Wi-Fi Walkman A wireless handhold that shares and recommend

Wi-Fi Walkman: A wireless handhold that shares and recommend music on peer-to-peer networks Jun Wang, Marcel J.T. Reinders, Johan Pouwelse, Reginald L. Lagendijk

Faculty of EWI, Delft University of Technology,

Mekelweg 4, 2628 CD Delft, The Netherlands

________________________________________________________________________ The Wi-Fi walkman is a new application that investigates the technological and usability aspects of human-computer interaction with personalized, intelligent and context-aware wearable devices in ad-hoc wireless environments such as the future home, office, or university campuses. It is a small handheld device with a wireless link that contains music content. Users carry their own walkman around and listen to music. All this music content can be shared using ad-hoc networking. The walkman naturally interacts with the users and it is situated in peer-to-peer environments. Without annoying interactions, it can learn the users’ music interest/taste and consequently provide personalized music resources according to the current situated context and user’s interest.

______________________________________________________________________

1. Introduction

Recently, with the rapid progress in information processing, communications, and storage technologies, the amount of information that we deal with in our daily lives has been rapidly increased and even more the types of information have been changed from homogeneous (textual) data only to heterogeneous data (audio, video, image, etc). We enjoy the entertainment and convenience brought to us by a variety of sources coming from digital TV, mp3 player, digital still image/video camera, etc., but are hampered to access this data by its sheer amount.

Not only the types of the information changed, but also the way people consume information changed. Peer-to-peer and ad-hoc networks, as new network topology, become a new way for people to distribute, exchange, and consume resources from their local storage devices in many different locations, such as the future home, office, or university campuses. There are two significant advantages of peer-to-peer and ad-hoc networks: 1) the replicas of the content among peers increase the content availability, 2) for the exchange of information, no requirements of centralized storage and management from the third parties make these networks have very low costs. Recently, those attributes attract a large body of people in the internet domain. For instance, through the internet peer-to-peer networks, such as Freenet[3] and Gnutella[4], a large number of people gained access to each other’s shared files. Furthermore, we believe that, in the recent future, the wireless communication technology will make those peer-to-peer networks wireless and exist in any place, and at any time.

In ad-hoc network environments, the volume of information is increasing far more quickly than our ability to digest it. The traditional textual keywords-based information retrieval approaches [5,6,7,8] encounter three major problems. Firstly, the transition from textual data to heterogeneous data requires large amount of textual Meta data on the one

hand. It is practically intractable to ask people to provide content as well as associated Meta data at the same time. On the other hand, automatic content analysis on the non-textual data is far from being efficient to get the Meta data that we need. Secondly, keywords are not semantically expressive enough to enable a seamless search, i.e. people hardly issue a textual query when they can not exactly express what they are looking for. Thirdly, in mobile environments, the user interface is constrained and consequently does not permit complex interactions between users and their handheld/wearable devices. Automatically assisting the user to acquire information and/or services that fits his/her interests is a non-trivial problem. Unfortunately, today’s computers merely act as information provider. One of the solutions to close this information gap is to increase the ability of computers to interpret the user’s interests and select relevant information on the user’s behalf.

To this regard, the research on information filtering is aroused to filter out, refine and systematically represent the relevant information. One of the solutions for overcoming the information overload is to provide personalized suggestions based on a history of a user’s likes and dislikes.

The Wi-Fi walkman is a case study that investigates the technological and usability aspects of human-computer interaction with personalized, intelligent and context-aware wearable devices in ad-hoc wireless environments such as the future home, office, or university campuses. It is a small handheld device with a wireless link that contains music content about the environment or from the user. Users carry their own walkman around and can both listen and record music content. All this music content can be shared using mobile ad-hoc networking. The walkman is situated in a peer-to-peer environment and naturally interacts with the users. Without annoying interactions with users, it can learn the users’ music taste and consequently provide personalized music resources to fit the user’s interest according to the user’s current situated context.

2. Related Work

Internet based peer-to-peer networks increase rapidly and it has given a large number of people the possibility of sharing resources in their local storage devices [1,2]. Recently, sharing resources in wireless networks has received some attention. In [2], the TunA system allows users to “tune in” to other nearby TunA music players and listen to what someone else is listening to. Another system, SoundPryer [1] allows drivers to jointly listen to music shared between cars on the road. Interestingly, these two applications show that the upcoming technologies start to take care about their social impact on everyday life, i.e. they bring people together that have been socially separated by the technologies for the last decades (such as TV, Internet, portable music player, etc.) Clearly, those technologies [1,2,3,4] are different from the traditional technologies in that they encourage people to make social interactions such as sharing and exchanging information. However, those applications are implemented far away from being called intelligent devices which aims to provide personalized services on user’s behalf. Differently, we present here a system which has the ability to steer the user’s interests/tastes and select/represent relevant information on the user’s behalf.

One of the most promising widely implemented and familiar technologies to understand user’s interest is collaborative filtering [9, 10, 11, 12]. Collaborative filtering based approaches utilize the correlations (commonalities) between users on the basis of their

ratings to predict and recommend items which have the highest correlations to the user’s rated /purchased items (user’s interest). Here, we will show how to use collaborative filtering to create a personalized music delivery system in a peer-to-peer environment.

3. Wi-Fi Walkman

The prototype of the Wi-Fi walkman on a Sharp Zaurus PDA is shown in Figure 1.

Fig. 1 The Wi-Fi walkman prototype on a Sharp Zaurus PDA

The Wi-Fi walkman allows to exchange music files (MP3 formatted) in a mobile network in a personalized way. The music files are stored on a local storage device (e.g. hard disk, fresh memory) of each Wi-Fi walkman and can be accessed through the Wi-Fi mobile network. The key issue in the Wi-Fi walkman is how to locate music files that the user will be interested in. To this regard, music recommendation is implemented as a user oriented music file filter to help the user to find relevant or desired music files according to current situated context and learned user interest.

3.1 Definition

In this section, we will define our research problems. Consider the case that users share music content in a peer-to-peer network. Each peer represents a Wi-Fi walkman used by a particular user.

Lets define the set of peers as:

{1}i P i M ,=,..., (1)

where M is the number of the peers currently online in the peer-to-peer network. That means they can be located and accessed with the sufficient bandwidth. Since the peers (Wi-Fi walkman) and users exist in pairs, we will use the term peer and user interchangeablely.

The music content in the network is defined as a set of items, denoted by the set I . Each item has a specific physical location, i.e.

,{{1}{1}}i j i I I i M j N =|=,...,;=,..., (2)

where N i is the number of items physically located in the local storage device by the peer P i .

User retrieves music content regarding to his/her own interests. In a particular time, user may have one particular interest/intent. The interest can be obtained either explicitly or implicitly. For instance, it could be explicitly obtained by asking users to rate items. Alternatively, this can also be implicitly indicated by the music items that the user is playing. In our Wi-Fi walkman, we use user’s music play-list to indicate the user’s music interest:

Formally, we use a vector ,{},{1}{1}i j q q i V v i M j N ==,...,;=,..., to represent the play-list of the user q P , where the element ,1i j q v =, if user q P played the item ,i j I ,

otherwise ,0i j q v =.

{{1}}q V V q M =|=,...,

,{), {1}{1}i j q q i V v i M j N ==,...,;=,..., (3)

We would like to note that generally the interest of the user will change over time. It is in fact depends on the current context. Therefore, the play-list (representing the current users’ interest) should ideally be dependent on the time also, i.e. ()q q V V t →.

We utilize a time window to forget the old music items users have played, as shown in Fig. 2. By doing so, system can more focus on the current user’s interest.

Fig. 2 Time window for forgetting. 3.2 Personalized Recommendation

In order to automatically recommend the appropriate music regarding to the user’s current interest (indicated by the play-list q V ), we denote a distance measurement between a music item and the play-list as ,(,)i j q d I

V to measure the likeness of the item ,i j I to the play-list q V . If d is smaller than a threshold D , the system will

recommend item ,i j I

to user q V . The list of recommendations for user q V is expressed

as follows:

,,Rec {|(, ) & } ,

{1}{1}

i j i j q q i I I d I V D i q i M j N =∈<≠=,...,;=,..., (4)

Play sequence

The current recommender system is implemented by using the collaborative filtering technique. Collaborative filtering utilizes the correlations (commonalities) between users on the basis of their ratings (i.e. the play-lists of users) to predict and recommend music items which have the highest correlations to the user’s preference (user’s current play-list).

The accuracy of the collaborative filtering directly relies on the number of users, who provide their ratings. In mobile networks, the density of peers may vary strongly depending on the local situation. (For instance, on the bus, there are only a dozen of people while in the airport there are thousands of peoples. Depending on the current density of peers, we perform recommendation by two different approaches, namely the flooding model and the client/server model.

Flooding Model

When the density of peers is large (i.e. thousands of users) and the play-lists from those users are enough to obtain a good recommendation, we use the flooding approach to find the correlations between users.

By using the correlation [12], the similarity between two play-lists q V and p V is calculated as follows:

(,)q p Sim V V = (5) ,1

i N M i j q q M i j i i v v N

=∑∑∑

where q v is the mean rating of the user q V .

The distance measurement between a music item ,i j I and a play-list from user q P can be calculated [12,13] as the weighted average rating. This is shown as follows:

,,,(,),(,)(,)()q p i j i j p q p q p q q sim V V T d I V v k

sim V V v v ?>=+?∑ (6)

where k is the constant value. In the flooding model, the play-list of the user is broadcasted to all its neighbors in order to determine the recommendation for that user. The neighboring peers check the similarity (using in Eq.(5)) between the received play-list and their own play-list. They decrease the TTL (Time to Live) field of the broadcasted play-list and then pass it to their neighboring peers again until the TTL count reaches 0. If one of the neighboring peers possesses that the similarity between its play-list and the broadcasted play-list is higher than T , the items in the play-list (including the locations) are sent back to the initial peer that posed the query. Finally all items received

by the querying peer are ranked according to the distance measurement (Eq.(6)) and consequently the top-N ranked items are recommended to the user.

Client/Server Model

When the density of the peers is small and consequently the play-lists (rating) from those users are not enough to obtain a good recommendation, we have to access a predefined rating database and use the database to calculate the recommendation. In this model, we assume the peer has a chance to access a server which has a rating database. The rating database stores the play-list of the users in the networks.

Fig.3 illustrates the procedure of obtaining the recommended play-list. In order to reduce the computational complexity, we apply the item-based recommendation algorithm proposed in [14] in the server part to calculate the recommendations.

In item-based recommendation, each music item can be represented by who has played it. More frequently, each item ,i j I can be represented by a vector ,i j U , where its element ,1q i

j u = if the item ,i j I has been played by the peer q P .

Item-based recommendation is then performed by exploring the correlations between the items rather than the correlations between users. Recommendations are created by finding items that are similar to other items that the user according to:

,',',',',','(,)(,)()()

i j i j i j i j i j i j Freq I I sim I I Freq I Freq I =× (7)

where ,()i j Freq I is the number of times that item ,i j I is in any of the play-lists. ,','(,)i j i j Freq I I is the number of times that item ,i j I and ','i j I are in the same play-list.

Due to the fact that the item-to-item matrix is relatively static, it is possible to compute this matrix offline. This extremely reduces the computational demands.

3.3 Demonstration Implementation

The Wi-Fi walkman demonstration is implemented on the Sharp Zaurus PDA by using C++, as shown in Fig. 1. It is running on an ad-hoc wireless network. It features audio playback, audio storage, audio recommendation, and ad-hoc wireless connectivity for audio exchange.

The Wi-Fi walkman itself contains an audio agent, a transport agent, and a wireless interface shown in Fig. 3. The audio agent is responsible for communication with the recommendation services, managing the MP3 files on the storage devices (e.g. a fresh card), and selecting which MP3 to play. The transport agent uses the wireless ad-hoc network to communicate with other transport agents and share the music files. Due to the dynamic nature of an ad-hoc network, the transport agents must keep track of the other walkmans around them. The standard 802.11 functionality is to transfer TCP/IP packets.

The enhanced ad-hoc wireless interface also informs the transport agent of new

Fig. 3 Illustration of the Wi-Fi Walkman in client/server model

Sending Play-list

recommended Play-list

Peer

Fig. 4 Recommendation in the client/server model.

The recommendation is implemented in the server part. We utilize a dataset in the AudioScrobbler1 community as our play-list dataset. Currently this dataset has 857.020 1 592f070879563c1ec5da7192/

tracks and 4.175.146 playback actions. The interaction between each peer and the server is illustrated in Fig. 4

Snap-shots of the Wi-Fi walkman application are shown in Fig. 6. The procedure to obtain the suitable music files to fit the user’s interest is illustrated in Fig. 5 and each step is described as follows:

Wi-Fi_Walkman()

Begin

V t to represent the user’s current interest from the play-list by

1.Create ()

q

utilizing a time window.

V t to the recommendation server.

2.Send ()

q

3.Get recommendation from server

4.Finding online peers and obtain the music item list from those peers

5.Select music items from the item list according to the recommendation

6.Locate the recommended items and download/stream them

7.Playback the obtained items

End.

4. Conclusions

In this paper, we introduce a new wireless application called Wi-Fi walkman. In this application, we investigate the technological and usability aspects of human-computer interaction with personalized, intelligent and context-aware wearable devices in ad-hoc wireless environments such as the future home, office, or university campuses.

Without bothering users for any annoying keywords input, the Wi-Fi walkman can steer user’s music interest and recommend appropriate music in the peer-to-peer networks.

In our framework, user’s interest is inferred by the play-list of a user. Based on collaborative filtering methods, system recommends music to users both in the blooding model and the client/server model depending on the local density of the peers.

Figure 5. System Diagram of the Wi-Fi Walkman application

Step. 5Recommended

play-list with the

peers and their

Locations in current ad-

hoc network

current ad-hoc

Fig. 6 Snap-shots of the Wi-Fi Walkman prototype

5. References

[1]M. ?stergren. “Sound Pryer Field Trials: Learning About Adding Value to Driving”,

in the workshop Designing for ubicomp in the wild: Methods for exploring the design of mobile and ubiquitous services, In the proceeding of MUM'2003., 2003.

[2] A. Bassoli, C. Cullinan, J. Moore, and S. Agamanolis. “TunA : a mobile music

experience to foster local interactions(poster)”, in UbiComp 2003 the Fifth International Conference on Ubiquitous Computing, Seattle, 12-15 October 2003.

[3]FreeNet, 592f070879563c1ec5da7192

[4]Gnutella, 592f070879563c1ec5da7192

[5]O. D. Gnawali. “A keyword set search system for peer-to-peer networks”, Master’s

thesis, Massachusetts Institute of Technology, June 2002.

[6]On J. Li, B. Loo, J. Hellerstein, F. Kaashoek, D. Karger, and R. Morris. “On the

feasibility of peer-to-peer web indexing”, In Proc. of the 2nd Int. Workshop on Peer-to-Peer Systems, 2003.

[7]Brian Cooper and Hector Garcia-Molina. “Studying search networks with SIL” In the

preceding of IPTPS, 2003.

[8]Bobby Bhattacharjee, Sudarshan Chawathe, Vijay Gopalakrishnan, Pete Keleher,

Bujor Silaghi. “Efficient peer-to-peer searches using result-catching”, In Proc. of the 2nd Int. Workshop on Peer-to-Peer Systems, 2003.

[9]U. Shardanand, P. Maes, 1995. “Social Information Filtering: Algorithms for

Automating ‘Word of Mouth’ ”, In Proceedings of the Conference on Human Factors in Computing Systems (CHI95), 210-217, Denver, Co, ACM Press.

[10]B. Sarwar, G. Karypis, J. Konstan, J. Riedl, 2001. “Item-based collaborative filtering

recommendation algorithms”, In Proceedings of WWW10 Conference, pages 285-- 295, Hong Kong.

[11]J. Konstan, Bo Miller, D. Maltz, J. Herlocker, L. Gordon, J. Riedl. “GroupLens:

Applying Collaborative Filtering to Usenet News”, Communications of the ACM, 40(3), pp. 77-87, 1997

[12]J. S. Breese, D. Heckerman, and C. Kadie, Empirical analysis of predictive

algorithms for collaborative filtering. In Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence (UAI-98). G. F. Cooper, and S. Moral, Eds.

Morgan-Kaufmann, San Francisco, Calif., 43-52. 1998.

[13]J. A., Konstan, B. N. Miller, D. Maltz, J. L. Herlocker, L. R. Gordon, and J. Riedl,

GroupLens: applying collaborative filtering to suenet news. Commun. ACM 40. 77-

87. 1997.

[14]K. George, Evaluation of item-based top-N recommendation algorithms, Technical

Report #00-046, Dept. of C.S., Univ. Of Minnesota, 1999.

本文来源：https://www.bwwdw.com/article/tzoq.html