Improvement in TF-IDF Scheme for Web Pages Based on the Contents of Their Hyperlinked Neighboring Pages

Transcription

1 Systems and Comuters in Jaan, Vol. 36, No. 14, 2005 Translated from Denshi Joho Tsushin Gakkai Ronbunshi, Vol. J87-D-II, No. 2, February 2004, Imrovement in TF-IDF Scheme for Web Pages Based on the Contents of Their Hyerlinked Neighboring Pages Kazunari Sugiyama, 1 Kenji Hatano, 1 Masatoshi Yoshikawa, 2 and Shunsuke Uemura 1 1 Graduate School of Information Science, Nara Institute of Science and Technology, Ikoma, Jaan 2 Information Technology Center, Nagoya University, Nagoya, Jaan SUMMARY The TF-IDF scheme is widely used to characterize documents in an information retrieval (IR) system based on the vector sace model. However, for documents having a hyerlink structure such as Web ages, the Web age contents can be characterized more accurately by using the contents of hyerlinked neighboring ages. Therefore, in this aer, we roose several techniques for using the contents of hyerlinked neighboring ages to imrove the TF-IDF scheme for Web ages and then verify the effectiveness of our techniques Wiley Periodicals, Inc. Syst Com Jn, 36(14): 56 68, 2005; Published online in Wiley InterScience ( DOI /scj Key words: WWW; information retrieval; TF-IDF scheme; hyerlink. 1. Introduction The World Wide Web (WWW) is a useful resource for users to obtain a great variety of information. Today, search engines store more than eight billion Web ages [1], and that number will clearly continue to grow in the future. Therefore, it will be increasingly more difficult for users to find relevant information on the WWW. Under these circumstances, Web search engines are one of the most oular methods for finding valuable information efficiently, and they are classified into two generations based on their indexing techniques [2]. A Web age is a semistructured document that can be reresented by a tree structure. Using this characteristic, in the first-generation search engines develoed in the early stages of the Web, only words surrounded by title tags that existed in a ortion near the root of a Web age were exloited as the indices of that age. Therefore, with this kind of characterization scheme, users were not satisfied with their retrieval accuracy. To deal with this roblem, in the second-generation search engines the hyerlink structures of Web ages are taken into consideration. For examle, PageRank [3] is the algorithm alied to the search engine Google (htt:// com), and HITS (Hyertext Induced Toic Search) [5] is the algorithm alied to the search engine in the CLEVER roject [6]. In these algorithms, weighting Web ages based on hyerlink structures achieves higher retrieval accuracy comared with the first-generation search engines. However, these algorithms have shortcomings that (1) the weight for a Web age is merely defined and (2) the relativity of contents among hyerlinked Web ages is not taken into consideration. Consequently, there still remains the roblem that Web ages irrelevant to a user s query are often ranked highly. Taking these oints into account, in order to rovide users with Web ages relevant to a user s query, it is necessary to develo a technique for reresenting the contents of Web ages more accurately. Therefore, feature vectors Wiley Periodicals, Inc.

2 should be calculated so that both Web ages linked to the target Web age (in-linked ages) and from the target Web age (out-linked ages) are taken into consideration. In this aer, we first roose several aroaches to imroving feature vectors of a target Web age created in advance based on the TF-IDF scheme [7] using both its in- and out-linked ages. Then we verify the effectiveness of our roosed aroaches. Comared with the second-generation search engines, our aroach is novel in characterizing Web ages more accurately by reflecting the contents of their hyerlinked neighboring Web ages. The rest of this aer is organized as follows. In Section 2, we review related work using hyerlink structures of the WWW. In Section 3, we roose novel methods for imroving feature vectors of Web age by using their hyerlinked neighboring ages. In Section 4, we resent the exerimental results for evaluating our roosed methods and discuss those results. In Section 5, we conclude the aer with a summary and directions for future work. 2. Related Work Hyerlink structures are one of the features of the WWW. Users can navigate the huge Web sace easily through the hyerlink structures; therefore, a great deal of research related to Web information retrieval has focused on the hyerlink structure of the WWW. In this section, we review related work of IR systems using the hyerlink structure of the WWW, esecially IR systems based on the concet of otimal document granularity and the two most oular Web age weighting algorithms, HITS and PageRank Information retrieval systems based on the concet of otimal document granularity With resect to this research area, we refer to the following works. Tajima and colleagues [8] roosed a technique that uses a concet called cuts (results of Web structure analysis) as retrieval units for the WWW. Moreover, they extended to rank search results that include multile keywords by (1) finding minimal subgrahs of links and Web ages including all keywords from the hyerlink structure of the WWW and (2) calculating similarities for keywords related to each minimal subgrah based on the locality of the keywords within it [9]. Following these works, Li and colleagues [10] introduced the concet of an information unit as one minimal retrieval unit for a document that consists of multile Web ages and roosed a novel framework for Web age retrieval based on these information units. However, these aroaches require considerable rocessing time to analyze hyerlink structures and to discover the semantics of Web ages. In addition, it is often the case that they find retrieval units that are irrelevant to the query terms secified by the user. As for these works, we do not believe that users could understand the search results intuitively, since the multile query keywords are scattered among several hyerlinked Web ages HITS algorithm The HITS algorithm [5] is alied to the search engine in the CLEVER roject [6]. This algorithm deends on the query, and considers the set of ages S that oint to or are redirected to, ages in the answer. In S, ages that have many links ointing to them are called authorities, while Web ages that have many outgoing links are called hubs. In other words, better authorities come from incoming edges from good hubs, and better hubs come from outgoing edges to good authorities. Let H() and A() be the hub and authority score of Web age, resectively. These scores are defined so that the following equations are satisfied for all ages : where H() and A() are normalized for all ages. These scores can be determined through an iterative algorithm, and they converge to the rincial eigenvector of the link matrix of S. However, the major roblem in this algorithm is that the Web ages that the root document oints to get the largest authority scores because the hub score of the root age dominates all the others when a Web age has few in-links but a large number of out-links. In order to resolve this roblem, several extended HITS algorithms have also been roosed [11 15] PageRank algorithm The PageRank algorithm simulates a user navigating randomly in the Web who jums to a random age with robability d or follows a random hyerlink with robability 1 d. It is further assumed that this user never returns to a reviously visited age following an already traversed hyerlink backwards. This rocess can be modeled with a Markov chain, from which the stationary robability of being in each age can be comuted. This value is then used as a art of the ranking mechanism used by Google. Let C(a) be the number of outgoing links from Web age a, and suose that Web ages 1 to n oint to Web age a. Then, the PageRank PR(a) of a is defined as 57

3 where the value of d is emirically set to about 0.15 to 0.2 by the system. The weights of other Web ages are normalized by the number of links in the Web age. PageRank can be comuted using an iterative algorithm, and corresonds to the rincial eigenvector of the normalized link matrix of the Web. The major roblems of this algorithm are that (1) the contents of Web ages are not analyzed, so the imortance of a given Web age is indeendent of the query, and (2) secific famous Web sites tend to be ranked more highly. Therefore, to yield more accurate search results, some algorithms that extend this algorithm have been roosed [16, 17]. 3. Proosed Method As we described in Section 2.1, the IR systems based on the concet of otimal document granularity have a roblem, in that the search results are incomrehensible for users. Moreover, HITS and PageRank algorithms also have roblems: (1) the weights for a Web age are merely defined and (2) the relativity of contents among hyerlinked Web ages is not considered. Based on these roblems, we believe that the feature vectors of Web ages should be generated by reflecting the contents of hyerlinked neighboring ages in order to reresent the contents of Web ages more accurately. Therefore, in this aer, we roose a method for imroving TF-IDF based feature vector of a target Web age by reflecting the contents of its hyerlinked neighboring Web ages. In the following discussion, let a target Web age be tgt. Then we define i as the length of the shortest directed ath from tgt. Let us assume that there are N i Web ages ( i1, i2,..., ini ) in the i-th level from tgt. Moreover, we denote the feature vector w tgt of tgt as follows: where m denotes the number of unique terms in the Web age collection, and t k (k = 1, 2,..., m) denotes each term. Using the TF-IDF scheme, we also define each element w tgt tk of w tgt as follows: (1) (2) and refer to this w g tgt as the imroved feature vector. In the following, we also define the distance dis(a, b) between two m-dimensional vectors a = (a 1, a 2,..., a m ) and b = (b 1, b 2,..., b m ) as In this aer, we roose three methods for imroving the initial feature vector w tgt defined by Eq. (1) based on the TF-IDF scheme defined by Eq. (2). In the following, the forward direction means navigating to a age that is linked from tgt [out-linked age; downward from tgt in Figs. 1(a), 2(a), and 3(a)], and the backward direction means navigating to a age that is linked to tgt [in-linked age; uward from tgt in Figs. 1(a), 2(a), and 3(a)] Method I In this method, we reflect the feature vectors of all Web ages at levels u to L (in) -th in the backward direction and levels u to L (out) -th in the forward direction from the target age tgt on the feature vector of target age tgt. Based on the ideas that (1) there are Web ages similar to the contents of tgt in the neighborhood of tgt, and (2) since on the one hand such Web ages exist right near tgt, on the other hand they might also exist far removed from tgt in the vector sace, we generate the imroved feature vector w g tgt by reflecting the distance between the initial feature vector w tgt and the feature vectors of the in- and out-linked ages of tgt in the vector sace on each element of initial feature vector w tgt. For examle, Fig. 1(a) shows that w g tgt is generated by reflecting the feature vectors of all Web ages at levels u to the second in the backward and forward directions from tgt on w tgt. In Fig. 1(a), ij(in) and ij(out) corresond to the j-th ages in the i-th levels in the backward and forward directions from tgt, resectively. In addition, Fig. 1(b) shows that imroved feature vector w g tgt is generated by reflecting each feature vector of in- and out-linked ages on the initial feature vector w tgt. In this method, each g element w tgt tk of w g tgt is defined by reliminary exeriments as follows: (4) where tf(t k, tgt ) is the frequency of term t k in the target Web age tgt, N web is the total number of Web ages in the collection, and df(t k ) is the number of Web ages in which term t k aears. Below, we refer to w tgt as the initial feature vector. Subsequently, we denote the imroved feature vector w g tgt of w tgt as follows: (3) (5) 58

4 Fig. 1. The imrovement of a feature vector as erformed by Method I [(a) in the Web sace, (b) in the vector sace]. Equation (5) shows that the roduct of w tgt tk (weight of term t k in Web age ij(in) ) and the recirocal of dis(w tgt, w ij(in)) (the distance between w tgt and w ij (in) in the vector sace), and similarly, the roduct of w ij(out) tk (weight of term t k in Web age ij(out) ) and the recirocal of dis(w tgt, w ij(out)) (the distance between w tgt and w ij (out) in the vector sace) is added to w tgt tk [weight of term t k in tgt, comuted by TF-IDF scheme defined by Eq. (2)] with resect to all Web ages at levels u to L (in) -th in the backward direction and levels u to L (out) -th in the forward direction from tgt. Dim, which denotes the number of unique terms, is introduced to revent the second and third terms of Eq. (5) from being dominant over the weight w tgt tk of the original index term Method II In this method, we first construct Web age grous G i(in) at each level u to L (in) -th in the backward direction, and G i(out) at each level u to L (out) -th in the forward direction from target age tgt. Then, we generate imroved feature vector w g tgt by reflecting centroid vectors of the clusters generated from G i(in) and G i(out) on the initial feature vector Fig. 2. The imrovement of a feature vector as erformed by Method II [(a) in the Web sace, (b) in the vector sace]. 59

5 Fig. 3. The imrovement of a feature vector as erformed by Method III [(a) in the Web sace, (b) in the vector sace]. w tgṭ This method is based on the idea that Web ages at each level in the backward and forward directions from tgt are classified into several toics in each level. In addition, we reflect the distance between w tgt and the centroid vectors of the clusters in the vector sace on each element of the initial feature vector w tgt. In other words, we first construct Web age grous G i(in) and G i(out), which are defined by Then, we roduce K clusters in each Web age grou G i(in) and G i(out) by means of the K-means algorithm [18]. The centroid vectors w g ic (in) and w g ic (out) (c = 1, 2,..., K) are roduced in G i(in) and G i(out), resectively. We generate the imroved feature vector w g tgt by reflecting the distance between each centroid vectors w g ic (in) and w g ic (out) and the initial feature vector w tgt on each element of w tgt. For instance, Fig. 2(a) shows that we construct Web age grous G 1(in), G 2(in), G 1(out), and G 2(out) at each level u to the second in the backward and forward directions from tgt, and generate an imroved feature vector w g tgt by reflecting the centroid vectors of each cluster roduced in each of these Web age grous on initial feature vector w tgt. In addition, Fig. 2(b) shows that imroved feature vector w g tgt is generated by reflecting the centroid vectors of each cluster on initial feature vector w tgt. In this method, g each element w tgt tk of w g tgt is defined by reliminary exeriments as follows: (6) (7) g Equation (8) shows that the roduct of w ic(in) tk (weight of term t k in centroid vector w g ic (in) of cluster c constructed from G i(in) ) and the recirocal of the distance dis(w tgt, w g ic (in)) (the distance between w tgt and w g ic (in) in the vector sace), and g similarly, the roduct of w ic(out) tk (weight of term t k in centroid vector w g ic (out) of cluster c constructed from G i(out) ) and the recirocal of the distance dis(w tgt, w g ic (out)) (the distance between w tgt and w g ic (out) in the vector sace) are added to w tgt tk (weight of term t k in tgt, comuted by TF-IDF scheme defined by Eq. (2)] with resect to all centroid vectors constructed at each level u to L (in) -th in the backward direction and each level u to L (out) -th in the forward direction from tgt. In addition, Dim, which denotes the number of unique terms, is introduced to revent the second and third terms of Eq. (8) from being dominant over the weight w tgt tk of the original index term Method III This method is based on the idea that Web ages at levels u to L (in) -th in the backward direction and levels u (8) 60

6 to L (out) -th in the forward direction from target age tgt are comosed of several toics. According to this idea, we cluster the set of all Web ages at levels u to L (in) -th in the backward direction and levels u to L (out) -th in the forward direction from tgt, and then generate the imroved feature vector w g tgt by reflecting centroid vectors of the clusters on initial feature vector w tgt. Furthermore, we reflect the distance between w tgt and the centroid vector of the cluster in the vector sace on each element of w tgt. In other words, we first create Web age grous G i(in) and G i(out) defined by G i(out) ) and the recirocal of dis(w tgt, w g c (out)) (the distance between w tgt and w g c (out) in the vector sace) are added to w tgt tk [weight of term t k in tgt, comuted by TF-IDF scheme defined by Eq. (2)], with resect to the number of clusters K. As mentioned in Methods I and II, in order to revent the value of the second and third terms of Eq. (11) from being dominant comared with the original term weight w tgt tk, we introduce Dim, which denotes the number of unique terms. 4. Exeriments 4.1. Exerimental setu Then, we roduce K clusters to create in G i(in) and G i(out) by means of the K-means algorithm. The centroid vectors w g c (in) and w g c (out) (c = 1, 2,..., K) are roduced in G i(in) and G i(out), resectively. Then we generate the imroved feature vector w g tgt by reflecting the distance between each centroid vector w g c (in) and w g c (out) (c = 1, 2,..., K) and initial feature vector w tgt on each element of w tgt. For examle, Fig. 3(a) shows that we construct Web age grous G 2(in) and G 2(out) at levels u to second in the backward and forward directions from tgt, and generate imroved feature vector w g tgt by reflecting the centroid vectors of clusters roduced in these Web age grous on initial feature vector w tgt. Furthermore, Fig. 3(b) shows that imroved feature vector w g tgt is generated by reflecting centroid vectors of each cluster on initial feature vector w g tgt. In this method, each element w tgt tk of w g tgt is defined by reliminary exeriments as follows: (9) (10) (11) We imlemented the three methods described in Section 3 using Perl on a workstation (CPU: UltraSarc-II 480 MHz 4; memory: 2 Gbytes; OS: Solaris 8) and conducted exeriments in order to verify the retrieval accuracy of each method using the TREC (text retrieval conference) WT10g test collection [19]. This test collection was created by Commonwealth Scientific & Industrial Research Organization (CSIRO; htt:// in Australia in 1997 using a ortion of the Web ages collected by the Internet Archive (htt:// This collection consists of aroximately 1.69 million Web ages (10 Gbytes) and information on in- and out-linked ages for each Web age in the test collection, sets of query toics, and set of relevant documents. The set of query toics was created for 50 toics based on the query log of Excite (htt:// and is comosed of a title field describing the search request, descrition field describing a sentence that satisfies the queries, and a narrative field describing the criteria for judging relevant documents, as shown in Fig. 4. Sto words were eliminated from all Web ages in the collection based on the stoword list (ft://ft.cs.cornell.edu/ub/smart/english.sto) and stemming was erformed using the Porter Stemmer [20] (htt:// We formulated query vector reresented as shown in Eq. (12) using the terms included in the title field for each of the 50 sets of query toics. Figure 4 shows one of the examles of a query toic. (12) In Eq. (12), t k (k = 1, 2,..., m) denotes an index term, and each element q tk is defined by g Equation (11) shows that the roduct of the weight w c(in) tk (weight of term t k in centroid vector w g c (in) of cluster c generated from G i(in) ) and the recirocal of dis(w tgt, w g c (in)) (the distance between w tgt and w g c (in) in the vector sace), g and similarly the roduct of element w c(out) tk (weight of term t k in centroid vector w g c (out) of cluster c generated from (13) where Qf(t k ), N web, and df(t k ) denote the number of index terms t k contained in query vector Q, the total number of 61

7 imroved feature vector w g tgt for initial feature vector w tgt of target age tgt with resect to the following cases: [Method I] (a) where the contents of all Web ages at levels u to L (in) -th in the backward direction from tgt reflect on initial feature vector w tgt (b) where the contents of all Web ages at levels u to L (out) -th in the forward direction from tgt reflect on initial feature vector w tgt (c) where the contents of all Web ages at levels both u to L (in) -th in the backward direction and u to L (out) -th in the forward direction from tgt reflect on initial feature vector w tgt. Fig. 4. An examle of toic descritions in the WT10g test collection. Web ages in the test collection, and the number of Web ages in which term t k aears, resectively. As reorted in Ref. 21, Eq. (13) is the element of a query vector that brings the best retrieval accuracy. We then comute the similarity sim(w g tgt, Q) between the imroved feature vector w g tgt described in Section 3 and query vector Q using the equation We evaluate our exerimental results in Section 4.3 using average recision based on the value of sim(w g tgt _, Q). In addition, we comute average recision P based on the equation where q i is the i-th (i = 1, 2,..., N q ) query, R qi is the number of relevant documents for q i, and Rel qi is the number of the to R qi relevant documents that system returns. In this aer, with regard to the set of 50 retrieval tasks (N q = 50), we aly Eq. (15) to evaluate the to 1000 Web ages that our roosed methods return based on the value of sim(w g tgt, Q) Exerimental methods (14) (15) In order to verify the effectiveness of the three roosed methods described in Section 3, we generated the [Method II] (a) where the centroid vectors of clusters generated by the grou of Web ages at each level u to L (out) -th in the backward direction from tgt reflect on initial feature vector w tgt (b) where the centroid vectors of clusters generated by the grou of Web ages at each level u to L (out) -th in the forward direction from tgt reflect on initial feature vector w tgt (c) where the centroid vectors of clusters generated by the grou of Web ages at each level both u to L (in) -th in the backward direction and u to L (out) -th in the forward direction from tgt reflect on initial feature vector w tgt. [Method III] (a) where the centroid vectors of clusters generated by the grou of all Web ages at levels u to L (in) -th in the backward direction from tgt reflect on initial feature vector w tgt (b) where the centroid vectors of clusters generated by the grou of all Web ages at levels u to L (out) -th in the forward direction from tgt reflect on initial feature vector w tgt (c) where the centroid vectors of clusters generated by the grou of all Web ages at levels u to L (in) -th in the backward direction and levels u to L (out) -th in the forward direction from tgt reflect on initial feature vector w tgt. We generated the imroved feature vector w g tgt and conducted exeriments to comare retrieval accuracies of our roosed methods by varying L (in) and L (out) for Methods I, II, and III so that 1 L (in) 5 for (a), 1 L (out) 5 for (b), and 1 L (in), L (out) 5 for (c) and also by varying the number of clusters K for Methods II and III so that 1 K Exerimental results and discussion Figure 5 illustrates the average recision when the values of L (in), L (out), or [L (in), L (out) ] vary in Method I (a), 62

8 Fig. 5. Average recision based on Method I. Fig. 7. Average recision based on Method II(a). (b), or (c), resectively. Figures 7 to 9 illustrate the average recision when the number of clusters K varies in Method II (a), (b), and (c), resectively. Figures 10 to 12 illustrate the average recision when the number of clusters K varies in Method III (a), (b), and (c), resectively. In order to facilitate the comarison of the average recision obtained by our roosed methods and TF-IDF scheme, we also show the retrieval accuracy obtained by our roosed methods and TF-IDF scheme in each figure. The recision of the TF-IDF scheme does not deend on the values of L (in), L (out), and the number of clusters K. Therefore, the retrieval accuracy obtained by the TF-IDF scheme is fixed for these values. We can observe the following findings in each method [22]. In Method I, as shown in Fig. 5, although the similarity between the target age tgt and Web ages at levels u to third (L (in) 3) in the backward direction from tgt and Web ages at levels u to third (L (out) 3) in the Fig. 8. Average recision based on Method II(b). Fig. 6. Distribution of average similarity. Fig. 9. Average recision based on Method II(c). 63

9 Fig. 10. Average recision based on Method III(a). Fig. 11. Average recision based on Method III(b). Fig. 12. Average recision based on Method III(c). forward direction from tgt is high, these neighboring Web ages of tgt contribute in reresenting the contents of tgt more accurately. However, the similarity between the target age tgt and Web ages from fourth (L (in) 4) level in the backward direction from tgt and Web ages from fourth (L (out) 4) level in the forward direction from tgt is low; therefore we cannot observe the effect of reresenting the contents of tgt more accurately. Here, in Fig. 6, we show the distribution of the average similarity between Web ages in the WT10g test collection and Web ages in the backward and forward directions from each Web age in the collection. The average similarity between a Web age in the collection and Web ages at levels u to third (L (in) 3) in the backward direction from, and Web ages at levels u to third (L (out) 3) in the forward direction from is relatively high. However, the average similarity between a Web age in the collection and Web ages from fourth (L (in) 4) level in the backward direction from and Web ages from fourth (L (out) 4) level in the forward direction from is low. We consider that these facts affect the average recision. In addition, as Fig. 5 shows, in the case of L (in) 4 or L (out) 4, the average recision tends to decline. Therefore, it is aroriate that we examined the average recision in the range of 1 L (in) 5 or 1 L (out) 5. In Method II, as shown in Figs. 7 to 9, the farther the distance from target Web age tgt, in other words, the larger the values of L (in) or L (out) are, the smaller the ga between the grah of the average recision obtained by our roosed methods and the grah of the average recision obtained by the TF-IDF scheme. In other words, the degree of imrovement in retrieval accuracy is small comared with the TF-IDF scheme. Thus, we found that with regard to the contents of Web ages, there is strong similarity between the feature vector of target age tgt, and the centroid vector generated by the Web age grous at each level u to first in the backward (L (in) = 1) and forward (L (out) = 1) directions from tgt. However, we also found that similarity between the feature vector of tgt and the centroid vector generated by the grou of Web ages at each level from tgt decreases as the value of i, which denotes the length of the shortest directed ath from tgt to its hyerlinked neighboring ages, increases. In Method III, as shown in Figs. 10 to 12, the best retrieval accuracy is obtained when we generate imroved feature vectors using each centroid vector of three clusters (K = 3) generated in a Web age grou roduced using all Web ages at levels u to second (L (in) 2, L (out) 2) in the backward and forward directions from the target age tgt. Since the best retrieval accuracy is obtained when the number of clusters is three (K = 3) in this method, we can infer that the toics of Web ages at levels u to second (L (in) 2, L (out) 2) in the backward and forward directions from tgt are usually comosed of three toics. 64

10 Furthermore, we found the following relations between the number of clusters in Methods II and III. First, in Method II, as shown in Figs. 7 to 9, we observed that the average recision tends to decrease in the case of K 3. On the other hand, in Method III, we obtain the same results as those of Method II, in the case of L (in) = 1 or L (out) = 1, as shown in Eqs. (8) and (11). However, in the case of L (in) 2 or L (out) 2, we observed that the average recision decreases gradually, when the number of clusters K is greater than 4. Therefore, we consider that it is valid that we examined the average recision in the range of 1 K 5 in Methods II and III. Table 1, which summarizes the results described above, illustrates the average recision when we generated the feature vector of Web age using TF-IDF scheme and the best average recision when we generated the feature vector of Web age using each of our roosed methods. In Method I, the best retrieval accuracy is obtained when we generate the feature vector of Web age by utilizing the contents of all Web ages at levels u to third (L (in) = 3) in the backward direction from the target age tgt. In Method II, the best retrieval accuracy is obtained when we generate imroved feature vectors using each centroid vector of two clusters (K = 2) generated in a Web age grou roduced using all Web ages at levels u to first (L (in) = 1) in the backward direction from the target age tgt. Moreover, in Method III, we obtain the best retrieval accuracy when we generate imroved feature vectors using each centroid vector of three clusters (K = 3) generated in a Web age grou roduced using all Web ages at levels u to second (L (in) = 2) in the backward direction from the target age tgt. Furthermore, as shown in Table 1, in any case of Methods I, II, and III, the best retrieval accuracy is obtained in exeriment (a), namely, in the case of using the contents of in-linked ages of a target age. We consider that this finding is obtained because we can easily reach Web ages similar to the target Web age when we follow the hyerlinks in the backward direction from the target age while we reach various Web ages that are not so similar to the Table 1. Comarison of the otimal search accuracy that was obtained using Methods I, II, and III % average recision % imrovement TF-IDF Method I (L (in) = 3) Method II (L (in) = 1, K = 2) Method III (L (in) = 2, K = 3) Table 2. Average recision of WT10g by using link information Grou Outline of each aroach contents of the target age when we follow the hyerlinks in the forward direction from the target age. In other words, Web ages have a characteristic that the in-linked ages of a target Web age have Web ages relevant to the contents of the target age. As described in Section 2.2, the HITS algorithm defines Web ages that have many outgoing links as hubs, and also defines the quality of a Web age as authority by considering the hubs as its in-linked ages of authority. In addition, focusing on the imortance of in-linked ages, the tool that enables navigation in the backward direction from a target age is also develoed [23]. We can consider that the results in Table 1 suggest the usefulness of in-linked ages even in our study Discussion related to search accuracy % average recision Ref. 27 Anchor text Anchor text + long query Ref. 26 Content-link gram content-link Ref. 24 Cocitation to Cocitation to HITS 4.88 Ref. 28 Okai + robabilistic augmentation Ref. 29 Anchor text Variant of anchor text Ref. 30 Back link frequency Ref. 25 Modified HITS 5.91 Modified HITS with weighted links 6.37 According to Ref. 19, in the TREC-9 Web Track, the outline of methods using link information by articiants and average recision are shown in Table 2. The reader is referred to the aers in Table 2 for their detailed methods. It is seen that the average recision based on HITS in Ref. 24 is 4.88%, and the average recisions based on modified HITS in Ref. 25 are 5.91 and 6.37%. These are oor results. On the other hand, as shown in Table 1, our best average recision is 16.23%. Since this result is comarable to the results of Refs. 26 and 24, we believe that our roosed methods are effective enough to characterize Web ages more accurately and the results obtained by using them are sufficiently solid. 65

11 5. Conclusions In this aer, in order to reresent the contents of Web ages more accurately, we roosed three methods for imroving the TF-IDF scheme for Web ages using the feature vectors of hyerlinked neighboring ages. Our method is innovative in imroving the TF-IDF based feature vector of a target Web age by reflecting the contents of its hyerlinked neighboring Web ages, which had not been done u to now. Naturally, if our scheme is used together with HITS or PageRank, which had already been roosed, a further imrovement in retrieval accuracy can be exected. In this aer, we focused on the hyerlink structures of the Web aiming at generating more accurate feature vectors of Web ages. However, in order to satisfy the user s actual information need, it is more imortant to find relevant Web ages from the enormous WWW sace. Therefore, we lan to address the technique to rovide users with ersonalized information. REFERENCES 1. Baeza-Yates R, Saint-Jean F, Castillo C. Web structure, dynamics and age quality. Proc 9th International Symosium on String Processing and Information Retrieval (SPIRE 2002), Broder A, Raghavan P. Combining text- and linkbased information retrieval on the web. SIGIR 01 Pre-Conference Tutorials. 3. Page L. The PageRank citation ranking: Bringing order to the web. htt://google.stanford.edu/ ~ backrub/ ageranksub.s, Brin S, Page L. The anatomy of a large-scale hyertextual web search engine. Proc 7th International World Wide Web Conference (WWW7), , Kleinberg JM. Authoritative sources in a hyerlinked environment. Proc 9th Annual ACM-SIAM Symosium on Discrete Algorithms (SODA 1998), IBM Almaden Research Center. Clever Searching. htt:// 7. Salton G, McGill MJ. Introduction to modern information retrieval. McGraw Hill; Tajima K, Mizuuchi Y, Kitagawa M, Tanaka K. Cut as a querying unit for WWW, netnews, . Proc 9th ACM Conference on Hyertext and Hyermedia (Hyertext 98), Tajima K, Hatano K, Matsukura T, Sano R, Tanaka K. Discovery and retrieval of logical information units in web. Proc 1999 ACM Digital Libraries Worksho on Organizing Web Sace (WOWS 99), Li W-S, Selçuk Candan K, Vu Q, Agrawal D. Retrieving and organizing web ages by Information Unit. Proc 10th International World Wide Web Conference (WWW10), , Chakrabarti S, Dom B, Raghavan P, Rajagoalan S, Gibson D, Kleinberg J. Automatic resource comilation by analyzing hyerlink structure and associated text. Proc 7th International World Wide Web Conference (WWW7), 65 74, Bharat K, Henzinger MR. Imroved algorithms for toic distillation in a hyerlinked environment. Proc 21st Annual International ACM SIGIR Conference (SIGIR 98), Li L, Shang Y, Zhang W. Imrovement of HITS-based algorithms on web documents. Proc 11th International World Wide Web Conference (WWW2002), Chakrabarti S. Integrating the document object model with hyerlinks for enhanced toic distillation and information extraction. Proc 10th International World Wide Web Conference (WWW10), , Chakrabarti S, Joshi M, Tawde V. Enhanced toic distillation using text, marku tags, and hyerlinks. Proc 23rd Annual International ACM SIGIR Conference (SIGIR 2001), Rafiei D, Mendelzon AO. What is this age known for? Comuting web age reutations. Proc 9th International World Wide Web Conference (WWW9), , Haveliwala TH. Toic-sensitive PageRank. Proc 11th International World Wide Web Conference (WWW2002), MacQueen J. Some methods for classification and analysis of multivariate observations. Proc 5th Berkeley Symosium on Mathematical Statistics and Probability, , Hawking D. Overview of the TREC-9 web track. NIST Secial Publication : The Ninth Text REtrieval Conference (TREC-9), , Porter MF. An algorithm for suffix striing. Program 1988;14: Salton G, Buckley C. Term-weighting aroaches in automatic text retrieval. Inf Process Manage 1988;24: Sugiyama K, Hatano K, Yoshikawa M, Uemura S. Refinement of TF-IDF schemes for web ages using their hyerlinked neighboring ages. Proc 14th ACM Conference on HyerText and Hyermedia (HT 03),

12 23. Chakrabarti S, Gibson DA, McCurley KS. Surfing the web backwards. Proc 8th International World Wide Web Conference, , Kraaij W, Westerveld T. TNO/UT at TREC-9: How different are Web documents? htt://trec.nist.gov/ ubs/trec9/aers/tno-ut.df, Crivellari F, Melucci M. Web document retrieval using assage retrieval, connectivity information, and automatic link weighting-trec-9 reort. htt://trec.nist.gov/ubs/trec9/aers/jhual.df, Clake CLA, Cormack GV, Kisman DIE, Lynam TR. Question answering by assage selection (MultiText exeriments for TREC-9). htt://trec.nist.gov/ ubs/trec9/aers/mt9.df, Fujita S. Reflections on Aboutness TREC-9 evaluation exeriments at Justsystem. htt://trec.nist.gov/ ubs/trec9/aers/ jscbt9w_aer.df, Savoy J, Rasolofo Y. Reort on the TREC-9 exeriment: Link-based retrieval and distributed collections. htt://tree.nist.gov/ubs/trec9/aers/ unine9.df, Singhal A, Kaszkiel M. AT&T at TREC-9. htt://trec.nist.gov/ubs/trec9/aers/att-trec9.df, McNamee P, Mayfield J, Piatko C. The HAIRCUT system at TREC-9. htt://trec.nist.gov/ubs/trec9/ aers/jhual.df, AUTHORS (from left to right) Kazunari Sugiyama (student member) graduated in 1998 with a secialty in comuter science from the Deartment of Engineering at Yokohama National University, comleted the first half of his comuter science doctoral course in 2000, and joined KDD Cor. (currently, KDDI) where he worked until In 2004, he comleted the second half of his doctoral course at Nara Institute of Science and Technology, receiving a Ph.D. degree in engineering. He then joined Hitachi, Ltd., Software Division, and is engaged in research on information retrieval. He is a member of the Information Processing Society of Jaan, the Jaanese Society for Artificial Intelligence, the Association for Comuting Machinery, the American Association for Artificial Intelligence, and IEEE. Kenji Hatano (member) graduated in 1995 with a secialty in recision machinery from the Deartment of Engineering at Kobe University, comleted the second half of his doctoral course in 1999, and became an assistant in the Graduate School of Information Science at Nara Institute of Science and Technology. He is engaged in research on XML databases and information retrieval. He holds a Ph.D. degree in engineering, and is a member of the Information Processing Society of Jaan, the Association for Comuting Machinery, and the IEEE Comuter Society. Masatoshi Yoshikawa (member) graduated in 1980 with a secialty in information science from the Deartment of Engineering at Kyoto University, comleted the second half of his doctoral course in 1985, and became a lecturer at Kyoto Sangyo University. After serving as an assistant rofessor at Sangyo University and at Nara Institute of Science and Technology, he has been a rofessor in the Information Technology Center at Nagoya University since He is engaged in research on XML databases and multidimensional sace indexes. He holds a Ph.D. degree in engineering, and is a member of the Information Processing Society of Jaan, the Association for Comuting Machinery, and the IEEE Comuter Society. 67

13 AUTHORS (continued) Shunsuke Uemura (member, fellow) graduated in 1964 with a secialty in electronics from the Deartment of Engineering at Kyoto University, comleted his master s course in 1966, and joined the Electrical Testing Institute of the Ministry of International Trade and Industry (MITI) National Research Institutes (currently, the Industrial Technology Research Institute). In 1988, he became a rofessor of mathematics and comuter science in the Deartment of Engineering at Tokyo University of Agriculture and Technology. He has been a rofessor in the Graduate School of Information Science at Nara Institute of Science and Technology since In , he was a visiting researcher at Massachusetts Institute of Technology. He is engaged in research on database systems, natural language rocessing, and rogramming languages. He holds a Ph.D. degree in engineering, and is a fellow of IEEE and the Information Processing Society of Jaan and a member of the Association for Comuting Machinery. 68