X Tang, Q Zeng/The Joumal of Systems and Software 85(2012)87-101 82.F: 12len(无权重).net(3 ble touton men marooner are spicate temsimuirkont m patten theorem mage compreinte moble not man body. ever function°ay°a e wsidncrete wavelet transfomer elements d nevapeech emoton recogretionseindour n是pnd· conceplnet ladde chart● ef reco focused web clawing technclopar mage de-ndimwen ° voice recognior pust20°swmm是sdho°dsde° negatve. instaneous, setion on e executive tock ordona Didam de ● very greed ide&knapsack pn Paticle swam optimization9eham°shis· compositon dingus如 man visus systeme fao●pmgm是 the gorhen· nerve baric ●mcas· dala wenshou. stud dalaba是tma· atficin emotion modeition iverneb data mirPcemoonerh &relata·o●eod·e non-hneatly doey° contex-awa'·osd° service· colocation● fequency● ame recod· oopleyng levereduction mode puring methcdreedomiaion m speech emoti@teating methodweless senao anyang metfcompenion image selectioapeech emotortasb's caca modfied binomial ° data ming° decsion tee°m· waelet transfodata单mc01a· trng kemet° dynamic prog9mth是 x kene● vaal navision and orienta2是 maren w reverse eroneenng amann Fig. 5. A weighted keyword graph of the subject"artificial intelligence"(edge weight threshold equals 2). If we assign a different value to the threshold s2, different e The method of ontology extension through keyword clustering gences of TGS then will be produced During our experiment, in holds two advantages. It allows subject ontology to be automati which 200 papers were used, we discovered that, when 1 was used cally and sensitively adaptive to the changes of research topics in as the threshold of edge weight, TGS is as shown in Fig 4: when any subject. However scientific research topics change, whatever 3 was used as the threshold of edge weight, most of the keyword new research hotspots appear or whichever old research contents nodes in TGS became extremely discrete and only three keyword fade out, for example, the changes and the newest status will be clusters, within which very few keyword nodes existed maintained. immediately reflected in the new formation as the keyword cluster- The selection of the edge weight threshold is strongly pertinent to ing is executed. The ontology extension also makes the user interest ve use 2 as the threshold when clustering the weighted keyword are able to place a new angle of view with more accurate classifica- graph in Fig 4 and the result is showed in Fig. 5. In this weighted tions on all subjects: in this way, users' interests will be captured keyword graph, aside from the"heterogeneous"topic, there are ten and recorded even more clearly. In the next section, we will discuss other newly emerged topics which are listed in Table 1. Therefore the construction of user interest profiles the subordinate classifications of the secondary subjectartificial intelligence"are engendered. 5. Approaches to generating user profiles based on the So far, we have finished the extension of the secondary subject extended subject ontology artificial intelligence. All the secondary subjects in the original subject ontology are extended by the keyword clustering pro- We now use the extended subject ontology to create user pro- cess. An illustrative sketch of the result of the ontology extension files. In this section, we present our refined ontological profiling through keyword clustering is provided in Fig. 6. approach to solve the problems of the low accuracy of recommen- dation and of the coarse granularity of user interest profiles Topics of the secondary subject"artificial intelligence traditional ontology-based profiling algorithms. This user interest Topic nam Connectivity of central profiling approach is able to distinguish between the different con- keyword node tributions of the papers on the same topicto the construction ofuse neural network interest profiles. Also, apart from the user profile obtained directly rom the user behavior data, which we call the" explicit interest rofile, we apply implicit profiles to infer possible interests that users may develop in the future, in order to describe user interests expert system more roundly and thereby improve recommendation. A user inter est profile therefore consists of two parts: an explicit interest profile ansform and an implicit interest profile. An arbitrary user has an explicit Inversion set interest in a certain topic if the user directly accesses one or more traveling salesman problem papers in the topic. By contrast, an arbitrary users implicit interest92 X. Tang, Q. Zeng / The Journal of Systems and Software 85 (2012) 87–101 Fig. 5. A weighted keyword graph of the subject “artificial intelligence” (edge weight threshold equals 2). If we assign a different value to the threshold ˝, different emer￾gences of TGS then will be produced. During our experiment, in which 200 papers were used, we discovered that, when 1 was used as the threshold of edge weight, TGS is as shown in Fig. 4; when 3 was used as the threshold of edge weight, most of the keyword nodes in TGS became extremely discrete and only three keyword clusters, within which very few keyword nodes existed maintained. The selection of the edge weight threshold is strongly pertinent to the content and the quantity of the papers used in this process. Now we use 2 as the threshold when clustering the weighted keyword graph in Fig. 4 and the result is showed in Fig. 5. In this weighted keyword graph, aside from the “heterogeneous” topic, there are ten other newly emerged topics which are listed in Table 1. Therefore, the subordinate classifications of the secondary subject “artificial intelligence” are engendered. So far, we have finished the extension of the secondary subject “artificial intelligence”. All the secondary subjects in the original subject ontology are extended by the keyword clustering pro￾cess. An illustrative sketch of the result of the ontology extension through keyword clustering is provided in Fig. 6. Table 1 Topics of the secondary subject “artificial intelligence”. Topic name Connectivity of central keyword node neural network 23 agent 20 SVM 16 genetic algorithms 14 expert system 11 hybrid attributes 10 ontology 9 wavelet transform 8 inversion set 6 traveling salesman problem 5 The method of ontology extension through keyword clustering holds two advantages. It allows subject ontology to be automati￾cally and sensitively adaptive to the changes of research topics in any subject. However scientific research topics change, whatever new research hotspots appear or whichever old research contents fade out, for example, the changes and the newest status will be immediately reflected in the new formation as the keyword cluster￾ing is executed. The ontology extension also makes the user interest profiles more precise and distinct. With the clustering method, we are able to place a new angle of view with more accurate classifica￾tions on all subjects; in this way, users’ interests will be captured and recorded even more clearly. In the next section, we will discuss the construction of user interest profiles. 5. Approaches to generating user profiles based on the extended subject ontology We now use the extended subject ontology to create user pro- files. In this section, we present our refined ontological profiling approach to solve the problems of the low accuracy of recommen￾dation and of the coarse granularity of user interest profiles in traditional ontology-based profiling algorithms. This user interest profiling approach is able to distinguish between the different con￾tributions of the papers on the same topic to the construction of user interest profiles. Also, apart from the user profile obtained directly from the user behavior data, which we call the “explicit interest” profile, we apply implicit profiles to infer possible interests that users may develop in the future, in order to describe user interests more roundly and thereby improve recommendation. A user inter￾est profile therefore consists of two parts: an explicit interest profile and an implicit interest profile. An arbitrary user has an explicit interest in a certain topic if the user directly accesses one or more papers in the topic. By contrast, an arbitrary user’s implicit interest