Recommendation in Education Portal by Relation Based Importance Ranking 45 3. Iteratively calculate the importance rank vector. Let S be the current stage, which is number n itera- tion, result of importance score. Then S’ = (A’) T S , S’ is the importance rank result in n+1 iteration. Calculate the variance of all the entities’ impor- tance rank in S’ and S by ∑ − ′ = i ii ssv )( . 4. if ( ε <v or n==m) goto 4; else repeat 3. 5. Output the importance rank of every entity in the education portal. Recommend the top n resources. Although in our approach, the importance of every entity can be achieved, we only recommend the top n resources to web page visitors. However, if we want to recommend other typed entities like users, it’s also straight forward to do that kind of recommendation. 4 Experiments In this section, we give experimental result in a real world data set. The real world data set is made up of academic papers published by three departments of Tsinghua University: Department of Computer Science, Department of Electronic Engineering and Department of Automation. The network can be constructed by the following elements: 1. Resource: paper is the resource in this experiment; 2. User: paper author is the user in this experiment; 3. Tag: keyword of the paper is the tag in this experiment; 4. Department: there are three departments in this experiment; 5. Category: research fields in these three departments are categories in this ex- periment; 6. User-resource relation: author of a paper will have this relation between him/her and the paper; 7. User-tag relation: author of a paper will have relations between him/her and the paper’s tags; 8. Tag-resource relation: tags of a paper will have relations between them and the paper; 9. Resource-category relation: the papers belong to one research field will have relations between them and the category represents the research field; 10. Category-department relation: the research fields belong to one department will have relations between them and the department; 11. Department-user relation: the department will have the relation to users if the users are belonged to this department. By this method, we extract 300 papers in these three departments. The papers are from the IEEE Explorer [7] web site. We get the network with the following numbers of nodes and relations: 46 X. Wang, F. Yuan, and L. Qi Table 1. The number of entities and relations Resource 300 User 82 Tag 412 Department 3 Entity Category 18 User-resource 950 User-tag 1330 Tag-resource 542 Resource-category 300 Category-department 18 Relation Department-user 82 The tags are prepared by extracting the keywords from the papers in the following two approaches: the first one is to extract the words in the “keyword” section in the paper. If the paper doesn't contain the “keyword” section, the tags are extracted by getting the first three frequent words in the papers’ abstract. The tags are extracted mainly in a manual way. The categories are prepared by using the papers’ conferences and journals in which the paper is published. The conferences and journals have been classified to different categories, for example, database, communications, machine learning. To evaluate the performance of our importance rank calculated by our approach, we use people to annotate the results output by our approach. The algorithm outputs 500 pairs which contain two resources, the first resource’s importance score is larger than the second one. If a tester agrees with this result, he/she will approve this result. Otherwise, he/she will decline this result. Because different testers may have different opinions about the results, the final result is achieved by a voting method. We use five undergraduate and graduate students as the tester and randomly generate 500 pairs of the results for them to annotate. The evaluation factor is the error rate of the evalua- tion result. )( )( resultnumber resultdeclinednumber err rate = (11) And we have repeated generation 500 pairs for five times, for each time, the error rate is as follows: Table 2. The error rate in five randomly generated test sets Test set 1 Test set 2 Test set 3 Test set 4 Test set 5 0.20 0.18 0.21 0.23 0.20 Recommendation in Education Portal by Relation Based Importance Ranking 47 The average error rate is 0.20, which is acceptable for users to find information which may interest him/her. We also try to use different λ ut and λ ur values to inspect how the values affect the results. We calculate the average error rate using the same method in above. We re- peated generation 500 pairs for five times and calculate the average error rate. The experimental results are as follows: Table 3. The result comparison in different values of λ ut and λ ur λ ut λ ur Average error rate 0.9 0.1 0.28 0.8 0.2 0.26 0.7 0.3 0.24 0.6 0.4 0.20 0.4 0.6 0.18 0.3 0.7 0.21 0.2 0.8 0.25 0.1 0.9 0.27 We can find that in this scenario, the results are better when λ ut is relatively a little smaller than λ ur . However, we can get the best result when λ ut = 0.4 and λ ur =0.6, which is 0.18. And it’s not much different from 0.20, which is obtained when we set λ ut =λ ur =0.5. 5 Related Works Recommendation has been a popular research topic in both industry and academic. However, there are still few researches about recommendation in education portal. Thus, we will review the current methods in other kinds of applications. Online sellers usually use recommendation to encourage customers to buy more items. For example. www.amazon.com, www.joyo.com and www.taobao.com all have their own methods on recommendations. Usually, such recommendation ap- proaches consider different aspects like discovering the associations of different items and users [8], mining the users’ profile and behavior logs [8], and the matching of item profiles [9]. Except online sellers, many web based applications use recommendation to push the potential interesting items to users when users are searching for something or viewing something. For example, www.youtube.com recommends related videos to users when they are viewing a video. http://news.sina.com.cn recommends related news when user is viewing a piece of news. Related item recommendation is usually implemented using the similarity calculation based on contents and metadata of the items [10]. 48 X. Wang, F. Yuan, and L. Qi 6 Conclusions In this paper, we have investigated the problem of recommendation in education portal. We have noticed the semantically explicit relations between entities in the education portal and used them to get better results of recommendation. We have formalized the recommendation as an importance ranking problem. By using the approach based on random walk, we have calculated the importance rank of each entity in the education portal. Finally, we have evaluated the results of our approach in real world data set. Future works can be summarized as follows: 1. We will apply this approach into the real education portal inside Tsinghua University, which is http://info.tsinghua.edu.cn, this web site is managed by our department and we are now working on adding this recommendation to it. The scenario could be as follows: When a student logins into Tsinghua’s edu- cation portal, the most important and useful courseware for the student will be recommended and displayed explicitly. 2. We will use this approach in more complex recommendation scenario. For example, to combine this approach in recommendation of entity query. In en- tity query, the importance rank will be combined with the content relevance to do recommendation. References 1. Michael, P., Daniel, B.: Content-Based recommendation systems. The Adaptive Web, 325–341 (2007) 2. Kyoung-jae, K., Hyunchul, A.: A recommendation system using GA K-means clustering in an online shopping market. Expert Systems with Applications 34(2), 1200–1209 3. Lin, W., Alvarez, S.A., Rujz, C.: Collaborative recommendation via adaptive association rule mining. In: WebKDD workshop (2000) 4. Zan, H., Wingyan, C., Hsinchun, C.: A graph model for e-commerce recommender sys- tems. Journal of the American Society for Information Science and Technology 55(3), 259–274 5. Sergey, B., Lawrence, P.: The anatomy of a large-scale hypertextual web search engine. Computer Networks and ISDN Systems 30(7), 107–117 6. Barber, M.N., Ninham, B.W.: Random and restricted walks. Gordon and Breach, New York (1970) 7. IEEE explorer, http://ieeexplore.ieee.org/ 8. Linden, G., Smith, B., York, J.: Amazon.com recommendations: item-to-item collabora- tive filtering. Internet Computing 7(1), 76–80 9. Markus, Z., Markus, J., Dietmar, J., Sergiu, G.: Comparing recommendation strategies in a commercial context. IEEE Intelligent Systems 22(3), 69–73 10. Resnick, P., Iacovou, N., Sushak, M., Bergstrom, P., Riedl, J.: GroupLens: An open archi- tecture for collaborative filtering of netnews. In: Proceedings of the 1994 Computer Sup- ported Collaborative Work Conference (1994) F. Li et al. (Eds.): ICWL 2008, LNCS 5145, pp. 49–58, 2008. © Springer-Verlag Berlin Heidelberg 2008 Research on Learning Resources Organization Model Qingtang Liu 1 and Zhimei Sun 2, * 1 Huangzhong Normal University, Wuhan, China, 430079 Liuqtang@mail.ccnu.edu.cn 2 Law and Business College of Hubei University of Economics, Wuhan, China, 430079 ccnuszm@126.com Abstract. The paper studies the adaptive learning content organization model and algorithm, which takes knowledge point as a unit. At first, it defines granu- larity of learning object, designs learning content level framework, which takes knowledge point as the meaning unit, builds up the reusable and sharable foun- dation of learning resources. The second is to provide the basic scheduling model for learning content organization. The model involves in pre-test, survey, knowledge point scheduling modules, summary, and post-test. Learning content scheduling algorithm is designed to choose learning content according to the learning content sequence, learner’s styles, learning evaluating outcomes and application context. Lastly, the scheduling algorithm is proved effectiveness and stable in E-learning Service Platform. Keywords: learning objects, Organization model, scheduling algorithm. 1 Introduction As an opening learning form, e-learning has broken the time and space limit in tradi- tional education, enables the learner to carry on the effective active learning. Research on control, management and evaluation of learning process, learning and teaching model, adaptive learning system etc. have become the hotspot in e-learning. The adaptive schedule of learning content has become the key of active, personalized, effective learning. Now, the learning content scheduling algorithm mainly focuses on subjective characteristic, learning content characteristic, learning strategy and intelli- gent decision-making. Along with the development of information technology, some problems come out as follows. The first problem is granularity of learning object [1]. In order to reuse and share learning resources, many e-learning technology standards organizations, such as IEEE, IMS[2], have formulated the correlative standards in abundance, in which they show the freedom definition for the granularity of learning object as a knowledge point, or a unit, or a curriculum. However, the effective scheduling and combination * Corresponding author. . basic scheduling model for learning content organization. The model involves in pre-test, survey, knowledge point scheduling modules, summary, and post-test. Learning content scheduling algorithm. effectiveness and stable in E-learning Service Platform. Keywords: learning objects, Organization model, scheduling algorithm. 1 Introduction As an opening learning form, e-learning has broken the. potential interesting items to users when users are searching for something or viewing something. For example, www.youtube.com recommends related videos to users when they are viewing a video.