat and specific conversations with someone but may not like to share information in the micro-blogosphere. If the diffusion path plans to pass through him/her, it will be easily interrupted. Due to the small character limit (140 characters) in the micro-blogosphere, a URL is frequently used to conduct information sharing behavior. On the other hand, a message is external information sharing from other sources, if it contains a URL in a micro-blogging message. The degree of daily sharing behavior of a social node is measured as:

( )_i ^http

post reply

sb u



    , (4.13)

where _post and _reply denote the total number of messages posted and the total number of message replied to others by

u respectively.

_i _http denotes the total number of messages containing at least a URL in _post and _reply .

According to previous survey [46], egoism and altruism are two significant motivations of users who are willing to share information. Egoism refers to users who would like to share the information for which they have preferences with their friends because they expect the sharing behavior to enhance their personal reputation. Altruism referred to users who are willing to increase the welfare of their friends without expecting returns, so users would like to share information with friends because they might know their friends’ preferences. The tendency of willingness-to-share of

u is defined as follow:

_i

( )_i ( )_i ( )_i

wts u



PF u



sb u

. (4.14)

-58- 4.1.4.2 Diffusion Path Analysis

In the proposed APPM, we have combined the probability of state transition, the tendency of willingness-to-share, and the diffusion reward function as treatments to explore the diffusion path with the highest diffusion reward. First, we define the diffusion reward function as:

( )_i ( ) (1_i ) ( )_i

DR u

 

IA u

  

RA u

, (4.15)

where  is the information diffusion strategy weighted to balance the performance of influenceability and reachability, which is determined by the focus of marketing strategies (business opportunity seeking or brand awareness). The direct reward coming from neighbor node i to starting node scan be formulated as:

   

_ , _{r s i} ( )_i ( )_i

Neighbor DR s i



P u u

 

wts u



DR u

. (4.16)

The total reward generated from diffusing the information through the planned optimal path, which starts from node s, is defined as :

 

_{ }

     

Path s consists of a sequentially selected key endorser node in the social network.

 

_ ,

Path Length s i denotes the path length between node s and node ˆl stands for the maximal length of a planned path. Notice that ^{TR s}

 

is the conservatively estimated reward of the diffusion process along the path starting from node s. That is, if the marketing information could be disseminated by following ^{Path s}

 

, the marketer could gain the diffusion reward at least as ^{TR s}

 

. If some of the nodes who are included in ^{Path s}

 

are additionally willing to pass the marketing information to other people who are not included in ^{Path s}

 

, the real diffusion reward will be greater than

 

TR s .

-59-

Note that the model could be easily extended to multiple paths starting from node s. For example, in Figure 4.1, if we can revise the reward function to use the maximal and sub-maximal values at the same time to plan the path, the path of e₂ would be extended to multiple paths (starting from k₁ and u₄, respectively), as shown Figure 4.4. However, the diffusion reward would be greater than that in the single path planning. Generally, the choice of the number of neighboring nodes to forward to is determined by the total cost of the incentive to induce message forwarding, which increases as the number of endorsers becomes larger.

Figure 4.4 Example of multiple path planning 4.2 Experiments

4.2.1 Experiment Source

In this section, we apply the proposed mechanism to the micro-blogging system to examine its effectiveness. Micro-blogging services are one of the top tools for social media marketing. We use Plurk, one of the most popular micro-blogging services, as the platform for conducting experiments. Currently, Plurk is very popular in Asia and the United States [7]. It allows users to send and respond to messages in short sentences (with a limitation of 140 characters). Besides, it attracts users to communicate with each other and share external information by embedding URLs. Because Plurk is popular and predominantly used for communicating and sharing, it is an excellent platform for marketers to conduct information diffusion while conducting social media marketing.

In the experiment, 131 active Plurk users were invited to be participants, and they were also the candidates for the start point of a diffusion path. Firstly, for the purpose of constructing the interaction network to obtain the transition probability, we collected the last 6 months’ micro-blogging messages (including post and response data) from

 

-60-

participants’ public Plurk interface. Then, with the purpose of constructing the friendship network to obtain the information influenceability and reachability, we recursively expanded friendships from the participants’ friend list. Finally, there were 4,832 social nodes included in the friendship network. The information on the collected social network data is outlined in Table 4.2. Figure 4.5 shows the visualization of the collected friend network.

Table 4.2

Data descriptions of the experiment Statistics of the experiment data

Number of invited participants 131

Average number of friends per participant 37

Average number of active social nodes per participant 11 Average number of monthly interactions per participant (6 months) 2,147

Figure 4.5 Visualization of collected social diffusion network 4.2.2 Experiment Design

In the experiment, we diffused 40 pieces of marketing information in total via 2 different marketing strategies: (1) seeking business opportunities and (2) establishing brand expression. According to previous studies, coupon promotions could cause an increase in product sales [8] and the product reviews from third parties might spread

-61-

good news/impressions of brands so that it can increase the effectiveness of firms’

advertising [18]. There were in total 20 product deals/coupon advertisements for seeking business opportunities and 20 product evaluation review articles for establishing brand expression. The former marketing information was collected from Yahoo! Shopping, which is one of the largest online shopping sites, and the latter was collected from Epinions, which is one of the most professional and famous product review platforms allowing users to share their product experiences and opinions. In order to perform the preference fitness analysis, the keywords that can best represent the product are needed. In our experiments, the keywords of marketing information were provided by an expert group made up of six senior graduate students and four doctoral students in business colleges. The advertisements were delivered with an online 5-star rating questionnaire for the marketing information receivers to feed back their acceptance and diffusion path tracking (Which friend was the marketing information received from?).

We evaluated our proposed mechanism by comparing with the following benchmark approaches: (1) random advertising without a path planning mechanism (Random), (2) random advertising with a path planning mechanism (Random+Path), (3) influencer advertising without a path planning mechanism (Influencer), and (4) influencer advertising with a path planning mechanism (Influencer+Path). According to Kiss and Bichler [62], out-degree centrality produces better performance in influencer identification, so we used out-degree influencer selection to select the starting point of information diffusion. Besides, the random advertising method randomly selects participants whose sb u( )0 as the starting point of the information diffusion process.

For each advertising method, we selected five participants as starting points for diffusing the marketing information.

4.3 Results and Evaluations

In order to evaluate the performance of different advertising methods, we use the click-through rate (CTR) of the advertisements and the receivers’ five-star acceptance rating feedback on the received marketing message as the evaluation indicators. The former is a popular practical indicator of advertising efficiency; the latter could evaluate the users’ impression of the marketing message received.

-62-

Intuitively, for the purpose of seeking business opportunities, it is expected to seek the potential customers with high ( four-star) acceptance of the product advertisement, and for establishing brand expression, it is expected to seek the potential customers with not the lowest ( two-star) acceptance of the product advertisement. We compare the performance using CTR with different star rating conditions.

4.3.1 Seeking Business Opportunities Strategy

Generally, business opportunities exist in the potential customers with high acceptance of product advertisement, which means that they have a higher chance of buying products. The CTR with the acceptance condition formula is defined as:

4

click star

ad

CTR    ^

  , (4.18)

where  denotes the total number of delivered advertisements, _ad _click denotes the total number of clicked/read advertisements, and _4star denotes the total number of receiver rating  four-stars acceptance.

Figure 4.6 shows the CTRs of each step with respect to the different benchmark methods. After 4 steps forward, the 20 advertisements in “Random” and

“Random+Path” respectively diffused 583 and 776 times in total and received 0.120 and 0.216 CTR, which means that our path planning mechanism improved by approximately 10% the chance for seeking business opportunities. The advertisements in the “Influencer” and “Influencer+Path” respectively diffused 852 and 1,067 times in total and received 0.264 and 0.347 CTR, which means that our path planning mechanism improved by approximately 8% the chance for seeking business opportunities.

-63-

Figure 4.6 CTR in seeking business opportunities.

Furthermore, a 95% significance level two-paired sample t-test is used to evaluate the overall performance of different advertising strategies. The results are shown in the following Table 4.3. First, the test results show that the proposed path planning mechanism significantly improved the benchmark advertising methods. Besides, the diffusion effectiveness was also significantly improved if the path planning started from qualified starting points.

Table 4.3 Statistical verification of the CTR under seeking business opportunities strategy

Random+Path V.S. Random 0.098 0.123 0.027 3.604 0.002 Influencer+Path V.S. Influencer 0.109 0.202 0.045 2.392 0.027 Influencer+Path V.S. Random+Path 0.172 0.224 0.050 3.433 0.003

4.3.2 Establishing Brand Expression Strategy

The purpose of this marketing strategy is to enhance (four to five stars) or reverse (two to three stars) the brand expression of customers. However, it is very hard to reverse the

0.120

-64-

brand expression of antis (zero to one star). It might even have the opposite effect in marketing strategies. The CTR with the acceptance condition formula is defined as:

2

click star

ad

CTR    ^

  , (4.19)

where _ad denotes the total number of delivered advertisements, _click is the total number of clicked/read advertisements, and _{2 star} denotes the total number of receiver ratings two-star acceptance.

The Figure 4.7 shows the CTR using different benchmark methods. The 20 advertisements in “Random” and “Random+Path” diffused in total 985 and 1,243 times and received 0.160 and 0.221 CTR, which means that our path planning mechanism improved by approximately 6% the chance for establishing brand expression. The advertisements in the “Influencer” and “Influencer+Path” respectively diffused 1,601 and 1,887 times in total and received 0.252 and 0.321 CTR, which means that our path planning mechanism improved by approximately 7% the chance for establishing brand expression. Finally, the result of the overall performance of different approaches is further evaluated by two-paired sample t-test and shown in Table 4.4. At the 95%

significance level, all the test results show that the proposed path planning mechanism significantly improves the other advertising approaches.

Figure 4.7 CTR in establishing brand expression.

0.160

-65-

Table 4.4 Statistical verification of the CTR under establishing brand expression strategy

Paired Group Mean Std

Deviation

Std Error

Mean T value

Sig.

(2-tailed) Random+Path V.S. Random 0.074 0.131 0.029 2.529 0.020 Influencer+Path V.S. Influencer 0.072 0.124 0.027 2.609 0.017 Influencer+Path V.S. Random+Path 0.076 0.106 0.023 3.206 0.005

4.3.3 Exposure Ability in Different Strategies

Advertisers are concerned about the effective exposure for their advertisements. The proposed APPM would plan a suitable diffusion path for advertisements following different strategies. In one of the diffusions, the total number of message receivers in addition to the people who are included in the planned diffusion path gives the message exposure range of path planning. For instance, as shown in Figure 4.1, the nodes

u ,

₁

u ,

₂

u , and

3

u are the exposure range of the planned diffusion path. Because the path was

₄ broken by node

k (

₂

k delivers the marketing information to nodes

₂

u and

₁

u rather

₂ than the planned node

k ) and the system respectively replans the diffusion path for

₃

u

₁ and

u , the planned diffusion paths of the diffusion would be adjusted as shown in

₂ Figure 4.8.

Figure 4.8 Adjusted diffusion path

However, the replanned diffusion paths still belong to the same marketing information diffusion process. The eventual number of message receivers of the diffusion is an important indicator for evaluating the performance of the planned diffusion path. The

 

-66-

exposure ability (EA) is the average number of receivers of marketing information and it is formulated as follows:

receivers mi

EA 

  , (4.20)

where _receivers is the total number of receivers in addition to the path nodes and

 denotes the total amount of delivered marketing information. EA is the average mi

number of receivers per marketing information.

From Figures 4.9 and 4.10, we observe that the proposed APPM could enhance the exposure ability of product advertisements, if we ignore the acceptance of product advertisement. For the random advertising method, after forwarding for 4 steps, the APPM respectively improves by approximately 33% and 26% the exposure ability of the random advertising method in the seeking business opportunities strategy and in the establishing brand expression strategy. For the influencer advertising method, the APPM respectively improves by approximately 25% and 22% the exposure ability of the random advertising method in the seeking business opportunities strategy and in the establishing brand expression strategy.

Figure 4.9 Exposure ability in seeking business opportunities strategy.

29.2

38.8

42.6

53.4

0 10 20 30 40 50 60

Random Random+Path Influencer Influencer+Path EA

Methods EA

-67-

Figure 4.10 Exposure ability in establishing brand expression strategy.

Here, the paired sample t-test is also performed to provide further confirmation of the significant difference in the results of the benchmark approaches under different strategies, as shown in Table 4.5 and Table 4.6. At the 95% significance level, all the test results show that the advertising strategies with the APPM significantly outperformed the advertising strategies without the APPM. Therefore, they prove that our proposed strategy is the best compared with other strategies.

Table 4.5 Statistical verification of the EA under seeking business opportunities strategy

Random+Path V.S. Random 9.65 14.01 3.13 3.080 0.006 Influencer+Path V.S. Influencer 10.75 18.81 4.21 2.556 0.019 Influencer+Path V.S. Random+Path 14.55 17.36 3.88 3.748 0.001

49.3

-68-

Table 4.6 Statistical verification of the EA under establishing brand expression strategy

Paired Group Mean Std

Deviation

Std Error

Mean T value

Sig.

(2-tailed) Random+Path V.S. Random 12.90 10.03 2.244 5.748 0.000 Influencer+Path V.S. Influencer 16.80 19.71 4.406 3.813 0.001 Influencer+Path V.S. Random+Path 32.20 20.06 4.487 7.176 0.000

4.3.4 Sharing Behavior Evaluation

This section further evaluates the sharing behaviors in different advertisement diffusion processes. As mentioned before, egoism and altruism are two of the significant factors of willing-to-share behavior. There are four delivery situations discussed, as shown in Table 4.7.

(1) Indicating that the forwarder expects to obtain positive recognition from receivers.

It is the most beneficial to both the business opportunities seeking strategy and the brand expression establishing strategy.

(2) Indicating that the forwarder expects to influence the impression of receivers on a specific product/brand. It may be helpful to the brand expression establishing strategy.

(3) Indicating that the forwarder expects to inform the receivers of some promotion information of products. It is most beneficial to the business opportunities seeking strategy.

(4) Although this could also indicate that the forwarders expect to obtain negative

(4) Although this could also indicate that the forwarders expect to obtain negative

在文檔中 社群商務服務計算機制之設計 (頁 62-0)