Context-Aware Voice-Based Mobile Community Soe-Tsyr Yuan and Kai-Hsiang Peng Department of Information Management Fu-Jen University, F2, No13, Ln77, Sec2, Roosevelt Rd., Taipei, Taiwan [email protected]

Abstract Mobile commerce has increasingly been recognized as one of the most significantly prosperous areas for deploying information technology. This paper aims to advance the value of the information by providing a novel voice-information sharing mechanism. The proposed mechanism is a combination of a location-based information service known as a Killer App and a virtual community that consequently becomes a WCP (Wireless Content Provider). This community embodies context aware intelligence by analyzing the context-sensitive behavior of the community members and; furthermore, enabling proactive and precise context sensitive voiceinformation sharing. This voice-information sharing mechanism is comprised of an IVR system, a location service, EPN (Euclidean distance with Positive and Negative Strength) Clustering, Naïve Bayesian Prediction, and a set of metrics for monitoring the progress of the community. The primitive results show that our mechanism satisfactorily reaches the goal of proactive, precise sharing of voice information between community members.

Keywords: Mobile Community, Context-Aware, Naïve Bayesian, Collaborative Filtering, Mobile Commerce

1

1 Introduction Wireless technology has been driving the advent of new applications and services [1]. For instance, wireless carriers continually strive to provide a variety of information to their subscribed users in this m-commerce era, and context-aware applications [2] have therefore increasingly become a focus of research too. Contextaware applications have moved beyond simple mobile services by overcoming the inherent input/output limitations of mobile devices, through higher degrees of automation, and thereby understanding the context within which their users operate, such as their locations, the mobile devices they use, the activities they are engaged in, etc. In this paper, we present a novel context aware application that combines voice technology with an adaptive mobile location service [3] in order to provide a highly interactive virtual voice-based mobile community (that will subsequently be referred to as CAVBMC (Context-Aware Voice-Based Mobile Community)). CAVBMC exhibits context aware intelligence that makes the functions of the community highly personalized in addition to being context-sensitive. In other words, CAVBMC records and analyzes the users’ locations, past interactions, and other users’ interactions at the same locations in order to provide valuable information services that are contextsensitive and personalized. Before manifesting the intended contributions of CAVBMC, we shall describe the current status of the relevant issues: (1) ICPs’

hardship in the contents generation:

Most ICPs (Internet Content

Providers) exert themselves to offer free contents about the likes of music, food, travel, news, etc. When designating WCPs (Wireless Content Providers) from the traditional ICPs, wireless carriers have to choose ICPs that have the capabilities to provide sufficient contents based on their past accomplishments. However, it is tough to come up with a formula that fairly apportions the profits and costs incurred in the process amongst the carriers and ICPs. (2) Limitation of voice recognition: Although the technology of voice recognition is continually being researched, it still suffers from the problem of background noise. 2

This so-called background noise is the noise occurring within the operator’s voice input environment. It is unusual for mobile users to actually seek out a noise-free environment in which to operate their voice input in order to eliminate background noise, and therefore the possibility for successful voice recognition from voice input of mobile users is marginal. (3) Triumph of Short Messages: According to a statistics report of the GSM Association, the volume of SMS (Short Message Service) transmitting through GSM wireless network amounted to 3 billion per month up to December 2001 [4]. In other words, SMS has been the prevailing approach for proactive delivery of messages. (4) Success of NTTDoCoMo’s i-Mode in Japan: According to statistics from NTTDoCoMo, the number of i-Mode subscribed users has reached 31 million up to March 2002 [5]. These users download ring tones, games and e-mails with their mobile phones. i-Mode has successfully evolved from a traditional voice-service provider to a leading wireless information-service provider. The success of i-Mode inspires existing wireless carriers to carry out information services in this wireless communication era.

In view of the indicated situations shown above, it naturally leads to a question: What approach should wireless carriers employ in order that proactive personalized information/contents services are provided inexpensively and attract their subscribed mobile users without placing a burden on their users? In this paper, one possible solution to the above question is provided. This solution is referred to as CAVBMC. CAVBMC aims to advance the value of the information by providing a novel voice-information sharing mechanism. The proposed mechanism is a combination of a location-based information service known as a Killer App, and a virtual community that consequently becomes a WCP (Wireless Content Provider). This community captures users’ preferences by analyzing the context-sensitive behavior of the community members. With these preferences, voice-information sharing intends to become proactive (with the help of SMS) and precise (i.e., a good quality analysis and prediction); furthermore, relevant product promotion can be deployed accordingly in

3

order to reach the objective of an efficient market. The ideas behind CAVBMC are six-fold:


Communities can be considered one source of the contents generation [6]. That is, the role of the ICP can be substituted through prudent exercising of communities. Mobile users of a community contribute context-sensitive contents to the community. As a result, information and contents are constructed inexpensively for wireless carriers based on the valuable user experiences of the subscribed users.




In order to contribute contents to a community, mobile users find voice input the most convenient and natural method as opposed to a keyboard entry which places a burden on the users. Due to the limitations of the technology of voice recognition, we employ an IVR (Interactive Voice Response) system together with stored data of the users’ messages recorded in a voice format. (Voice recognition can be used when the problem of background noise is resolved in the future).




SMS is employed for proactive message delivery.




Shared voice information is notable owing to its location sensitivity and its having been contributed to by different users at the same location. In other words, information, location, and people, are associated in order to increase the value of information as opposed to the shared information captured in traditional Internetbased communities [7][8] whose contents are context-insensitive and restricted to text and image.




Deploy suitable analysis and prediction algorithms in order to furnish mobile users with a precise information-sharing experience. These algorithms are EPN (Euclidean distance with Positive and Negative Strength) Clustering and Naïve Bayesian Prediction and will be described in Section 3. Furthermore, devise a set of community metrics for monitoring the progression of the community in order to maintain the attraction of mobile users to the community with appropriate community promotion exertion such as free coupons, etc.




The dynamic preferences of a user are captured in a user profile that records the community forums in which the user participates and the corresponding levels of

4

participation in terms of participation frequency in the forums. Subsequently, these preferences of users can be utilized for location-based wireless targetmarketing if necessary. We anticipate that CAVBMC could become a new profitable business model for wireless carriers.

The remainder of the paper is organized into five sections. In Section 2, we discuss the characteristics and architecture of CAVBMC. Section 3 presents the analysis and prediction methods used in CAVBMC. In Section 4, we describe the metrics we have devised for monitoring community progression. Section 5 then evaluates CAVBMC in terms of the quality of its analysis and prediction. Finally, a discussion and a conclusion are made in Section 6 and Section 7, respectively.

2 CAVBMC In this section, we begin by describing the functions of CAVBMC. These overall functions provide a picture about why CAVBMC is one possible solution for wireless carriers to inexpensively provide proactive personalized information/contents services that attract their subscribed mobile users without placing a burden on them. The architecture of CAVBMC is subsequently described.

2.1 The Functions of CAVBMC CAVBMC binds information, locations, and people together in order to advance the value of information by providing a novel voice-information sharing platform that utilizes the valuable users’ experiences. That is, the voice information is locationsensitive and is contributed to by a collection of users whose experiences are locationspecific. CAVBMC has two modes of function: the take mode and the offer mode. The take mode enables a user to acquire from CAVBMC, voice information that is personalized and location-sensitive, while the offer mode allows a user to quickly render to CAVBMC his/her valuable experience that is expressed in voice format and is location-tagged. 5

(a)

(b)

(c) Figure 1. A Scenario of CAVBMC’s functions: the offer mode function and the take mode function

6

For instance, in the scenario shown in Figure 1, if Rick invokes the offer mode function of CAVBMC at 5:30 p.m. in order to report a nice Italian restaurant in SOGO Mall, then, after having received Rick’s message together with the location information, CAVBMC predicts the possible forums to which Rick’s sharing might belong (based on the analysis of his historical interaction records) in order to speed up the sharing process. On the other hand, if, at 8:40 p.m. in SOGO Mall, Jerry invokes the take mode function of CAVBMC for the purpose of obtaining good restaurant recommendations, CAVBMC would then furnish Jerry with some location sensitive information (i.e., good restaurant recommendations in SOGO Mall) that is also based on Jerry’s personalized preferences (learned from the historical interaction records captured in the user’s profile) and the experiences of users similar to Jerry. It is likely that Rick’s experience could be shared with Jerry if Rick is considered a similar user to Jerry and Rick’s message is worthwhile. Due to the physical limitations of mobile phones, the method of Top-N display is exerted (the user chooses and listens to the Top-N messages recommended by CAVBMC). For simplicity, Top-3 display is used in this paper, whereby the user can first listen to the Top-3 messages in part and then decide whether to listen to the three complete messages or to choose other topics of interest.

2.2. The Architecture of CAVBMC Figure 2 shows the architecture of CAVBMC, which is made up of several components, each of which is responsible for particular tasks facilitating the process of voice-information sharing. Figure 2 also exhibits the flow of the process in both the offer mode and the take mode of information sharing. CAVBMC’s voice-information sharing platform is accomplished chiefly by those components, together with an IR system, a Location Service system, and an SMS gateway. IVR is a telephony technology by which someone uses a touch-tone telephone to interact with a database and acquires information from or enters data into that database. IVR technology does not require human interaction over the telephone as the user's interaction with the database is predetermined by what the IVR system will 7

allow the user access to. For example, banks and credit card companies use IVR systems so that their customers can receive up-to-date account information instantly and easily without ever having spoken to anyone. IVR, together with the voice information database, enables information sharing by simple voice input. Location Service is a location gateway that obtains a user's location and presence data from multiple sources (including Cell-ID, A-GPS, AOA, E-OTD and others) and delivers that data to location-based applications. Location Service enables information sharing to be tagged by location. SMS is a service for sending messages of up to 160 characters (224 characters if one is using a 5-bit mode) to mobile phones that use GSM (Global System for Mobile) communication. SMS messages are transmitted within the same cell or to anyone with roaming service capabilities. They can also be sent to digital phones from a website equipped with PC Link or from one digital phone to another. An SMS gateway is a website that lets you enter an SMS message to someone within the cell served by that gateway or that acts as an international gateway for users with roaming capabilities. The SMS gateway renders the CAVBMC’s functions proactive. The following are the descriptions for the components of CAVBMC:


Community Main Function: Community Main Function controls the operation of the community such as activating the execution of its internal modules based on users’ demands.




Forum Module: For the flexibility of community-forum maintenance, we modulate the community forums in order that each forum of topics can be inserted/deleted easily within the community.




Rating Module: For the purpose of maintaining the quality of shared voice information, a rating module is used to regard the shared messages. A user can rate a piece of a message after experiencing what the message has said. To prevent them from forgetting, the user can request a reminding prompt service after the lapse of a previously selected amount of time. SMS is then employed to proactively remind the user of this rating task. For instance, the topic of restaurants is exercised in our evaluation: a user can’t logically rate his/her dining experience at a restaurant suggested by a message that this user listens to

8

unless he/she has already started dining. It is believed that this prompting service will increase the user’s intention to participate in the rating.


EPN Module: In order to provide a precise experience of information sharing (in both the take mode and the offer mode), the EPN module supplies recommendations

to

users

by

predicting

the

users’

preference

of

messages/forums to take/offer by analyzing users’ past behaviors such as their message ratings between forums recorded in their user profile. The EPN Module consists of two methods: the EPN (Euclidean distance with Positive and Negative Strength) Clustering method (Section 3.1) and the Naïve Bayesian Prediction method [9][10] (Section 3.2).

Offer Mode

3.

Voice database

2.

Location Service

10.

4. 8.

Forum Agent User Profile

Community Agent

5. 6.

12.

9.

EPN Agent

11.

1.

7. Rating Agent

Mobile Phone System 14.

13. SMS Gateway

IVR system

Figure 2. The architecture of CAVBMC and the process flow of the offer mode function

What follows is the flow of the process of the offer mode of information sharing from Step 1 to Step 11 (the process flow for the take mode is similar to the process flow of the offer mode except that it needs three additional steps: Step 12 to Step 14): (1) A user activates CAVBMC. (2) The mobile phone system dispatches the user’s phone number to Location Service. (3) Location Service attains the data of the user’s location with the user’s mobile

9

phone number. (4) Location Service sends the user’s location data to the mobile phone system. (5) CAVBMC redirects the user to the IVR (Interactive Voice Response). (6) Community Main Function receives the request from the user and activates the execution of EPN Module. (7) EPN Module predicts the possible forums that this sharing intends, using the information stored in the users’ profiles of the cluster of users to which this user belongs. Afterwards, a Top-3 list of forum recommendations is presented to the user. (8) If the user rejects these forum recommendations, then he/she may jump to the toplevel forums and select the ones with which this user would like to share. (9) Confirm the forum choice of the user and return to the main function. (10)

Record into the voice database the user’s voice message together with the

location data obtained from Location Service. (11)

Add this interaction record into the user’s profile.

(12)

When the user activates the reminding prompt service of SMS for the coming

rating task for the purpose of evaluating a message the user has just taken, the control is then redirected to Rating Module. (13)

Rating Module dispatches the request of the prompt service to SMS Gateway

when a specified amount of time has elapsed. (14)

SMS Gateway transmits this reminding prompt to the user through the Mobile

Telephone System. (15)

Mobile Telephone System passes the prompt to the user’s mobile device.

(16)

The user activates the rating function.

(17)

Enter the IVR system.

(18)

IVR acquires the rating value provided by the user, who selects a rating value

from the mobile device’s keypad, and then stores the rating value back to the voice database.

In summary, the architecture shown in Figure 2 manifests the ideas behind the context-aware application of CAVBMC:

10




Shared voice information is tagged by location with the function of Location Service.




Information sharing is exerted in a highly personalized manner by means of the EPN Module, User Profile, and Rating Module.




People are linked together (and in a sense interact with each other in terms of rating, taking, and offering messages) by the shared information that is contributed to by a collection of people whose experience takes place at indicated locations.




Mobile communication devices unfold the breakthrough of context-aware applications.

3. EPN Module Since the core of the CAVBMC rests on the EPN module (composed of two methods: the EPN Clustering method and the Naïve Bayesian Prediction method), this section provides the descriptions of the methods in Section 3.1 and Section 3.2, respectively. Before we detail and rationalize the two methods, we will first explain the dilemma associated with the existing relevant approaches. CF

(Collaborative

Filtering)

[11][12][13]

is

a

technique

of

providing

recommendations based on statistical matches of people's tastes. A CF system maintains a database of users and keeps records of items they have rated. These vectors of ratings are compared to the vectors provided by other users and subsequently people with similar tastes can be discovered and the prediction of a user’s behavior can be made based on the information of the cluster to which the user belongs. Existing CF methods presume that the rating data will be plentiful so that clusters of users with similar tastes can be computed with ease. However, this assumption does not always hold, especially for the rating data in CAVBMC (as there would be quite a number of community forums, each of which would then consist of a great number of messages, and thus the possibility of a large number of messages unrated is high). Table.1 demonstrates a simple case of such a situation, that is, with users U1, U2, …, etc., their message ratings across forums identified by capital letters such as 11

A, B, etc. (representing forums) followed by numbers (representing messages within a forum). For instance, A13211 represents the 3211th message in the A1 forum. (This way of representing the messages enables the analysis described later ranging across messages and forums.) The Liker-Scale 1-5 levels of rating are implemented for the users to rate the messages, that is, a user enters a digit ranging from 1 to 5 by means of the keypad on the mobile phone to record the quality of a message based on his/her experience. The symbol “-“ in Table.1 represents that the user has not rated the message yet. In the instance of large, sparse data [14][15], there will be bias about the clusters of people with so-called similar tastes if the clusters formed are based merely on those sparse rating data. To overcome the problem of sparse rating data within the CF framework, this paper devises a novel clustering method (the EPN Clustering method) together with a Naïve Bayesian Prediction method so that the bias mentioned above can be reduced. The Naïve Bayesian Prediction method differs from existing Naïve Bayesian methods in two ways: (1) Existing Naïve Bayesian methods compute the probability of what a user will do in a certain situation based on his/her past behaviors. (2) However, our Naïve Bayesian method estimates this probability based on the behaviors of the whole clustered group in order to overcome the problem of the insufficient rating data of the single user. Since users’ preferences (in terms of their ratings) would change as time goes by, the EPN module periodically clusters the data in order to ensure clusters of people bearing similar tastes at all times.

Table1. Raw Rating Table A13211 A11123 B12321 A34121 B22112 B14001 D21153 F13121

K12112 K21212

U1 U2

-

3

2

3

-

2

4

3

-

2

1

3

-

-

2

3

3

2

1

3

U3

-

-

5

3

-

3

3

5

-

1

U4

1

-

-

-

-

2

-

-

1

3

U5

3

2

4

2

4

-

4

5

3

2

U6

-

-

4

4

4

4

5

5

3

12

3.1 The EPN Clustering Method To surmount the problem of sparse rating data, the EPN (Euclidean distance with Positive and Negative Strength) Clustering method combines Euclidean distance with positive/negative strength to compute the similarities between users. Positive strength between a user Uq to another user Up (who is under prediction) represents the ratio of the messages that are rated by Uq and with the same rating as Up’s in light of the ratio of the rated messages to the total messages. While negative strength denotes the ratio of messages that are rated by Uq but with different ratings from Up’s in light of the ratio of the unrated messages to the total messages. In addition to the traditional Euclidean-distance similarity measurement, both positive strength and negative strength reveal important information about the degree of similarity between users. EPN Clustering Algorithm 1. 2. 3.

The user under the prediction is ( U p ). Compare the rating records of the other users ( U q ) with that of the user under the prediction ( U p ). Calculate the positive strength (P) and the negative strength (N). (U q , U p ) : ( P, N ) = (

S R R−S A− R × ， × ) A R A R

Up：

the user under prediction U q ：other users(q=1~n) S ：number of the Uq’s rated messages with the same rating values as ( U p ) R ：number of messages already rated by user ( U q ) A ：total number of the messages 4. Calculate the Euclidean Distance (that only considers the ratings of the messages rated by both users): E=

∑ (r

pj

− rqj ) 2

r ：rating j

5.

：message index

r pj ：Rating

value of

Up

toward the jth message.

rqj ：Rating

value of

Uq

toward the jth message.

Calculate the value of the similarity measurement between Similarity = w1*P + w2*N + w3* (1/E)

Uq

and

Up.

w1, w2, w3: Relative weight

6.

13 Set the users with top 5 values of similarity as the cluster of similar users to

Up.

The justification behind the necessity of these measurements will be revealed by means of a simple example shown in Table 2 and a demonstration of the EPN Clustering method (the algorithm of which is shown above) based on the example shown in Table 1: Table 2. Comparisons with Similarity A13211 A11123 B12321 A34121 B22112 B14001 D21153 F13121 K12112 K21212 U3

-

-

5

3

-

3

3

5

-

1

U4

1

-

-

-

-

2

-

-

1

3

U5

3

2

4

2

4

-

4

5

3

2




Since U4 has more unrated data than U5, who almost completes the ratings, we consider U4 more similar to U3 than U5 due to U4’s greater chance of being similar to U3. (Of course, the amount of U4’s unrated data will change over time, and thus periodical clustering of the unrated data is necessary in order to reassess the degree of similarity.)




In the EPN Clustering method, the similarity measurement of two users is a weighted sum of their positive strength, negative strength, and Euclidean distance. The weights (w1, w2, w3, respectively for positive strength, negative strength, and Euclidean distance) are determined based on the ratio of the rating data. For deficient rating data, the weights of positive strength and negative strength should be larger than that of the Euclidean-distance’s so as to avoid the bias lead by a big Euclidean-distance due to the case of a few extreme ratings. On the contrary, when the rated data is sufficient, the weight of Euclidean distance should be larger than those data of positive strength and negative strength because then it will not be as easily affected by a case of extreme ratings.




Considering the users shown in Table 1 and for the purpose of finding users that are similar to U3 (who is under prediction), the method would compute the similarity measurement between any other user and U3. Take an example of User1 and User3 in Table 1: the positive strength is 1/7 * 7/10 because User1 has rated seven messages out of the total ten messages and only one of the seven messages has the same rating as User3’s. The negative strength is 6/7 * 3/10 14

because User1 has three unrated messages out of the total ten messages but there are six of U1’s messages (out of the seven rated messages) whose ratings are different from User3’s. The intuition behind the computation of positive strength and negative strength is two-fold: (1) The positive correlation from User1 to User3 (1/7) has to be weighted by the percentage of the messages rated by User1 (7/10) in order to tune this positive correlation to a correct scale. (2) The negative strength considers the ratio of the messages unrated by User1 (3/10) as the greater number of User1’s unrated messages and hence the greater probability of User1 looking similar to User3; however, this ratio has to be weighted by the percentage of the differentiation from User1 to User3 (6/7). The positive strength and the negative strength from U1 to U3 are then represented as follows: 1 7 6 3 (U 1 ,U 3 ： ) ( P, N ) = ( × ， × ) 7 10 7 10

Euclidean distance between User1 and User3 is as follows:

(5 − 2) 2 + (3 − 3) 2 + (3 − 2) 2 + (3 − 4) 2 + (5 − 3) 2 + (1 − 2) 2 = 16 Likewise, the similarity measurement between any other user and User3 can be computed as follows: E (U 1 , U 3 ) =

16

(U 2 , U 3 ) : ( P , N ) =

E (U 2 , U 3 ) =

13

E (U 4 , U 3 ) =

5

E (U 5 , U 3 ) =

4

E (U 6 , U 3 ) =

7

(U 4 , U 3 ) : ( P , N ) = (U 5 , U 3 ) : ( P , N ) = (U 6 , U 3 ) : ( P , N ) =




3 1 7 6 × ， × ) 7 10 7 10 2 8 6 2 ( × ， × ) 8 10 8 10 6 0 4 4 ) ( × ， × 4 10 4 10 1 1 9 8 ) ( × ， × 9 10 9 10 1 7 6 3 ( × ， × ) 7 10 7 10

(U 1, U 3 ) : ( P , N ) = (

From the above measurements, we find that U2 is most similar to U3. In a comparison between U1 and U6, U6 is more similar to U3 than U1 because the Euclidean distance of U6 and U3 is smaller although the positive strength and the negative strength of U1 to U3 and U6 to U3 is the same. This method can find K most similar users to U3 and form a cluster of people with tastes similar to those of U3. When K equals 3, U2, U4, and U6 are the cluster of people with

15

tastes similar to those of U3.

3.2 The Naïve Bayesian Prediction Method The Naïve Bayesian Prediction method aims to predict a message’s rating value for a user by estimating the conditional probabilities of the possible rating values and selecting the rating value with the largest conditional probability. Existing Naïve Bayesian methods compute a conditional probability based on the belief that a user will do a certain thing in a certain situation based on his/her past behaviors. However, with only a meager amount of the user’s rating data (i.e., a limited amount of the user’s past behaviors) an alternative approach of Naïve Bayesian is employed. Our Naïve Bayesian method estimates the conditional probability from the past behaviors of the cluster of people of similar taste instead of merely the past behaviors of the single user. (The cluster of people of similar taste to which a user is added, is acquired by the EPN Clustering method as mentioned in Section 3.1. For instance, with the example shown in Section 3.1, a cluster of people of taste similar to that of U3 is shown in Table 3.) That is, the past behaviors of people of similar taste are used to make up for the lack of information in the past behaviors of the single user. After the rating values of the messages are predicted, the n messages with the top n largest rating values are taken as the Top-N recommendations for the user. What follows is a demonstration of the Naïve Bayesian Prediction method (the algorithm for which is shown below) in the example shown in Table 3:


Since the five-scale method is used to rate messages, we have to individually compute the probability of each possible rating (1 to 5) if we want to predict the behavior of User3 to the k12112 message. For instance, the probability of the rating 1 is computed as follows: The probability of a rating with the value i (i = 1 to 5) from the clustered rating table: P( v j ) The probability of a rating with the value i (i = 1 to 5) occurring in the k12112 message: P( a j | v j )

v NB = arg max P(v j ) ∏ P(a i | v j ) 16

From Table 3, the probability of the k12112 message that is rated 1 is 5/25 * 2/5 because there are five 1 ratings out of twenty-five ratings and two 1 ratings occurring in the k12112 message out of the five 1 ratings. The probability that User3 rates 1 for the k12112 message is:

5 2 2 × = 25 5 25

Similarly, we can compute the other probabilities individually: 3 − The probability of the rating of 2: × δ 25 3 We use the δ to substitute for 9 − the value of 『-』in the The probability of the rating of 3: δ × 25




9

computation and it represents a

δ

The probability of the rating of 4:

4 − × 25 4

The probability of the rating of 5:

4 1 1 × = 25 4 25

very small value。

Accordingly, the most possible rating of User3 for the k12112 message is 1 that has the biggest conditional probability compared with the other rating values.




Similarly, the method predicts User3’s behaviors toward the other messages. Among the messages with the predictive ratings, select the n messages with the top n largest rating values as the Top-N recommendations for the user.

Table 3. Rating Table clustered with EPN A13211 A11123 B12321 A34121 B22112

B14001 D21153 F13121

K12112 K21212

U2

1

3

-

-

2

3

3

2

1

3

U3

-

-

5

3

-

3

3

5

-

1

U4

1

-

-

-

-

2

-

-

1

3

U6

-

-

4

4

4

4

5

5

3

17

Naïve Bayesian Prediction Algorithm 1. Given a user under prediction and the cluster of his/her similar users and their messages, select the messages that are associated with the location under inquiry, resulting in a rating cluster table. 2. Calculate the total number of ratings available in the rating cluster table that is obtained after the execution of the EPN Clustering method. T : The number of ratings in the clustered rating table. 3. Calculate the total number of ratings with their values being i . Ri :

The number of ratings with values being i ( i =1~5)

4. Calculate the number of the m th message’s ratings with their values being i . th Rm : The number of the m message’s ratings with their values being i 5. Compute the probability of the m th message with the rating value being i . Ri Rm × Pi = T Ri Pi ：The probability of the message’s rating value being i ， i =1~5 6. Compare all of the computed probabilities and make the prediction that will be the rating value of the biggest probability.

4. Community Metrics In this section, we describe the metrics we have devised in order to monitor community progression so as to maintain the attraction of the community to mobile users by exerting appropriate community promotion. These exertions from wireless carriers can be the likes of a certain amount of free airtime, service upgrading, coupons of their business partners, etc. As addressed in Section 1, for wireless carriers, communities are good sources of the content generation and CAVBMC is expected to become a new, profitable business model. Accordingly, certain incentives to the users provided by wireless carriers are worthwhile. Existing non-profit communities are mainly sustained by certain key users who possess enthusiasm as well as highly professional knowledge in the relevant topics [16] and who intensively share valuable information in certain forums of the community. However, sole dependence on key players might jeopardize the community when they lose their motives. Appropriate incentive strategies would bring about the quality and the popularity of a community. 18

The following is a list of the metrics and their descriptions:


Quality of Forum (QF): This QF metric aims to adequately reflect the quality in the shared information of a forum in the community. Existing community forum metrics mainly emphasize the quantity of pieces of shared information. However, the size of the shared information does not actually represent the quality of a community forum. For example, the quality of a community forum is bad if the community forum is full of advertisements instead of valuable information. The metric QF we devise (QF takes into account both the size and the merit of the shared information), is defined as follows: QF =

R G−B ( ) M R

Number of all the messages: M Number of rated messages: R Number of messages of good quality: G (messages with rating values being 3-5 are considered good messages) Number of messages of bad quality: B (messages with rating values being 1-2 are considered bad messages)


Popularity of Forum (PF): This PF metric intends to quantify the popularity of a community forum. Existing community forums consider the quantity of responses to certain messages as the index of community forum popularity. However, this only reveals information about the popularity of the designated messages (as opposed to the popularity of the community forums themselves). The metric PF we devise (PF regards both good discussion quality and high number of ratings as the important clue for community popularity), is defined as follows: PF = QF ×

v V

Quality of Community Forum: QF Number of ratings in the forum:

19

v

Number of all the ratings in the community:


V

Personal Reputation (PR): The PR metric aims to identify the key players in a community and to provide an understanding of community users’ contributions in terms of the size and the quality of their shared information. Accordingly, incentives offered by wireless carriers can be apportioned to the users appropriately [17]. The PC value of a user is defined as follows: PR =

n 2 ∑ Ri × N n

Number of the user’s shared messages: n Number of all messages: N Average value of the ratings for a certain message shared by the user: Ri

5. Evaluations In this section, we provide the evaluation results on how the precise sharing of voice information in CAVBMC can be reached (Section 5.1). We also show how helpful the community metrics devised in this paper are in monitoring the community’s progression (Section 5.2). In evaluating the precision of the information (i.e., the quality of the EPN Clustering method combined with the Naïve Bayesian Prediction method), the measurement of predictive accuracy is employed; while in evaluating the usefulness of the community metrics (i.e., the abilities of the metrics), the approach of look-and-feel is applied.

5.1 Precision of the EPN Clustering Combined with the Naïve Bayesian Prediction In this section, we first define the measurement of predictive accuracy and then present subsequent evaluations on predictive validity, which is exerted in two ways: (1) the progression of the measurements along with the size of ratings made by a user type (Section 5.1.1) (2) the ability of the measurements to discriminate between two 20

extreme types of users (Section 5.1.2). The measurement of predictive accuracy for a user is defined as the ratio of the number of the accurate predictive ratings to the total number of predictive ratings, from all of the ratings predicted by our method for the user. By accurate predictive ratings, we mean the predicted rating values that fall into the behavior pattern of the user. In this paper, we employ a model of user types [18][19][20] that classifies users into six types: A-type (Introversion), B-type (Neuroticism or Emotional Stability), Ctype (Agreeableness), D-type (Conscientiousness), E-type (Intellect or Openness to Experience), F-type (Negative). Based on the characteristics (and thus the behavior patterns) of these types, Table 4 then shows the resulting behavior patterns of the six different types of users [20][21]. Each type of user posits certain behavior patterns such as the range of the possible rating values and the frequency of the ratings among all the different types of users.

Table 4. Six personalities Personality

Possible Rating Values

Rating Frequency

A-Type (Introversion)

1,2,3

5%

B-Type (Neuroticism or Emotional Stability) C –Type (Agreeableness) D-Type (Conscientiousness)

1,2

5%

4,5 3,4,5

30% 30%

E –Type (Intellect or Openness to experience) F-Type (Negative)

3,4,5

60%

1,2

60%

5.1.1 The progression of the Predictive Accuracy Measurements This set of experiments aims to understand if the size of a user’s ratings influences the predictive accuracy of the user. For simplicity, the prior condition of the community is presumed as follows: there are 100 messages across the community forums and 100 users (that are composed of 10% A-type, 10% B-type, 30% C-type, 30% D-type, 10% E type, and 10% F-type, 21

based on the general normal distribution of the types of users within society). Accordingly, there is a collection of ratings generated and distributed following the user behavior patterns shown in Table 4. This set of experiments is then conducted as follows: (1) For each user type, we construct five independent experiments. (2) In each of such experiments (i.e., for a new user of the given type), a designated number (such as 1, 5, 10, 20, and 30) of ratings of the user are randomly generated (based on the user pattern) as the given ratings by the user. (3) Our prediction method (the EPN Clustering method (with the cluster size of 5) combined with the Naïve Bayesian Prediction method) predicts the remaining certain number of ratings (based on the rating frequency allowed by the user type). These are subsequently matched against the range of the rating values allowed by the user type in order to calculate the measurement of predictive accuracy. (4) When the five experiments are completed, an average of the five predictive accuracy measurements can be acquired for the user type. The experiment results are shown in Figure 3 and prove that the size of ratings given by a user does influence predictive accuracy: (1) The average predictive accuracy, in general, reaches a satisfactory state (70%-80% of predictive accuracy) after 10 ratings given by a user. (2) The more ratings a user provides, the higher predictive accuracy our method achieves. For instance, after giving 30 ratings, a user can obtain CAVBMC’s recommendations with 90% of predictive accuracy. (3) Conversely, there is low predictive accuracy when the number of ratings is less than 10, an initial inquired users’ interests would help improve the predictive accuracy of the recommendations at the early phase. In order to investigate if the distribution of user types influences the predictive accuracy results, we conduct the same experiments except we do not presume a normal distribution of user types. Instead, random distributions of user types are employed as long as the resulting size of each user’s type is greater than 5 as the process of EPN clustering generates a cluster of size 5 for a user under prediction. Figure 4 then shows that our method with random distributions of user types still exhibits a similarly good quality of predictive accuracy compared to the method with a normal distribution of user types.

22

Averaged Predictive Accuracy

100.00% Type_A

80.00%

Type_B

60.00%

Type_C

40.00%

Type_D Type_E

20.00%

Type_F

0.00% 1 rating

5 ratings

10 ratings 20 ratings 30 ratings

Number of Ratings

Figure 3. The progression of the predictive accuracy measurements for normal distribution of user types

100.00% 90.00% Type_A Type_B Type_C Type_D Type_E Type_F

Averaged Predictive Accuracy

80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% 1 rating

5 ratings 10 ratings 20 ratings 30 ratings Number of Ratings

Figure 4. The progression of the predictive accuracy measurements for a random distribution of user types

At this point we would like to investigate if a cluster of size 5 is indeed appropriate in our method. Therefore, alternative sets of experiments are conducted in order to

23

investigate if cluster size influences predictive accuracy. In this set of experiments, the same experiment setting as indicated in the first set of experiments shown in Section 5.1.1 is employed except that we make the most popular user type (C-Type) as the target of users under investigation. Seven sizes (1, 2, 3, 4, 5, 6, 10) of clusters are experimented on in a hope of justifying that clusters of size 5 are appropriate in our method.

100.00% 90.00% 80.00% 70.00% Averaged Predictive Accuracy

60.00%

size=1

50.00%

size=2

40.00%

size=3

30.00%

size=4

20.00%

size=5

10.00%

size=6

0.00%

size=10 1 rating

5 ratings 10 ratings 20 ratings 30 ratings Number of Ratings

Figure 5. The relationship of cluster size and predictive accuracy

Figure 5 shows the results that indicate the following observations: (1) Extremely small sizes (1, 2) of clusters exhibit fluctuating predictive accuracy growth. (2) Size 3 of clusters reveal a steady predictive accuracy growth. (3) Larger sizes (5, 6, 10) of clusters unfold better predictive accuracy growth. (4) Larger sizes (5, 6, 10) of clusters exhibit similar predictive accuracy growth. (5) Between cluster sizes 5, 6 and 10, the choice of size 5 is accordingly made for gaining more efficiency in Naïve Bayesian Prediction.

24

Averaged Predictive Accuracy

100.00% size=3

80.00%

size=4

60.00%

size=5

40.00%

size=6

20.00%

size=10

0.00% 1 rating 5 ratings

10 ratings

20 ratings

30 ratings

Number of Ratings

Figure 6. The relationship between cluster size and predictive accuracy for a larger experiment scale

However, in order to allay fears that size 5 of clusters making good predictive accuracy growth only applies to a particular experiment scale (100 messages and 100 users), an alternative set of experiments of enlarged scale (500 messages and 500 users) is conducted. Figure 6 then shows the results. From Figure 6, we have the following observations: (1) Larger sizes (5, 6, 10) of clusters in a large scale experiment still exhibit better predictive accuracy growth. (2) Larger sizes (5, 6, 10) of clusters still exhibits similar predictive accuracy growth. (3) As a result, the choice of cluster size 5 is believed to be appropriate in our method regardless of the scale of the experiment.

5.1.2 The Discrimination Ability of the Predictive Accuracy Measurement This set of experiments aims to understand if the size of a user’s ratings influences the measurement’s ability to discriminate between extreme user types. In this set of experiments, the extreme user types are E-type and F-type. For easily discerning the prediction discrimination between extreme user types, we recast the predictive accuracy measurement to an averaged prediction value. As in the experiment setting described in Section 5.1.1, at the initial state of the community, there are 100 messages across the community forums and 100 users (that 25

are composed of 10% A-type, 10% B-type, 30% C-type, 30% D-type, 10% E type, and 10% F-type based on the general distribution of the types of users within society). Accordingly, there is a collection of ratings generated and distributed following the user behavior patterns shown in Table 4. This set of experiments is then conducted as follows: (1) For each of these two user types, we construct ten experiments. (2) In each of these experiments (i.e., for a new user of the given type), a designated number (such as 1, 5, 10, 20, and 30) of ratings of the user is randomly generated (based on the user pattern) as the given ratings of the user. (3) Our prediction method makes a single prediction (for a randomly chosen unrated message) right after the user gives the ratings of size the designated number. (4) When the ten experiments are completed, an average of the predicted rating value of the ten experiments can be computed for the user type. Figure 7 then shows the experiment results and reveals that the discrimination

Rating value

ability of the measurement generally also reaches a satisfactory state after 10 ratings.

4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

type_e type_f

1 rating 5 ratings

10 ratings

20 ratings

30 ratings

Number of ratings

Figure 7. The Discrepancy of the Measurements between Two Extreme Types of Users

26

5.2 Usefulness of the Community Metrics This section aims to show the usefulness of the community metrics (i.e., the abilities of the metrics) with the approach of look-and-feel, that is, to justify that these metrics endow the community manager the ability to decide when and to whom the community promoting strategies should be exerted. By the look-and-feel approach, we mean a collection of graphs of metric statistics (look) obviously manifest the answers (feel) to the above questions of when and to whom. In this experiment, each of eight forums (identified by A1, A3, B1, B2, D2, F1, K1, and K2) of the community is presumed to have certain number of messages (M), we randomly designate messages that are to be rated (and thus R messages to be rated), and rating values are randomly assigned to those designated messages (and thus G good messages and B bad messages). Figure 8(a) then shows a look-and-feel graph of the QF statistics. It is very obvious for the community manager to understand the quality of all the community forums and to decide if this moment is the right time to exert promoting strategies on certain forums. Similarly, Figure 8(b) shows a look-and-feel graph of the PF statistics. It is quite straightforward for the community manager to perceive the popularity of all the community forums and decide if this moment is the right time to exert promoting strategies on certain forums in order to boost their popularity. Figure 8(c) then allows the community manager to award the right people (i.e., the users with high PR) so as to maintain the quality and the popularity of the community forums.

27

0.2 A1 A3 B1 B2 D2 F1 K1 K2

0.15 0.1 0.05 0 -0.05

(a) 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 -0.01 -0.02

A1 A3 B1 B2 D2 F1 K1 K2

(b) 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 PR

U1 U2 U3 U4 U5 U6 U7 U8 U9 U10

(c) Figure 8. (a) Ability of QF (b) Ability of PF (c) Ability of PR

28

6. Discussion In this section, a few questions are discussed to investigate the feasibility of the proposed CAVBMC: (1) In the case of information content that is not consistent with location (i.e., a user activates the offer mode function at a location for sharing a message that is associated with a different location), the consequence is that other users would then rate this message poorly at the take mode function and thus this message would possibly not be selected as one of the Top-N recommendations. (2) Lengthy, verbose messages might adversely affect their values but this depends purely on how users in general would respond to messages of a verbose nature (rather than how CAVBMC would cope with it). (3) The period of time in which messages stay inside the community might influence the values of the messages but this also depends on how users in general respond to such messages. That is, if a message is outdated (for instance, a recommended restaurant that does not exist anymore), chances for the message to get exposed to Top-N recommendations decrease gradually (or even reach zero chance shortly). (4) Although the user behavior patterns (that are shown in Table 4 and used for our evaluations) only show one possible set of behavior patterns, we believe our method would work well for other sets of behavior patterns as long as there actually are behavior patterns. (5) To make any sort of IVR based audio browsing system even remotely comfortable and useful is a major challenge, and a thorough user study of the proposed user interface of CAVBMC has to be conducted in the future. (6) Besides the functions of the offer mode and the take mode, it is worthy of further study to see if other functions are also helpful to mobile users in sharing experiences within CAVBMC (for instance, the function of editing messages).

7. Conclusions In this paper, we raise an interesting question – what approach should be employed for wireless carriers in order that proactive personalized information/contents services 29

are provided inexpensively and attract their subscribed mobile users without placing a burden on the users? We also present one possible solution – CAVBMC. CAVBMC aims to advance the value of given information by providing a novel voiceinformation sharing mechanism that is a combination of a location-based information service and a virtual community that consequently becomes a WCP. This community captures users’ preferences by analyzing the context-sensitive behavior of the community members. With the preference of the members, voice-information sharing intends to become proactive (with the help of SMS) and precise (with a good quality EPN Clustering method and the Naïve Bayesian Prediction method that we have devised); furthermore, relevant product promotion can be subsequently deployed in order to reach the objective of an efficient market. Our future jobs include seeking out the opportunity of field-testing and evaluation and working on further user studies of the proposed user interface and possible refinements of our method.

30

References: [1]

ITU

(International

Telecommunication

Union)

(2001)

http://www.itu.int/ti/industryoverview/index.htm [2] Romm, CT, Pliskin, N. and Clarke, RJ (1997) ”Virtual Communities and Society: Toward an Integrative three Phase Model”, International Journal of Information Management Vol.17,No.4,P261-270, 1997. [3] Ericsson (2001) http://www.ericsson.com.tw/ericsson/technology/t0605.htm [4] GSM Association (2001) http://www.gsmworld.com/news/statistics/index.shtml [5] NTTDoCoMo (2002) http://www.nttdocomo.com/html/subscriber_growth.html [6] Tiwana, Amrit and Bush, Ashley (2000) ”Peer-to-Peer Valuation as Mechanism for Reinforcing Active Learning in Virtual Communities: Actualizing Social Exchange Theory”, Proceedings of the 33rd Hawaii International Conference on System Sciences. [7] Rheingold, H. (2000) ”Virtual Community: Homesteading on the Electronic Frontier”, First MIT Press ed. Cambridge, Mass. : MIT Press. [8] Brown, Peter J., Bovey, John D., and Chen, Xian (1997) “Context-Aware Applications: from the Laboratory to the Marketplace”,

IEEE Personal

Communications, 4(5):58-64. [9] Kim, Yu-Hwan, Hahn, Shang-Yoon, Zhang, Byoung-Tak (2000) ”Text Filtering by Boosting Naïve Bayes Classifiers”, International ACM SIGIR Conference on Research and Development in Information Retrieval. [10] Chen, Yung-Hsin and George, Edward I. (1999) ”A Bayesian Model for Collaborative Filtering”, Department of MSIS, University of Texas at Austin, USA. [11] Pennock, David. and Horvitz, Eric (2000)

”Collaborative Filtering by

Personality Diagnosis: A Hybrid Memory- and Model-Based Approach” , 6th International Conference on Uncertainty in Artificial Intelligence. [12] IBM (2001) ” Collaborative Filtering: Community Values”, Centers for IBM ebusiness Innovation . [13] Resnick, P., et al. (1994) “GroupLens: An Open Architecture for Collaborative Filtering of Netnews”, In Proceedings of ACM CSCW’94 Conference on Computer31

Supported Cooperative Work, P.175-186. [14] Ester, M., Kriege, H.P., Sander, J., and Xu, X. (1996) ”A Density-Based algorithm for Discovering Clusters in Large Spatial Databases”,

In Proc. of

International Conference of Knowledge Discovery and Data Mining, P226-231, Portland, OR, USA. [15] Resnick, Paul, Iacovou, Neophytos, Suchak, Mitesh, Bergstrom, Peter and Riedl, John (1994) ”GroupLens ： An Open

Architecture for Collaborative Filtering of

Netnews”, In Proceedings of ACM Conference on Computer Supported Cooperative Work, New York, P175-186. [16] Yang , Ti-ya, (1990) “The Role of Members in Internet Virtual Community & the Investigation of Communication and Interaction ”, Department of Master Business Administration, National Chung Cheng University, Taiwan. [17] Wu, You-shun, (1990) “The Research of the Process in Internet Community Knowledge Sharing – Illustrated by Business Teaching Wbsite”, Department of Master Business Administration, National Cheng Chi University, Taiwan. [18] Costa, P.T., and McCrea, R.R. (1992) The Five-Factor Model of Personality and its Relevance to Personality Disorders, Journal of Personality Disorders, Vol. 6, P. 343-359. [19] Costa, P.T. (2001) http://www.personalityresearch.org/bigfive/costa.html [20] Lee, Wen-Ti (1999) ”The Personality of Taiwan Children”, Tech. Report, National Pingtung Teachers College, Taiwan. [21]Yam

Co.

(2000)

"A

survey

of

Internet

users

http://survey.yam.com/life2000/index.html, Yam Digital Technology.

32

in

Taiwan”,