Academia.eduAcademia.edu

Figure 6 – uploaded by Samuel Kaski

See full PDFdownloadDownload figure
Fig. 6. The keyword search mode was utilized to find information on color displays. 30 best-matching units were marked on the display with circles the size of which indicates the goodness of the match. As seen from the map display, the matches are distributed into several tight clusters found in different regions of the map. From two of these clusters the partial contents of a sample matching unit are shown in the insets. Closer inspection of the units reveals different aspects of color and displays. Unit a features a considerable number of abstracts about color filters used in building LCD displays, whereas in b one finds technology related to displaying colors when printing documents (the Descriptive-words lists were found for each map unit using the automatic keyword selection method introduced in [51]). The user who probably did not have printing in mind when formulating the query can then concentrate on the other clusters.

Figure 6 The keyword search mode was utilized to find information on color displays. 30 best-matching units were marked on the display with circles the size of which indicates the goodness of the match. As seen from the map display, the matches are distributed into several tight clusters found in different regions of the map. From two of these clusters the partial contents of a sample matching unit are shown in the insets. Closer inspection of the units reveals different aspects of color and displays. Unit a features a considerable number of abstracts about color filters used in building LCD displays, whereas in b one finds technology related to displaying colors when printing documents (the Descriptive-words lists were found for each map unit using the automatic keyword selection method introduced in [51]). The user who probably did not have printing in mind when formulating the query can then concentrate on the other clusters.

Related Figures (7)

Fig. 1. Overview of the construction and operation of the WEBSOM2 system. phases have been explained, e.g., in Honkela et al. [37]. Certain reasons expounded below, however, recently led us to abandon the semantic maps and to encode the word histograms by the newly developed projection methods. The results reported in the present article are totally based on these new develop- ments, which has given us a reason to call the present system the WEBSOM2. An overview of the WEBSOM2 system is depicted in Fig. 1.
Fig. 1. Overview of the construction and operation of the WEBSOM2 system. phases have been explained, e.g., in Honkela et al. [37]. Certain reasons expounded below, however, recently led us to abandon the semantic maps and to encode the word histograms by the newly developed projection methods. The results reported in the present article are totally based on these new develop- ments, which has given us a reason to call the present system the WEBSOM2. An overview of the WEBSOM2 system is depicted in Fig. 1.
It can be shown that the similarity of pairs of projection vec- tors (x;,x,), measured by the inner products, is the same on the average as the similarity of the pairs of corresponding orig- inal document vectors (n;,n,;), and the error thereby made is inversely proportional to m [41]. It was demonstrated experi- mentally [41] for dimensionalities of m exceeding 100 (cf. also Table I) that the classification accuracy is practically as good as with the vector-space method while, with the decreasing dimen- sionality of the document vectors, the time needed to classify a document is decreased radically.
It can be shown that the similarity of pairs of projection vec- tors (x;,x,), measured by the inner products, is the same on the average as the similarity of the pairs of corresponding orig- inal document vectors (n;,n,;), and the error thereby made is inversely proportional to m [41]. It was demonstrated experi- mentally [41] for dimensionalities of m exceeding 100 (cf. also Table I) that the classification accuracy is practically as good as with the vector-space method while, with the decreasing dimen- sionality of the document vectors, the time needed to classify a document is decreased radically.
collection, each document was mapped onto one of the grid points of each map, and all documents that represented a minority class at any grid point were counted as classification errorss .
collection, each document was mapped onto one of the grid points of each map, and all documents that represented a minority class at any grid point were counted as classification errorss .
Fig. 2. Illustration of the relations of the model vectors in a sparse (s, solid lines) and dense (d, dashed lines) grid. Only partial grids are shown in the figure. Here m,,’‘4 shall be interpolated in terms of the three closest sparse models m,'‘), m,’¢), and m,;’*?, respectively.
Fig. 2. Illustration of the relations of the model vectors in a sparse (s, solid lines) and dense (d, dashed lines) grid. Only partial grids are shown in the figure. Here m,,’‘4 shall be interpolated in terms of the three closest sparse models m,'‘), m,’¢), and m,;’*?, respectively.
Fig. 3. Finding the new winner in the vicinity of the old one, whereby the old winner is directly located by a pointer. The pointer is then updated.
Fig. 3. Finding the new winner in the vicinity of the old one, whereby the old winner is directly located by a pointer. The pointer is then updated.
Fig. 4. Distribution of four sample subsections of the patent classification system on the document map. The gray level indicates the logarithm of the number of patents in each node.
Fig. 4. Distribution of four sample subsections of the patent classification system on the document map. The gray level indicates the logarithm of the number of patents in each node.
Fig. 5. Content addressable search was utilized to find information on laser surgery on the cornea of the eye. The best matching locations are marked with circles. Zooming on the area reveals a small cluster of map units that contains patent abstracts mostly about the cornea of the eye, and of surgical operations on it. Several abstracts concemed with the description of interest, i.e., laser surgery on the cornea, are found in the best matching units.
Fig. 5. Content addressable search was utilized to find information on laser surgery on the cornea of the eye. The best matching locations are marked with circles. Zooming on the area reveals a small cluster of map units that contains patent abstracts mostly about the cornea of the eye, and of surgical operations on it. Several abstracts concemed with the description of interest, i.e., laser surgery on the cornea, are found in the best matching units.
Related topics:
High Dimensional DataSelf Organized Map
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved