JP2011216021A - Clustering device, clustering method and clustering program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve ambiguity of a tag, and to prevent an abrupt increase of the number of tags during reuse of the tags.SOLUTION: A dendrogram is constructed by executing hierarchical clustering to an entire tag group, and a bottom-up index in which an upper layer can be specified from a lower layer is generated in advance, and the entire tag group is clustered into a plurality of partial tag groups by referring to the generated index when a high-order application makes a request.

Description

  The present invention relates to a technique for reusing a classification axis assigned to resource information in a cooperative classification system.

  Recently, a variety of resource information such as URLs (bookmarks), photos, moving images, still images, books, and papers is classified and arranged by a third party different from the primary creator, and the results of the classification are widely shared. Collaborative / social tagging systems that can provide high-level information to viewers are prosperous.

  FIG. 15 is a diagram illustrating a classification example of classification axes (hereinafter referred to as tags) in the cooperative classification system. A URL is shown as an example of resource information. A plurality of tags “news”, “soccer”, and “interesting” related to the description content are generated by the third party classifier based on the description content “soccer news“ today's result ”” of the URL. .

  FIG. 16 is a diagram for explaining a flow of reusing tags generated by a classifier. The collaborative classification system allows a classifier to freely generate a tag and attach (associate) the generated tag to the URL, and stores it in the classification result storage database 12 for reuse. That is, in the process of collecting tags that can also be classified results of resource information by the classifier and generating the presentation information provided to the viewer, the tags given for the purpose of describing the description contents of the URL are classified result storage database 12 is read out and reused.

  On the other hand, problems such as ambiguity of tags and explosion of the number of tags have become apparent because tags can be freely generated by a third party and assigned to classification targets. The tag is ambiguous means that one tag has a plurality of meanings (ambiguity tag) and that the notation is different but the meaning is the same (synonymous tag). The number of tags due to such ambiguity, influences of third-party (ie, classifier) emotions and preferences at the time of tagging, trends in the world, trends in tags given by other third parties, etc. As a result of the explosion, it is difficult to reuse the tag (see Non-Patent Document 1).

  However, in spite of such difficulties, if a tag given by a third party is used, the meaning of the resource information is a tag unique to the third party viewpoint that is not necessarily included in the presentation information originally. Can be expressed accurately, and subjective tags that express impressions and opinions that are difficult to collect by auto-tagging (for example, “This is amazing”, “cool”, “read later”, etc.) Is often used to describe tagged information.

  For example, “flicker” (http://filckr.com) is a service that can tag photos, and “del.icio.us” (http://del.icio.us) and “Hatena Bookmark” ( http://b.hatena.ne.jp), “go bookmark” (http://bookmark.goo.ne.jp), and “youtube” (http://youtube.com) , "US Amazon" (http://www.amazon.com) as a service for tagging books, "CiteULike" (http://www.citeulike.org/) as a service for tagging academic papers, etc. Has been. In an actual service, a required number of tags are selected from the assigned tag set in order of frequency of use and are used to describe the presentation information.

  Here, the problem is how much number of tags should be used to generate the presentation information. In general, in product recommendation, there is a hypothesis that recommending more various products leads to improved user satisfaction. Based on the hypothesis, a method for diversifying the recommendation results has been proposed (Non-Patent Document 2). Also in web search and image search, a method of diversifying recommendation results based on similar hypotheses has also been proposed (see non-patent documents 3 and 4 for web search and non-patent documents 5 and 6 for image search). ). In either case, after recommending various products to the user, the recommended products are clustered on-demand based on the user's request (Post-Processing). In FIG. 17, as such a conventional method, after a request from the user, a distance of | m | or less from the request point r is calculated, and an area within the distance is divided in a specific direction to cluster products. The technique to do is shown. Note that clustering refers to making a whole set a part or all of a subset.

  That is, when the product is associated with tags, it can be said that it is a conventional technique to generate provided information using as many tags as possible.

Scott Golder, 1 other, "The Structure of Collaborative Tagging Systems", Journal of Information Science, 2006 Cai-Nicolas Ziegler, 3 others, “Improving recommendation lists through topic diversification”, Proc. WWW, 2005 Filip Radlinski, 1 other, “Improving personalized web search using result diversification”, Proc. SIGIR, 2006 Rakesh Agrawal, 1 other, "Diversifying search results", Proc WSDM, 2009 Kai Song, 3 others, “Diversifying the image retrieval results”, Proc. ACM Multimedia, 2006 Reinier H. van Leuken, 3 others, "Visual diversification of image search results", Proc. WWW, 2009 Toshihiro Kamisu, “Clustering Method for Data Mining Field (1)”, Journal of Population Intelligence Society, Vol. 18, No. 1, January 2003

  However, when many tags are used at the time of tag reuse, there are a plurality of tags including ambiguity as described above, and a plurality of tags representing similar contents may be collected and displayed. , There is a problem that the satisfaction of the viewer is lowered and the reusability of the tag is lowered.

  The present invention has been made in view of the above problems, and it is an object of the present invention to eliminate tag ambiguity and prevent explosion of the number of tags during tag reuse.

  The present invention according to claim 1 constructs a dendrogram that sequentially merges the classification axes having a high similarity among the plurality of classification axes characterizing predetermined information in a hierarchical manner, and searches the dendrogram to find a lower layer. And hierarchical clustering means for generating an index that can identify the upper layer from the storage means and storing it in the storage means.

  According to the present invention, it is possible to construct a dendrogram that sequentially merges high-similarity classification axes among a plurality of classification axes that characterize predetermined information, and search the dendrogram to identify the upper layer from the lower layer Since a simple index is generated and stored in the storage means, it is possible to speed up the subsequent clustering process.

  The present invention as set forth in claim 2, wherein the index read out from the storage means is referred to, and the upper layer repeats clustering by merging the same classification axis until the desired number of clusters is reached. , Further comprising.

  According to the present invention, referring to the index read from the storage means, the upper layer repeats clustering by merging the same classification axis until the desired number of clusters is reached, thereby eliminating tag ambiguity, It becomes possible to prevent explosion of the number of tags.

  The present invention according to claim 3 constructs a dendrogram that sequentially merges the classification axes having a high degree of similarity among a plurality of classification axes characterizing predetermined information in a hierarchical manner, and searches the dendrogram for a lower layer. And generating an index that can identify the upper layer from the storage means and storing it in a storage means.

  According to the present invention, it is possible to construct a dendrogram that sequentially merges high-similarity classification axes among a plurality of classification axes that characterize predetermined information, and search the dendrogram to identify the upper layer from the lower layer Since a simple index is generated and stored in the storage means, it is possible to speed up the subsequent clustering process.

  The present invention described in claim 4 refers to the index read from the storage means, and repeats the upper layer to merge and cluster the same classification axis until the desired number of clusters is reached. Furthermore, it is characterized by having.

  According to the present invention, referring to the index read from the storage means, the upper layer repeats clustering by merging the same classification axis until the desired number of clusters is reached, thereby eliminating tag ambiguity, It becomes possible to prevent explosion of the number of tags.

  The present invention described in claim 5 is characterized by causing a computer to execute each step described in claim 3 or 4.

  According to the present invention, at the time of tag reuse, tag ambiguity can be resolved and explosion of the number of tags can be prevented.

It is a figure explaining the concept of clustering. It is a figure which shows an example of the dendrogram constructed | assembled by hierarchical clustering. It is a figure explaining the merit and demerit which describe presentation information with various tags. It is a figure which shows schematically the functional block structure of the whole system. It is a figure which shows roughly the functional block structure of a clustering apparatus. It is a figure which shows the whole processing flow of a clustering apparatus. It is a figure explaining the clustering to a partial tag set. It is a figure which shows the processing flow of a hierarchical clustering part. It is a figure which shows an example of a dendrogram and a bottom-up index. It is a figure which shows the processing flow of a partial clustering part. It is a figure which shows an example of the processing flow of a partial clustering part. It is a figure explaining the transition of partial clustering. It is a figure explaining the transition of the cluster to which a tag belongs. It is a figure which shows an example of the tag set clustered. It is a figure which shows the example of a classification | category of the classification axis | shaft in a cooperative classification system. It is a figure explaining the flow which reuses the tag produced | generated by the classifier. It is a figure explaining the concept of the conventional clustering.

  Hereinafter, an embodiment for carrying out the present invention will be described with reference to the drawings. However, the present invention can be implemented in many different modes and should not be construed as being limited to the description of the present embodiment.

  Before describing the configuration and processing of the clustering apparatus according to the present embodiment, the outline of the present invention will be described in advance in order to facilitate understanding. This embodiment is generated by a third party when generating provided information to a viewer using a given classification axis (hereinafter referred to as a tag) in the cooperative classification system described in the background art. In addition, by selecting various tags as quickly as possible, the ambiguity of the tags at the time of tag reuse is resolved and the explosion of the number of tags is prevented.

  That is, as shown in FIG. 1, all tags are clustered in advance in consideration of a plurality of tags other than the target tag shown in FIG. 17 (Pre-Processing), and the result of the pre-clustering is used as an index. By using it, tags are diversified faster than before.

  Clustering includes a division optimization method and a hierarchical method (see Non-Patent Document 7). In this embodiment, a hierarchical method is used. That is, a dendrogram on a tag set constructed by hierarchical clustering is used. Here, the features of the dendrogram constructed by hierarchical clustering will be described below.

  FIG. 2 is a diagram illustrating an example of a dendrogram constructed by hierarchical clustering. This dendrogram has a binary tree structure. When tags with high similarity are grouped one by one, the whole tag set is finally obtained (aggregation type). Conversely, the entire tag set is halved. When the operation of dividing is repeated, the feature is that each becomes a single tag at the end (branch type). The present invention can be applied to any of the aggregation type and branch type algorithms, and the branch type will be described.

  The numbers given to the intermediate nodes in FIG. 2 represent the order in which the entire set is divided in order from the upper layer node to the lower layer node. For example, if the dendrogram is cut at the second intermediate node from the root node, the entire set is divided into two. Subsequently, when cutting at the third intermediate node, the entire set is divided into three. Further, when the cut is made at the fourth intermediate node, it is divided into four (A to D shown in FIG. 2), and as a result, the whole set is clustered with the numbers of the intermediate nodes. Due to this property, once a dendrogram is constructed on a tag set, it can be cut to an arbitrary number less than the total number of tags. Note that “cut” means clustering.

  On the other hand, in a normal division optimization method, clustering by cutting a tag set into several pieces depends on data and is one of the items to be tuned. In the hierarchical clustering used in this embodiment, once a dendrogram is constructed on a tag set, an arbitrary number of clustering results can be obtained only by cutting the dendrogram. That is, constructing a hierarchical clustering to generate a dendrogram is equivalent to generating an arbitrary number of clusters.

  This embodiment is based on the premise that various tag sets are effective. The set of various tags is a set of tags having meanings that are not similar to each other, or a set of tags that include tags that are similar but have different meanings. As a set of various tags, “Web”, “essay”, “Google”, “read later”, “HTTP” can be cited as examples. On the other hand, a non-various set of tags represents content that is semantically close, and examples thereof include “Web”, “Web 2.0”, “WWW”, “Internet”, and “HTTP”.

  For reference, the merits and demerits of describing presentation information with various tags are shown in FIG. Rather than presenting tag sets with high similarity such as “soccer”, “goal”, “news”, and “soccer”, a variety of tag sets with low similarity such as “soccer”, “Brazil”, “news”, and “rare play” Thus, it is possible to appropriately grasp information from the presented information.

  Next, an overall system configuration having the clustering apparatus according to the present embodiment will be described. FIG. 4 is a diagram schematically showing a functional block configuration of the entire system. The overall system is a collaborative classification system that can communicate with a client terminal 5 used by a tag classifier or a viewer of presentation information (hereinafter collectively referred to as a user) and a client terminal 5 via a communication network 3 such as the Internet. 1.

  The cooperative classification system 1 is provided as a network service, and the user can use the network service through a web browser or a client application installed in the client terminal 5. Specifically, it comprises a communication unit 11 such as a LAN or a router, a classification result storage database 12 for storing data, and a clustering device 13 for actually realizing processing.

  As described in the background art, the cooperative classification system 1 includes an application that allows a classifier to freely generate a tag and assign the generated tag to a classification target. Description of specific functions and processing contents related to the application is omitted. An example of the cooperative classification system 1 is a social bookmark service (Social Bookmark Service: a system that shares bookmarks of users through a network).

  The communication unit 11 realizes communication such as external or internal communication or inter-internal communication by the cooperative classification system 1 at the network level.

  As described in the background art, the classification result storage database 12 includes tags and resources that are freely generated by a classifier based on various resource information such as URL (bookmark), photograph, moving image, still image, book, and paper. It has a function of storing tags attached to information in a readable manner.

  In the present embodiment, it is sufficient that tags are generated in advance and stored in the classification result storage database 12, and the clustering device 13 can use any method to generate tags. Even with various types of tags, the same effect can be obtained regardless of the tag generation method and the tag type by executing the processing described later.

  As shown in FIG. 5, the clustering apparatus 13 includes a communication interface 131, a hierarchical clustering unit 132, a partial clustering unit 133, a representative tag selection unit 134, and a storage unit 135.

  The communication interface 131 has a function of mediating communication with the communication unit 11 and the classification result storage database.

  The hierarchical clustering unit 132 includes a clustering unit 132a that hierarchically clusters a plurality of tags (hereinafter referred to as an entire tag set) to construct a dendrogram, and an index generator that searches the dendrogram and generates a bottom-up index. Part 132b.

  The partial clustering unit 133 includes an index reference unit 133a that refers to the generated index, and a sorting unit 133b that clusters the entire tag set into a plurality of partial tag sets using the reference index.

  The representative tag selection unit 134 has a function of selecting an arbitrary representative tag from each partial tag set.

  The storage unit 135 has a function of storing the constructed dendrogram and the generated index in a readable manner. Such a storage unit 135 can be realized by a memory such as a ROM or a RAM or a storage device such as a hard disk.

  In addition, each said function part which comprises the clustering apparatus 13 can also be implement | achieved by a single server apparatus, and can also be implement | achieved by the structure which distributed and arranged each function in the some server apparatus. It is.

  Next, a processing flow of the clustering apparatus 13 having the above configuration will be described. FIG. 6 is a diagram showing an overall processing flow of the clustering apparatus.

  First, the hierarchical clustering unit 132 hierarchically clusters the entire tag set to construct a dendrogram, and a bottom-up index is generated in advance (S1).

  Next, when there is a request from the client terminal 5 or another upper application, the partial clustering unit 133 refers to the index generated in S1, and the entire tag set is clustered into a plurality of partial tag sets (S2). .

  Finally, the representative tag selection unit 134 selects the most frequent representative tag from each of the partial tag sets clustered individually, and returns the selected k various tag sets to the requesting higher-level application. (S3).

  Here, the reason why the generated index is bottom-up and clustering is performed using the bottom-up index will be described.

  It is also possible to generate an index from the top down and cluster it into a partial tag set. For example, as shown in FIG. 7, a case will be described in which clustering into a partial tag set is performed using hatched tags (T1, T4, T7, T9, T13, T14) instead of the entire tags (T1 to T15). In this case, since all the tags (T1 to T15) can be clustered with a smaller number of cuts than when clustering into partial tag sets, the division order of intermediate nodes (numbers surrounded by circles in FIG. 7) is as shown in FIG. Unlike the case shown in Fig. 5, the number of partial tag sets to be clustered is also different.

  Then, when a set is divided at a certain intermediate node, it is possible to perform clustering into a partial tag set by searching whether or not a hatched tag is included in the divided subset.

  However, since it cannot be determined whether the diagonal tag exists in the partial tag set unless the dendrogram is traced to the leaf node (the lowest tag), it cannot be efficiently clustered into the partial tag set.

  Therefore, in this embodiment, the dendrogram on the entire tag set is materialized in advance as a bottom-up index, and clustered into partial tag sets in the bottom-up manner.

  Next, the processing flow of the hierarchical clustering unit 132 will be specifically described. FIG. 8 is a diagram illustrating a processing flow of the hierarchical clustering unit.

  First, the clustering unit 132a reads a plurality of tags from the classification result storage database 12, and constructs a dendrogram (dendrogram) in which tags with high similarity are sequentially merged in a hierarchical manner (S11).

  Finally, the index generator 132b searches the dendrogram constructed in S11, generates a bottom-up index that can identify the upper layer from the lower layer, and stores it in the storage unit 135 (S12).

  From the above, a dendrogram (binary tree on the whole tag set (T1 to T15)) as shown in FIG. 9A is constructed by S11, and a bottom-up index as shown in FIG. 9B is obtained by S12. Is generated.

  In a normal use scene of hierarchical clustering, this dendrogram itself is only used for visualization for the purpose of grasping the relationship between tags, but in this embodiment, this hierarchy is bottom-up. Keep as an index. A bottom-up index is an index with a lower layer node in a dendrogram as a key, and an upper layer node (hereinafter also referred to as a parent node) is obtained from a lower layer node (hereinafter also referred to as a child node). Can do. If a dendrogram is constructed, a bottom-up index can be generated by scanning the dendrogram once.

  That is, the hierarchical clustering unit 132 hierarchically clusters the entire tag set to construct a dendrogram, and a bottom-up index is generated in advance, so that the clustering process by the partial clustering unit 133 in the subsequent stage is accelerated. Is possible.

  Note that both split and merged algorithms can be used as hierarchical clustering algorithms, but a merged algorithm can generate a bottom-up index at the same time as building a dendrogram, so a branching algorithm is used. It is possible to generate a bottom-up index at a higher speed.

  Next, the processing flow of the partial clustering unit 133 will be specifically described. FIG. 10 is a diagram illustrating a processing flow of the partial clustering unit.

  First, the index (see FIG. 9B) generated by the hierarchical clustering unit 132 is read from the storage unit 135 by the index reference unit 133a (S21).

  Finally, the sorting unit 133b refers to the index read in S21, and the upper layer (parent) merges the same tags to form a partial tag set (clustering). It repeats until it becomes (the number of clusters) (S22).

  That is, the partial clustering unit 133 first considers one tag as one cluster, initializes the number of clusters, and bottoms up by merging clusters every time a merge is found while climbing the dendrogram bottom up. It is characterized by performing various processes. Hereinafter, an example of the processing flow in S22 will be described below with T1, T4, T7, T9, T13, and T14 as clustering target tags. FIG. 11 is a diagram illustrating an example of a processing flow of the partial clustering unit.

  First, the input of the designated cluster number k requested by the application, the partial tag set T ′ to be clustered, and the bottom-up index IDX acquired in advance is received (S31). Assume that the designated cluster number k is 3, the partial tag set T ′ is T1, T4, T7, T9, T13, T14, and the bottom-up index IDX is FIG. 9B.

  Next, the number of temporary clusters c (= | T ′ |) at that time, the parent node list P obtained by obtaining the parent node division order of the partial tag set T ′ from the bottom-up index IDX and sorting in descending order, and the parent node The position pointer cp in which the division order of the parent node of the node having the largest division order in the list P is set as a temporary variable (S32). Since the partial tag set T ′ is T1, T4, T7, T9, T13, and T14, at this time, c = 6, P = 5, 7, 8, 11, 14, 15, and cp = 13 are set. The

  Next, the temporary cluster number c is compared with the designated cluster number k (S33). If the temporary cluster number c is larger than the designated cluster number k, the following is performed until the temporary cluster number c matches the designated cluster number k. The processes of S34 to S39 to be described are repeated.

  Next, as a result of the comparison in S33, if the temporary cluster number c is larger than the designated cluster number k, the division order of the parent node of the node having the largest division order in the parent node list P is determined from the bottom-up index IDX. It is acquired and set in the position pointer cp (S34). Since the parent node list P is not changed, the same cp = 13 as the initial temporary value is set (see time point A shown in FIGS. 12 and 13).

  Next, it is determined whether or not the position pointer cp newly set in S34 is included in the parent node list P (S35). Referring to time point A in FIGS. 12 and 13, cp = 13 is not included in P.

  Next, as a result of the determination in S35, if the position pointer cp is not included in the parent node list P, the division order of the parent node of the node having the largest division order in the parent node list P is set to the bottom-up index IDX. And the rearrangement order of the largest node is exchanged with the obtained parent node division order and sorted in descending order, and the process returns to S33 (S36). Thereby, P = 5, 7, 8, 11, 13, and 14 are set.

  Thereafter, cp = 12 is set by the processing of S33 and S34 (see time point B shown in FIGS. 12 and 13). Similarly, P = 5, 7, 8, 11, 12, 13, and cp = 12 are set by the processes of S36, S33, and S34 (see time point C shown in FIGS. 12 and 13).

  Next, as a result of the determination in S35, if the position pointer cp is included in the parent node list P, another partial tag set of the same parent node as the parent node of the partial tag set that has been processed until now is displayed. Since it can be determined that it exists, by deleting the division order of the largest node in the parent node list P, the two partial tag sets are merged (S37).

  Next, the parent node list P is rearranged in descending order (S38), 1 is subtracted from the temporary cluster number c (S39), and the process returns to S33.

  Thereafter, cp = 4 is set by the processing of S33 and S34 (see time point D shown in FIGS. 12 and 13). Similarly, it is assumed that the current processing time point is time point E shown in FIGS. 12 and 13 by repeating the processing of S33 to S39.

  Next, as a result of the comparison in S33, if the temporary cluster number c is not larger than the designated cluster number k, a partial tag set clustered into k is output (S40). Thus, a partial tag set (T13 and T14) having P = 3 as a parent node, a partial tag set (T1 and T4) having P = 4 as a parent node, and a partial tag set having P = 5 as a parent node (T7 and T9) are output. As described above, after the processing of S40, the representative tag selection unit 134 selects the most frequent representative tag from each partial tag set, and the selected k various tag sets are the upper-level applications of the request source. Returned to

  As described above, as shown in FIG. 14, it is possible to provide various tag sets that increase the user satisfaction of the viewer at high speed from a large number of tag sets assigned by a third party.

  According to the present embodiment, a dendrogram is constructed by performing hierarchical clustering on the entire tag set, and a bottom-up index that makes it possible to identify the upper layer from the lower layer is generated in advance and requested from the upper application. When there is, the entire tag set is clustered into a plurality of partial tag sets with reference to the generated index, so clustering into the partial tag sets can be executed at high speed. It becomes possible to eliminate the ambiguity of the tags and prevent the explosion of the number of tags.

  In addition, the set of tags collected in the collaborative classification system can be divided into two categories: objective tags agreed upon by everyone, and subjective tags that are noisy as classification axes, or manual differentiation. By supporting this, it is possible to facilitate the utilization of the tag of the service provider.

  As a method of selecting tags used in a normal social bookmark service, there is a method of obtaining k items from the top in order of the number of tagging. However, various tag sets as described in the present embodiment are not necessarily selected in the order of simple tagging times.

  As another method, a supervised method of manually building a category dictionary to which tags belong and selecting tag sets belonging to as many different categories as possible is conceivable. However, in a collaborative classification system, third parties continue to tag every hour, so a dictionary-based approach is to guess the category of unknown tags or unclear tags (for example, emoticons, pictograms, etc.) Is difficult. According to the present embodiment, since the clustering is performed using the distance between the tags related to the similarity only from the result of the classification of the tags by a third party, various tags can be used without a teacher and without a teacher. A set can be obtained.

  Finally, the clustering apparatus described in the present embodiment is configured by a computer, and each process of each functional block is executed by a program. In addition, the clustering device described in the present embodiment is recorded as a program in a readable manner on a recording medium such as an optical storage device or a magnetic storage device, and this recording medium is incorporated into a computer or a program recorded on the recording medium is recorded. Of course, each processing operation described above can be made to function as a clustering apparatus by downloading to a computer via an arbitrary communication line or installing from a recording medium and operating the computer with the program. .

DESCRIPTION OF SYMBOLS 1 ... Cooperative classification system 11 ... Communication part 12 ... Classification result storage database 13 ... Clustering apparatus 131 ... Communication interface 132 ... Hierarchical clustering part 132a ... Clustering part 132b ... Index generation part 133 ... Partial clustering part 133a ... Index reference part 133b ... Sorting part 134 ... Representative tag selection part 135 ... Storage part 3 ... Communication network 5 ... Client terminal S1 to S3, S11 to S12, S21 to S22, S31 to S40 ... Step

Claims (5)

Build a tree diagram that sequentially merges high-similarity classification axes among a plurality of classification axes characterizing predetermined information, and searches the tree diagram to generate an index that can identify the upper layer from the lower layer Hierarchical clustering means stored in the storage means,
A clustering apparatus characterized by comprising: A partial clustering unit that refers to the index read from the storage unit and repeats clustering the upper layer by merging the same classification axes until the desired number of clusters is reached;
The clustering apparatus according to claim 1, further comprising: Build a tree diagram that sequentially merges high-similarity classification axes among a plurality of classification axes characterizing predetermined information, and searches the tree diagram to generate an index that can identify the upper layer from the lower layer And storing it in the storage means,
A clustering method characterized by comprising: Referring to the index read from the storage means, and repeating the upper layer merging the same classification axis and clustering until the desired number of clusters,
The clustering method according to claim 3, further comprising:   A clustering program that causes a computer to execute each step according to claim 3 or 4.

Images