A New Recommender For The Mobile Web
- 1. Expert Systems with Applications Expert Systems with Applications 32 (2007) 143–150 www.elsevier.com/locate/eswa MONERS: A news recommender for the mobile web a,* b H.J. Lee , Sung Joo Park a Sloan School of Management, Massachusetts Institute of Technology, NE20-336, 3 Cambridge Center, Cambridge, MA, 02142-1347, USA b Graduate School of Management, Korea Advanced Institute of Science and Technology, Seoul 130-722, Republic of Korea Abstract Mobile web news services, which served by mobile service operators collecting news articles from diverse news contents providers, provide articles sorted by category or on the basis of attributes, such as the time at which they were posted. The mobile web should pro- vide easy access to the categories or news contents preferred by users because user interface of wireless devices, particularly cell phones is limited for browsing between contents. This paper presents a mobile web news recommendation system (MONERS) that incorporates news article attributes and user pref- erences with regard to categories and news articles. User preference of news articles are estimated by aggregating news article importance and recency, user preference change, and user segment’s preference on news categories and articles. Performance of MONERS was tested in an actual mobile web environment; news organized by category had more page hits, while recommended news had a higher overall article read ratio. Ó 2005 Elsevier Ltd. All rights reserved. Keywords: Recommendation system; Mobile web; News 1. Introduction recommendations, using the existing web (Billsus & Paz- zani, 2000; Goldberg, Nichols, Oki, & Terry, 1992; Kon- The mobile web allows users navigate the Web using stan et al., 1997; Shepherd, Watters, & Marath, 2002). wireless devices such as cell phones and personal digital News can be filtered using collaborative filtering, which is assistants (PDAs). During early days of the mobile web, based on similarities among user preferences, or by using ring tones and wallpaper were the most commonly accessed information filtering, which uses news article keywords content, but in recent years there has been a significant and user profiles (Aas, 1997; Belkin & Croft, 1992; Foltz increase in the use of data content such as music, movie & Dumais, 1992; Konstan et al., 1997; Mostafa, Mukho- information, and news (Kelly, Gray, & Minges, 2003; padhyay, & Lam, 1997). Kim, Lee, Cho, & Kim, 2004). Because news services on User interface, user behavior, and browsing between the mobile web are so ubiquitous, there are many users. content are different on the mobile web than they are in a Currently, most mobile web news services are sorted by web-based service (Brunato & Battiti, 2003; Ho & Kwok, category or by news article attributes, such as recency. An 2003; Kim et al., 2004); therefore, the mobile web requires inconvenient user interface may constitute a barrier to its own recommendation method, adapted to its features. browsing between contents; therefore, mobile web news This paper presents a mobile web news recommendation services should provide easy access to the categories or con- system, Mobile News Recommendation Systems (MON- tent preferred by users. Researchers have conducted many ERS), which incorporates news article attributes, user pref- studies on personalizing news services in order to make erences with regard to category and news articles, and user segments. MONERS incorporates the news article’s impor- * Corresponding author. Tel.: +82 2 958 3647; fax: +82 2 958 3604. tance, as well as its recency, calculated by the difference E-mail address: fastbat@gmail.com (H.J. Lee). between the time it was posted and the present. It also 0957-4174/$ - see front matter Ó 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2005.11.010
- 2. 144 H.J. Lee, S.J. Park / Expert Systems with Applications 32 (2007) 143–150 incorporates a user segment that is focused on user profiles, (Billsus & Pazzani, 2000; Billsus, Pazzani, & Chen, reading patterns of news articles, changes in user interest, 2000). and usage patterns. Users of the actual mobile web tested the news recommender’s performance, and its performance In collaborative filtering, the nearest neighbor algorithm was measured and analyzed. requires computation that grows with both the number of Section 2 summarizes studies related to news article rec- customers and the number of products and has a sparsity ommendation and the mobile web. Section 3 introduces problem; if there are few user preferences, its recommenda- MONERS’ flow and recommendation algorithms. Section tion performance is low (Claypool et al., 1999; Sarwar 4 presents the experiment used to analyze the news recom- et al., 2000). Information filtering also has some inherent mendation method performance, and an analysis of the flaws, since it is applicable only to products that include results. Section 5 discusses this study’s implications and text-based information, and understanding the context of conclusions. keywords can be difficult. With regard to solving the short- comings of these two methods, researchers have presented 2. Related work solutions that use both methods simultaneously (Claypool et al., 1999; Kim et al., 2004). Others have presented ways 2.1. News recommendation to reduce the dimensions of the keyword vector or the user preference vector (Billsus & Pazzani, 1998), or to calculate Personalization and recommendation involving news similarity among users using a neural network or associa- services on the web use the following approaches: tion rules (Jennings & Higuchi, 1992; Mobasher, Dai, Luo, & Nakagawa, 2001; Shepherd et al., 2002). • Collaborative filtering: This is a method for calculating expected user preference for a product, using evaluation 2.2. Recommendation on the mobile web by, or the preferences of, other users who have experi- enced the product (Billsus & Pazzani, 1998; Goldberg Many studies have been undertaken on providing per- et al., 1992; Konstan et al., 1997). Collaborative filtering sonalized web service to users or on recommending person- is currently widely applied and is used not only for news alized products to web users (Billsus & Pazzani, 1998; but also for various products such as music or movies Billsus & Pazzani, 2000; Claypool et al., 1999; Goldberg (Billsus & Pazzani, 1998; Goldberg et al., 1992; Konstan et al., 1992; Konstan et al., 1997; Mobasher et al., 2001; et al., 1997). The basic input data consist of the prefer- Shepherd et al., 2002). Since a cellular phone provides a ence matrix between users and products; to collect expli- limited user interface and low data storage and business cit user preferences for this input data, a purchasing logic for clients, there is little use for keyword-based intention or implicit preferences, such as an inquiry or searches; furthermore, browsing between pages is inconve- visit, may be used. Similarity among users is calculated nient (Brunato & Battiti, 2003; Ho & Kwok, 2003; Varsh- by the Pearson correlation coefficient or the cosine mea- ney, 2003). VISCORS is a wallpaper-recommending system sure (Konstan et al., 1997; Mild & Natter, 2002; Sarwar, on the mobile web that applied collaborative and informa- Karypis, Konstan, & Riedl, 2000). Based on the similar- tion filtering to the choice and purchase of a wallpaper for ity calculation, we can find neighbors with similar ten- relieving users’ enormous navigation (Kim et al., 2004). dencies to a particular user. We can calculate the Mobile web users can be identified by contexts that include user’s expected preference for a product based on his their location and their access time and day of the week or her average preference for other products and his (Brunato & Battiti, 2003; Korpipaa, Mantyjarvi, Kela, ¨¨ ¨ ¨ or her neighbors’ preference for the product (Konstan Keranen, & Malm, 2003; Varshney, 2003; Varshney, Vet- ¨ et al., 1997; Mild & Natter, 2002; Sarwar et al., 2000). ter, & Kalakota, 2000). At present, mobile service provid- • Information filtering: Unlike collaborative filtering, ers operate various location-based services, such as information filtering is a method for finding similar ‘searching friends’, to notify users about where their friends products by comparing user information—a user pro- are, or to make recommendations about nearby restau- file—with product information. Generally, a user profile rants. Efforts are also being made to use cellular phone can be generated based on documents or information equipment to extract information about contexts, such as that the user has created or visited. After the user profile user environment, user behavior, and mechanical cellular has been compared with product information, only phone action (Korpipaa et al., 2003). ¨¨ products with a high similarity level are recommended The limitations of cellular phones and mobile devices (Aas, 1997; Belkin & Croft, 1992; Billsus & Pazzani, necessitate personalization in order to recommend appro- 2000; Claypool et al., 1999; Foltz & Dumais, 1992; priate and well-timed content to users. A mobile phone Kim, Lee, & Park, 2004; Mostafa et al., 1997). Term can also provide a more interactive service by fulfilling Frequency-Inverse Document Frequency (TF-IDF) ‘push’ service through short message services (SMS) or indexing is applied to the vector presentation, and meth- other interactive channels (Ho & Kwok, 2003). With ods for categorizing documents and information are regard to the mobile web, researchers recognize that con- applied to enhance performance of information filtering tent recommendation is one of the most important applica-
- 3. H.J. Lee, S.J. Park / Expert Systems with Applications 32 (2007) 143–150 145 tions for future development, along with mobile auctions, as there is no content usage history; in this case, they can be mobile offices, and mobile education (Varshney, 2003). placed in a segment with a similar demographic. Based on these three points, an equation to calculate 3. MONERS news article preference for an individual user might take this form: 3.1. Considerations for mobile web news recommendation pij ¼ f ðpriorityj ; recencyj ; userj ; segj ; seg weightj Þ ð1Þ Three points must be considered when designing news where pij is i user’s preference for j news article, priorityj is content personalization for the mobile web. The first is the importance of j news article, recencyj is the weight for news content. On the mobile web, news services focus on the recency of j news article, userj is the weight of prefer- the distribution of current news rather than on past or ence change with respect to the user for the category to related news articles; searching past or related articles which news article j belongs, segj is the preference of the and browsing between pages is not easy, so mobile web segment of user i for the category to which news article j news services mainly provide current reports or important belongs, and seg_weightj is the weight of preference calcu- articles sorted by news category. It is therefore important, lated by the ratio of users who have read news article j in when personalizing news, to consider the recency and the segment of user i. importance of news articles. The next section describes MONERS’ flow and how The second is the news service. On the mobile web, the each parameter of Eq. (1) can be calculated. purpose of a news service is not to provide past news arti- cles but to provide current articles. Even though it is possi- 3.2. Calculation of news article preference by user ble to learn about preferences for past news articles, it is difficult to apply the collaborative filtering method when Fig. 1 shows the flow of MONERS using the attributes there is no user preference for current news articles, or referred to in Section 3.1. When a user accesses MONERS when few users have read the present day’s articles. There- news service with a cellular phone, the system determines fore, learning user preference requires learning preferred whether he/she is a new user. If so, he/she is temporarily news categories. Preferred category information can be placed in a similar segment on the basis of demographics. used to recommend articles by selecting current news arti- If he or she has used the system before, his or her individual cles belonging to preferred categories. Information about preferences for news categories are retrieved. The news ser- changes in user interest is also helpful when recommending vice presents articles to the user after weighting their news articles, in that articles from recently preferred cate- importance and recency, determined by comparing the time gories may be selected. the article was posted on the mobile web with the present The third point relates to the social behavior of users. time. A user’s segment and the news category preferences Users can be classified into a user group by article usage of the user, as well as his or her preferences, are regularly pattern and demographic information. User segments with updated after collecting user data; calculations about news similar content or similar usage patterns can be used to articles can be made only with a user’s request in real time. make recommendations by selecting articles preferred by The following are MONERS’ methods of calculation for the segment. It is not easy to place new users in a segment, each parameter of Eq. (1): Fig. 1. Mobile news recommendation systems (MONERS) flow.
- 4. 146 H.J. Lee, S.J. Park / Expert Systems with Applications 32 (2007) 143–150 • News importance News articles are rated from level 1–7, according to their importance; the level determines the article’s weight. For example, the weight of headline news, at the ‘most impor- tant’ level 1, is 1.5; the weight of ‘important’ level 2 is 1.45. News content providers such as off-line newspapers supply information on the importance of news articles. • News article recency Recent news has more weight and, as time passes, its weight is lowered. Eq. (2) calculates the weight decrease in news article recency, shown in Fig. 2. 2 timepresent À timeposted recencyj ¼ À þ1 ð2Þ 86; 400 where timepresent is the present time and timeposted is the Fig. 3. Weight of category preference change. time article j was posted to the mobile news service. The difference between the two times is calculated by As shown in Fig. 3, Eq. (3) is devised so that the value of the second unit and a denominator 86,400, meaning sec- the category preference change weight becomes 1 when onds per 24 hours. If the article-posted time is short-term and long-term preferences are the same, 2:37:14 PM and the current time is 4:13:5 PM, the re- and the weight’s value of the weight is between 0.5 and cency weight of j news article is Àð58;385À52;634Þ þ 1 ¼ 86;400 1.6 even if the short-term category preference increases 0:9334375. If the difference between the present and or decreases sharply. If userj,s_days is 0.5 and userj,l_days posted time is over 24 hours, the recency weight of the is 0.25, the preference change weight value is 1.2262; article becomes zero. i.e., if recent preference for the category doubles, the • Change in user interest change in preference weight value increases 1.2262 times. When applying users’ category preferences, more weight • User segment category preference is given to a news category that has drawn recent user This is the news category preference of the segment to interest. The following equation gives more weight to which a user belongs. A user segment is formed through a category that has an increased preference over a desig- k-mean clustering using Euclidean distance measure. nated period of time: Clustering is performed using a user’s category prefer- userj;s days 1ÀDF ence, age, and sex. userj ¼ log10 ð10 À 1Þ þ 1 þ DF userj;l days Supposing that users in the same segment prefer similar ð3Þ kinds of news, the segment’s preference for a news cate- gory is calculated as follows: where userj is the preference change weight for the cate- gory to which j news article belongs, and userj,s_days is rsc segsc ¼ PN ð4Þ the user’s category preference for the past s_days; s_days j¼1 rsj refers to a short-time preference and l_days to a long- where rsc is the number of times that users in segment s time preference. MONERS uses a 1-week preference as read news articles in category c, N is the total number of a short-time preference and a 3-week preference as a categories, segsc is segment s’s preference for category c, long-time preference. DF is a differential factor with a the value of which is obtained by dividing the number of value of 0.5; when the DF value is 0.5, the weight value times that users read the articles in category c by the can be obtained as shown in Fig. 3 according to the ratio user days number of times that users read articles in the total num- of userj;s days . j;l ber of categories. In Table 1, seg11 is 12 ¼ 0:4. 30 It is difficult to assign new users to a segment via clus- tering because they have no or little experience using mobile web news services. Therefore, a temporary seg- ment is established for new users; after gaining a cer- tain amount of experience, they can be assigned to a Table 1 Category preference by segment Category 1 Category 2 Category 3 Total Segment 1 12 8 10 30 Segment 2 5 10 15 30 Segment 3 12 18 10 40 Total 29 36 35 Fig. 2. Weight decrease for news article recency.
- 5. H.J. Lee, S.J. Park / Expert Systems with Applications 32 (2007) 143–150 147 segment via clustering. The temporary segment for new When the news importance weight for article j is 1.5, pij users is determined by age and sex; the most suitable (user i’s expected preference for news article j) can be deter- segment for their age and sex can be found through mined by using the weights that are calculated using the the following equation: above equations as follows: U age À Ai;age U sex À Ai;sex pij ¼ 1:5  ð1 þ 0:4Þ Â 0:9334375  1:2262  1:08 MIN þ ð5Þ S i;age S i;sex ¼ 2:5959: where Uage and Usex are the user’s age and sex; males are indicates with 1, and females with 0. Ai,age is the average 4. Experiments and results age in segment i and Si,age is the standard age deviation in segment i. Similarly, Ai,sex is the average sex value in 4.1. Experiments segment i and Si,sex is the standard sex deviation in seg- ment i. An experiment on the use of this news recommendation • User segment article preference service was carried out on service members who were reg- If only the weight of category preference for all users in istered with a Korean mobile service provider’s intelligent the same segment is applied, all new articles in the same wireless service. The news recommendation service’s usage category are given the same preference weight. There- log from October 2003 to the end of April 2004 was fore, more weight is given to an article read by more analyzed. users in the same segment, making it more likely that When a news service is accessed, screens similar to ones this article will be recommended. The weight of article in Fig. 4 appear. The root menu of a news service consists preference for users in a segment is calculated as follows: of recommended news, news by category, and current rj news. Recommended news provides personally recom- seg weightj ¼ 1 þ Pm ð6Þ s¼1 r s mended news articles from all categories; current news where rj means the number of times users in a segment arranges news articles according to recency; news by cate- read article j (1 6 j 6 m), and m is the total number of gory organizes news into categories such as politics, busi- articles. When users in a segment read a total of 50 arti- ness, sports, etc.; if a user selects a category, articles in cles 250 times and read article j 20 times, seg_weightj is that category are displayed. 20 ð1 þ 250Þ ¼ 1:08. 4.2. Comparison of recommendation performance The following equation calculates each user’s preference for each news article, when all the above factors are taken Fig. 5 shows the number of times users accessed each into consideration: menu; this information was gathered by analyzing the news service usage log. pij ¼ priorityj  ½1 þ segsc Š Although the number of times users accessed each menu 2 # timepresent À timeposted varied, recommended news and news by category were  À þ1 more popular than current news. 86; 400 Various measures have been used to compare the perfor- userj;s days mance of news service recommendations; these have been  log10 ð101ÀDF À 1Þ þ 1 þ DF userj;l days based on errors between users’ evaluation scores and expected preference (Breese, Heckerman, Kadie, 1998; rj  1 þ Pm Sarwar et al., 2000). It is difficult to show whether an article s¼1 rs has been read on the mobile web because it is not easy to ð7Þ obtain a user’s explicit rating of an article, and each article Fig. 4. Examples of news service screens: (a) news service, (b) recommended news, (c) current news, (d) news categories.
- 6. 148 H.J. Lee, S.J. Park / Expert Systems with Applications 32 (2007) 143–150 Fig. 7. Article read ratio by menu. one article consisting of several pages. An article is usually composed of four or five pages; if users read the entire arti- Fig. 5. Number of times users accessed each news menu. cle, it probably indicates that they find it interesting. An ANOVA test verified the difference between news hits by menu and by article read ratio. The ANOVA test’s null hypothesis, that verifies the difference of news hits by menu, is H0: l1 = l2 = l3, where li constitutes average news hits on menu i. As shown in Table 2, results indicated that the null hypothesis must be rejected because the value of F is 180.751 and the p-value is 0.0000. This suggests that there is a difference between news access hits by menu. Duncan’s test using Post Hoc analysis indicates that there were more hits for news sorted by category than for recom- mended news, and that there were a similar number of hits for current news and recommended news. The ANOVA test’s null hypothesis for verifying the dif- ference of the article read ratio by menu is H0: l1 = l2 = l3 Fig. 6. Number of news accesses by menu. and li is the article read ratio for menu i. As shown in Table 2, the null hypothesis is rejected because the value of F is 15.970 and the p-value is 0.000. Rejection of the null consists of several pages. This study was able to measure hypothesis indicates that there is a difference between arti- how many articles and how many pages of an article on cle read ratios by menu. Duncan’s test using Post Hoc anal- the mobile web were read. ysis suggests that the article read ratio for recommended Fig. 6 shows the number of news articles read when a news is higher than that for news by category or recency. user accessed one menu. News by category had more hits The article read ratio for news by category is higher than than did recommended news or current news. News hits that for news by recency. are the number of pages accessed; if a user sees two pages of an article consisting of four pages, it results in two news 5. Discussion hits. The number of menu sessions shows how many times a user accessed the news selection menu (current, catego- 5.1. Performance of MONERS recommendation ries, and recommended), rather than news content. Fig. 7 shows the article read ratio of news articles by As the comparison between performances in Section 4.2 menu; this measures how far users read when they access indicates, news by category had the most news hits, and Table 2 Results of ANOVA test News hit Article read ratio Average Std. Average Std. Category 2.21 1.899 2.01 1.623 Recommended 1.17 0.593 2.27 1.175 Current 1.11 0.411 1.74 1.377 ANOVA (F-value) 180.751 (sig. = 0.000) 15.970 (sig. = 0.000) Duncan’s test Category Recommended, Current Recommended Category Current
- 7. H.J. Lee, S.J. Park / Expert Systems with Applications 32 (2007) 143–150 149 current and recommended news had a similar number of mation filtering alone is applied to news recommendation, hits. News by category enables users to read various arti- it can be difficult to recommend important current news, cles in one category; they can access news categories inter- including new keywords or hot news recently accessed by esting to them such as MLB news or entertainment. Users many users. Therefore, news services on the mobile web often selected several news articles in one category; how- can recommend or present articles based on keywords that ever, if a news article consisted of several pages, they can either be directly input by users or picked up using tended to read more pages of recommended articles than information filtering. For example, if ‘‘U2’’ is a user’s pre- articles sorted by category or recency. Recommended news ferred entertainer and ‘‘Curt Schilling’’ is the preferred had the highest article read ratio, followed by news by cat- athlete, articles that include these keywords will be recom- egory; news sorted by recency had the lowest ratio. A high mended or presented to the user on the mobile web. article read ratio indicates that users find an article interest- ing; it is useful, therefore, to personalize news categories for 6. Conclusion and limitations users, and to present it in the mobile device’s upper news menu. It can also be useful to provide personalized recom- This paper presented the MONERS news recommenda- mended news articles within various news categories as rec- tion system, a system that incorporates the characteristics ommended news in the upper menu. of users as well as the nature of the mobile web news service and content. MONERS calculates the distance between the 5.2. Collaborative filtering and information filtering as present time and the time an article was posted to deter- methods for recommending news mine news recency; it also incorporates the importance of news articles. MONERS also considers changes in user The application of collaborative filtering to news rec- preference with regard to news categories. It can identify ommendation on the mobile web requires learning about preference for news category by user segment and prefer- preferences for news articles or categories. This will allow ence for particular articles by users in the segment. for detailed measurements of similarities among users, as For real-time recommendation, regular analysis, which well as expected preferences. When dealing with a great would include learning about various preferences and seg- deal of information, sparse data can lead to low perfor- ment composition, is carried out by batch work. The func- mance (Billsus Pazzani, 1998; Mild Natter, 2002; Sar- tion of providing personalized news content is performed war et al., 2000). As it is inconvenient to search or browse only when users access a news service. This study measured past articles using the mobile web, only current news is the performance of recommendations using the article read provided. The most current news and news with high ratio; this measure is only suitable when a news article con- importance get the most hits on the mobile web. When sists of several pages. While news by category had more users have read the current articles, preferences about page hits, the article read ratio for recommended news arti- news articles can be learned, and collaborative filtering cles was higher than for news articles sorted by recency or can be applied; however, it is difficult to recommend a news category. news article to users if the article has not been read or This study focused only on news service; future research only a few users have read it. should devise a method to expand this personalized recom- In collaborative filtering preferences, learning about mendation service to general content supplied by the news categories reduces the scalability or sparsity problem mobile web, as well as to content only the mobile web more effectively than does learning about preferences with can supply. regard to news articles. However, the level of recommenda- tion is limited to news categories; since the preference value is calculated on the level of news category, all articles in References one news category such as baseball news or entertainment Aas, K. (1997). A survey on personalized information filtering systems for news have the same preference value. In this situation, arti- the World Wide Web. Norwegian Computing Center. cles in a news category are arranged according to their Belkin, N. J., Croft, W. (1992). Information filtering and information recency or importance without considering the preferences retrieval: two sides of same coin. Communications of the ACM, 35(12), of users. MONERS learns preferences about news catego- 29–38. ries with user segments, and calculates preferences for news Billsus, D., Pazzani, M. J. (1998). Learning collaborative information articles in categories by utilizing an article’s recency and filters. In Proceedings of 15th international conference on machine learning (pp. 46–54). Madision, Wisconsin. importance, along with the number of hits by users in the Billsus, D., Pazzani, M. J. (2000). User modeling for adaptive news same segment. access. User Modeling and User-Adapted Interaction, 10, 147–180. When information filtering is applied to news, keywords Billsus, D., Pazzani, M. J., Chen, J. (2000). A learning agent for wireless from news articles read by users are extracted to create a news access. In Proceedings of IUI 2000 (pp. 33–36). New Orleans, USA. user profile, which is then compared with registered news Breese, J. S., Heckerman, D., Kadie, C. (1998). Empirical analysis of articles to allow for recommendation of news articles with predictive algorithms for collaborative filtering. In Proceedings of the a high similarity to a user profile (Belkin Croft, 1992; fourteenth conference on uncertainty in artificial intelligence (pp. 43–52). Foltz Dumais, 1992; Mostafa et al., 1997). When infor- Madison, Wisconsion.
- 8. 150 H.J. Lee, S.J. Park / Expert Systems with Applications 32 (2007) 143–150 Brunato, M., Battiti, R. (2003). PILGRIM: A location broker and Konstan, J. A., Miller, B. N., Maltz, D., Herlocker, J. L., Gordon, L. R., mobility-aware recommendation system. In Proceedings of IEEE Riedl, J. (1997). GroupLens: applying collaborative filtering to PerCom2003 (pp. 265–272). Fort Worth, Texas, USA. USENET news. Communication of the ACM, 40(3), 77–87. Claypool, M., Gokhale, A., Miranda, T., Murnikov, P., Netes, D., Korpipaa, P., Mantyjarvi, J., Kela, J., Keranen, H., Malm, E.-J. (2003). ¨¨ ¨ ¨ ¨ Sartin, M. (1999). Combining content-based and collaborative filters in Managing context information in mobile devices. IEEE Pervasive an online newspaper. In Computer science technical report (pp. 1–11). Computing, 2(3), 42–51. Worcester Polytechnic Institute. Mild, A., Natter, M. (2002). Collaborative filtering or regression models Foltz, P. W., Dumais, S. T. (1992). Personalized information delivery: for Internet recommendation systems? Journal of Targeting, Measure- an analysis of information filtering methods. Communications of the ment and Analysis of Marketing, 10(4), 304–313. ACM, 35(12), 51–59. Mobasher, B., Dai, H., Luo, T., Nakagawa, M. (2001). Effective Goldberg, D., Nichols, D., Oki, B. M., Terry, D. (1992). Using personalization based on association rule discovery from web usage Collaborative filtering to weave an information tapestry. Communica- data. In Proceedings of WIDM 2001 (pp. 9–15). Atlanta, GA, USA. tion of the ACM, 35(12), 61–70. Mostafa, J., Mukhopadhyay, S., Lam, W. (1997). Multilevel approach Ho, S. Y., Kwok, S. H. (2003). The attraction of personalized service for to intelligent information filtering: model, system, and evaluation. users in mobile commerce: an empirical study. ACM SIGecom ACM Transactions on Information Systems, 15(4), 368–399. Exchanges, 3(4), 10–18. Sarwar, B., Karypis, G., Konstan, J., Riedl, J. (2000). Analysis of Jennings, A., Higuchi, H. (1992). A personal news service based on a recommendation algorithms for e-commerce. In Proceedings of Elec- user model neural network. IEICE Transactions on Information and tronic Commerce’00 (pp. 158–167). Minneapolis, Minnesota. Systems, E75-D(2), 198–209. Shepherd, M., Watters, C., Marath, A. T. (2002). Adaptive user Kelly, T., Gray, V., Minges, M. (2003). Broadband Korea: internet case modeling for filtering electronic news. In Proceedings of the 35th annual study. International Telecommunication Union. Hawaii international conference on system sciences (p. 102b). Hawaii, Kim, C. Y., Lee, J. K., Cho, Y. H., Kim, D. H. (2004). Viscors: a visual- USA. content recommender for the mobile web. IEEE Intelligent Systems, Varshney, U. (2003). Location management for mobile commerce 19(6), 32–39. applications in wireless internet environment. ACM Transactions on Kim, J. W., Lee, H. J., Park, S. J. (2004). Intelligent knowledge Internet Technology, 3(3), 236–255. recommendation methods for RD knowledge portals. Journal of Varshney, U., Vetter, R. J., Kalakota, R. (2000). Mobile commerce: a Electronic Science and Technology of China, 2(3), 80–85. new frontier. IEEE Computer, 33(10), 32–38.