X Tang Q Zeng/The Journal of Systems and Software 85(2012)87-101 aI. F: \LAB FORKS\Pajek Network model\dat abase 12)en, net (377) 区 ladom daab °d ° customers telatonthp managem°nds·te°aske● dynamic recoweryoene segment ●bee ●出 °ech●d一 tee eml heterogeneou &mobie comouiroowwallpaten● web pee techno erase block● abetone是ygpg● nand Hach bas bourdoubodegoup ma ches chicle ° informalonizahioninto quanity° beg mach componet·os t Hon ne analyted oceanog best gene ° ao smal abbe界 ne key dgal courte是 port vector mogod ertl not car● patial odered ter te mechaniam event chane ntercepors-med是 ohton ser是mnd ° coped dala°dt● semantic relatcriimage annolationregon tense frequent pellen high dimention● ste schema° dEabetes data va●s ● diluted dula ° debate wnct是okse. daa enorEe ane. privacy preserva dwa pupose DMacyewae是 treta dyad·shu·m embeded dyal 4°yee° solte house pro.20 managemetapesty°o点 detection°aps° hammers hidden makow mulidmensionated associatonlemovng emended poddntbuted reali edenstybased cueing agof path east squares suppeatid odered tempastial ordered tesfempord database gene"geredride Fig. 11. The weighted keyword graph of"databases"whose edge weight threshold value is 2. threshold value; the corresponding weighted keyword graph is pre- interest profiling algorithm. This experiment lasted for 30 days, and sented in Fig. 11. As shown in Fig. 10, there are 2 clusters(topics). 5 master students were engaged in it. These students were asked while the number of clusters is 7 in Fig. 11. Small newly emerged freely and naturally browse, download, comment on or collect the keyword clusters and their corresponding connectivities are pre research papers of the subjects of"databases"and"artificial intel ented in Table 2. Since the value 2 is proved to be a rather decent ligence"within SPRS. Because we mainly focused on the validity of alue of the edge weight threshold we then used it as the edge our profiling approach at the detailed classification level, we did not weight threshold. Thus, the seven topics plus the"heterogeneous" include other subjects. We also conducted a training process to find topic, constitute the new classifications of the secondary subject the best value of the parameter i in the new user interest profiling 'databases". In this way, the databases"subject of the original algorithm. During the training process, we found that using 2 as A ontology is extended shows the best performance of recommendation, and we therefor used2 as the value of入 6.2. 2. Experiment 2: user interest profile construction In order to make a comparison, we als Based on the extended ontology formed from the clustering ests using a traditional ontology-based profiling method. In both esults, user interest profiles were calculated according to our user methods, the parameter dp is set as 1, which equals the value of this 0.74 0.75 4 ■ Artificial Intelligen 0.25 21 sers Fig. 12. Computational results of the users' interests on the secondary subject level(explicit interests).X. Tang, Q. Zeng / The Journal of Systems and Software 85 (2012) 87–101 97 Fig. 11. The weighted keyword graph of “databases” whose edge weight threshold value is 2. threshold value; the corresponding weighted keyword graph is pre￾sented in Fig. 11. As shown in Fig. 10, there are 2 clusters (topics), while the number of clusters is 7 in Fig. 11. Small newly emerged keyword clusters and their corresponding connectivities are pre￾sented in Table 2. Since the value 2 is proved to be a rather decent value of the edge weight threshold, we then used it as the edge weight threshold. Thus, the seven topics plus the “heterogeneous” topic, constitute the new classifications of the secondary subject “databases”. In this way, the “databases” subject of the original ontology is extended. 6.2.2. Experiment 2: user interest profile construction Based on the extended ontology formed from the clustering results, user interest profiles were calculated according to our user interest profiling algorithm. This experiment lasted for 30 days, and 5 master students were engaged in it. These students were asked to freely and naturally browse, download, comment on or collect the research papers of the subjects of “databases” and “artificial intel￾ligence” within SPRS. Because we mainly focused on the validity of our profiling approach at the detailed classification level, we did not include other subjects. We also conducted a training process to find the best value of the parameter in the new user interest profiling algorithm. During the training process, we found that using 2 as shows the best performance of recommendation, and we therefore used 2 as the value of . In order to make a comparison, we also compute the user inter￾ests using a traditional ontology-based profiling method. In both methods, the parameter ˚ is set as 1, which equals the value of this Fig. 12. Computational results of the users’ interests on the secondary subject level (explicit interests).