JP3315230B2 - Similarity search device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a similarity retrieval apparatus for retrieving similar data, and more particularly to a similarity retrieval apparatus for automatically retrieving and extracting similar data that accurately reflects a search intention from existing data.

[0002]

2. Description of the Related Art A device for retrieving data similar to specific data from stored data is called a similarity retrieval device. ),
It is applied to various fields such as image search of image processing and dictionary search of natural language processing.

FIG. 24 shows an example of a schematic configuration diagram of a conventional similarity search apparatus for searching all data. In this similarity search device, first, information given as a problem case is input to an input device 171 including a keyboard, a mouse, and the like. The input information is decrypted by the input / output control unit 172, and the decrypted content is captured. Based on this data, the inference processing unit 173 instructs the search unit 174 to search. The search unit 174 calls up all the data related to the problem case from the database 175 storing the data represented by the attribute-attribute value pair and the data of the similarity, and the similarity r = Σ (weight of the attribute ×× A similarity search is performed by comparing the magnitude of the similarity between the cases as represented by the similarity d) between the attribute values. Next, the search result is derived by the inference processing unit 173, and the input / output control unit 172 outputs the search result to an output device 176 including a CRT display device or the like.

However, in such a conventional similar search apparatus, when the number of data is large, the amount of calculation and the time required for the similar search become enormous.

[0005] On the other hand, FIG. 25 shows a schematic configuration diagram of a conventional similarity search apparatus that searches a large amount of data after narrowing it down with an artificially created index.

Here, when a data search is performed by a similar search device, what is used as a search key is called an "index". The two indexes used in the following description, that is, the first index and the second index are distinguished. For clarity, we define them. First, the first index is used to narrow down the number from a large amount of data in the database, and is an index that must strictly satisfy the conditions. And
A similarity search is performed on the data group that satisfies the first index. At this time, the data used to evaluate the similarity between the input problem data and the searched data in detail is the second search. It is an index. The second index is the weight of each attribute.

[0007] In the similarity search apparatus of FIG. 25, first, the problem case data and the first index are input to the input device 181. The input data is decrypted by the input / output control unit 182, and the inference processing unit 183 instructs the first search unit 184 to narrow down the data. The first search unit 184 calls data on the problem case from the database 185 and calls data satisfying the first index. The inference processing unit 183 sends this data group to the second search unit 186,
Then, a search is performed from the narrowed data by the same means as the search unit 174 of the similar search device in FIG. Next, this search result is derived by the inference processing unit 183, and the input / output control unit 1
82 outputs the search result to the output device 187.

However, it is difficult to search for similar data as desired even with the first index set as described above. This is because, in the conventional similar search apparatus, the user cannot change the search width because the user fixes the first index.

Therefore, as one method of changing the first index, the user can use the given second index (attribute weight) and the number of search data as criteria for strengthening or relaxing the first index. Can be considered. For example, if you want to increase the number of search data,
The first index is relaxed from its current state in order to relax restrictions on attributes having a small second index value, that is, attributes having low importance. Conversely, if it is desired to further reduce the number of search data, the first index is strengthened from the current state in order to increase the constraint on the attribute having the large second index value, that is, the attribute having high importance.

However, in this method, when the first index is changed, there is no index that can be used as a criterion as to which attribute should be performed or which attribute should be relaxed or strengthened to what level. Since the change of the first index is all a trial and error method by the user, it is difficult to adjust the first index in a well-balanced manner, and it is not possible to perform a similar data search that accurately reflects the search intention. there were.

[0011]

The conventional similar search apparatus described above has a problem that it takes a long time to search for all the data as a search target as in the former case, while a large amount of data as in the latter case. If the search is performed after narrowing the first index by the first index set by the user through trial and error, it is difficult to set the first index in a well-balanced manner, so it is not possible to narrow down the data accurately, and the search intention There was a problem that it was not possible to perform a similar data search that accurately reflected.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and an effective similarity search that accurately reflects a search intention by setting a first index in a well-balanced manner according to a data accumulation state. It is an object of the present invention to provide a similarity search device capable of performing a search.

[0013]

According to the present invention, there is provided a similarity search apparatus according to the present invention, wherein: a data storage means for storing a plurality of sets of data represented by a plurality of sets of attributes and attribute values; A similarity evaluation knowledge storage unit for storing similarity evaluation knowledge representing a similarity between a structure and the attribute value; and a plurality of data stored in the data storage unit based on the similarity evaluation knowledge. Classifying means for classifying according to a second index indicating a range of attribute values for each and an externally given importance of each attribute, and a target similarity generation for generating a target similarity from a data classification result by the data classifying means Means, first index generating means for generating a first index indicating a range of attribute values of each attribute from the generated target similarity, and using the first index Wherein for a plurality of data stored in the data storage means, a first search means to narrow the amount of data by performing a search, using the second index,
A second search unit that extracts data similar to data having a specific attribute value from the narrowed-down data.

Further, in the similarity search device according to the present invention (claim 2), a data storage means storing a plurality of data represented by a plurality of sets of attributes and attribute values, a hierarchical structure of the attribute values and the attribute A similarity evaluation knowledge storage unit that stores similarity evaluation knowledge representing a degree of similarity between values; and, based on the similarity evaluation knowledge, a plurality of data stored in the data storage unit. Data classification means for classifying according to a second index indicating the importance of each attribute given from the range and externally provided, a classification result presentation means for presenting the data classification result to a user, and instructed based on the classification result First index generation means for generating a first index indicating a range of attribute values of each attribute from the target similarity, and the data storage means using the first index. For a plurality of data stored, and the first search means to narrow the amount of data by performing a search,
A second search unit configured to extract data similar to data having a specific attribute value from the narrowed data using the second index.

Preferably, in each of the above structures,
The apparatus further includes a target similarity changing unit that interactively changes the target similarity.

On the other hand, in the similarity search device according to the present invention (claim 4), a data storage means for storing a plurality of data represented by a plurality of sets of attributes and attribute values, a hierarchical structure of the attribute values and the attribute A similarity evaluation knowledge storage unit that stores similarity evaluation knowledge representing a degree of similarity between values; and, based on the similarity evaluation knowledge, a plurality of data stored in the data storage unit. Data classification means for classifying according to the second index indicating the range of the attribute and the importance of each attribute given from the outside, and generating a first index indicating the range of the attribute value of each attribute from the data classification result by the data classification means Using the first index generating means and the first index, performing a search on a plurality of data stored in the data storage means to narrow down the data amount. A first search unit that writes, using the second index, and characterized by including a second search means for extracting data similar to data having a specific attribute value from among the narrowed-down data.

Preferably, in each of the above structures,
The apparatus further comprises a first index changing means for interactively changing the first index.

Preferably, in each of the above structures,
The apparatus further comprises means for storing the data classification result.

[0019]

In the similarity search device according to the present invention (claim 1),
A first index used for a search (strict search) for narrowing down a data amount of a plurality of data stored in the data storage unit, which is performed by the first search unit, is empirically determined by a user as in the related art. It is generated automatically instead of being determined by trial and error.

That is, first, the data classification means classifies the plurality of data stored in the data storage means according to the range of the attribute value for each attribute and the second index. The target similarity generating means determines a target similarity to be satisfied by the data at the time of the strict search from the classification result. Then, the first index generation means determines a first index indicating the range of the attribute value of each attribute from the generated target similarity. In other words, the first index of which attribute should be relaxed or strengthened and which attribute value should be relaxed, and how far the relaxed or strengthened level should be performed, is used for the classification result of the data stored in the database. After the reflection, it is automatically generated in a well-balanced manner.

By performing a strict search using the first index generated according to the data accumulation state in this manner, data can be narrowed down sufficiently reflecting the search intention, and further performed by the second search means. A similar search is an effective one that accurately reflects a search intention.

Further, in the similarity search device according to the present invention (claim 2), a search (strictly) for narrowing down a data amount of a plurality of data stored in the data storage means performed by the first search means. The first index used for the search is semi-automatically generated, instead of being determined by the user by empirical, trial and error methods.

That is, first, the data classification means classifies the plurality of data stored in the data storage means according to the range of the attribute value for each attribute and the second index. The classification result presentation means presents the data classification result to the user. The first index generation means includes:
The user waits for the user to specify a desired target similarity with reference to the classification result, and generates a first index indicating the range of attribute values of each attribute according to the instruction. In other words, while the user has conventionally determined the first index by empirical and trial and error, the classification result presenting means presents the data classification result. One index can be specified.

For this reason, by performing a strict search using the first index specified as described above, it is possible to narrow down data that sufficiently reflects the search intention.
Further, the similarity search performed by the second search means is an effective one that accurately reflects the search intention.

On the other hand, in the similarity search device according to the present invention (claim 4), the first index is automatically generated directly from the classification result without generating the target similarity.

By performing a strict search using the first index generated according to the data accumulation state as described above, the data can be narrowed down sufficiently reflecting the search intention, and further performed by the second search means. A similar search is an effective one that accurately reflects a search intention.

[0027]

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing a schematic configuration of a similarity search apparatus according to one embodiment of the present invention. Here, a case where the similarity search device of the present invention is applied to a similarity search in a case-based reasoning system will be described. Case-based inference is inference in which past problem-solving experiences are accumulated as cases, and similar cases are searched for and corrected for new problems to reach conclusions. In the present embodiment, specifically, a description will be given of generating a well-balanced index from hierarchical relation data (tree structure) and performing a simple and accurate similarity search according to the data accumulation status. Further, as a specific example, a case will be described in which an example similar to the “banquet decision item” of “Sukiyaki in June in Tokyo” is searched from the case data in the database.

As shown in FIG. 1, the similarity search apparatus of this embodiment
Input device 101, input / output control unit 102, inference processing unit 10
3. First search unit 104, first index generation unit 10
5, a configuration including a data classification unit 106, a target similarity generation unit 107, a classification data storage unit 108, a database 109, a similarity evaluation knowledge 110, a second search unit 111, and an output device 112.

In the above-described configuration, the input device 101 inputs data representing attributes and attribute values as "attribute-attribute values" as problem cases, for example, "region-Tokyo", "time-June",
Enter "Food-Sukiyaki". This input data is decoded by the input / output control unit 102, and the decoding result is output to the inference processing unit 1
03 is sent. In response, the inference processing unit 103 instructs the first search unit 104 to perform a strict search. Here, the strict search refers to a search for data that must be strictly satisfied in order to narrow down the number from a large amount of data in the database.

First search unit 104 instructed to perform a strict search
First instructs the first index generation unit 105 to create a first index in order to generate an index used for an exact search.

The first index generation unit 105 determines the data accumulation status based on the case data registered in the database 109, the evaluation knowledge registered in the similarity evaluation knowledge 110, and the second index specified by the user. Generate an appropriate first index accordingly.

Hereinafter, the generation of the first index performed by the first index generation unit 105 will be described. The procedure for generating the first index by the first index generation unit 105 is shown in the flowchart of FIG. A flowchart of FIG. 3 shows a procedure for classifying data, and a flowchart of FIG. 4 shows a procedure for generating a target similarity required for generating the first index in the target similarity generator 107.

FIG. 5 shows some case data. As shown in FIG. 5, the case data is composed of a plurality of (attribute-attribute value pairs). For example, case 1 has attribute values of (region-Kanagawa), (time-August), (food-pan). I understand.

In the present embodiment, as evaluation knowledge, a hierarchical relationship between attribute values in each attribute and a similarity indicating a degree of similarity between attribute values are used. FIGS. 6, 7 and 8 show the hierarchical relationship of attribute values for each attribute. In the case of the attribute of the region in FIG. 6, there are Kanto and Kansai below the whole of Japan,
It also indicates that there is a hierarchical structure of tree structure such as Tokyo and Kanagawa under Kanto. 9 and FIG.
FIG. 11 shows the similarity between attribute values for each attribute. The larger the similarity, the stronger the degree of similarity. The similarity between Tokyo and Kanagawa is 0.40, and the similarity between Tokyo and Kyoto is 0.01. 9 In the present embodiment, since the input question data is the data of the terminal node, the similarity is defined only between the terminal nodes. However, the input problem data does not need to be the terminal node. It is possible to set the similarity corresponding to the problem.

First, in step S23 in FIG. 2, the data in the database is classified using the data classification unit 106 for the input question. The details of this data classification will be described according to the flowchart of FIG.

In step S31, a range t of similarity between attribute values in the tree structure for each attribute is checked. For example, when examining the area, in the Kanto range (excluding Tokyo) from the attribute value “Tokyo” in question (excluding Tokyo), it is 0.40 for Kanagawa with the highest similarity and 0.05 for the smallest Tochigi, so 0.05 ≦ t ≦ 0.40. FIG. 12 shows the area, FIG. 13 shows the time, and FIG. 14 shows the food.

In step S32, a second index w, which is the weight (importance) of each attribute, is input. Here, the weight of the attribute of the area is 0.2, the weight of the attribute of the time is 0.5, and the weight of the attribute of the food is 0.3.

In step S33, w × t (w is the second index, t is the range of similarity between attribute values) is calculated (hereinafter, s = w × t), and the range of s is examined.

For example, when examining the area, in the range from Tokyo to Kanto, Kanagawa with the highest similarity is 0.4%.
0, the smallest Tochigi is 0.05, so Kanagawa's second index 0.2 x the similarity between attribute values 0.40 and s is 0.08, and Tochigi's second index 0.2 x the attribute value At the similarity of 0.05, s becomes 0.01. That is, the range of s is 0.01 ≦ s ≦ 0.08. this
(a1). FIG. 15 shows the area, FIG. 16 shows the time, and FIG. 17 shows the food.

In step S34, data (cases) are grouped for each similarity range r. The similarity range r is obtained from the combination of s obtained in step S33. FIG. 18 shows a summary of these.

If the above grouping is satisfied in step S35, the classification of the data is determined.
If the grouping is not satisfactory, the process returns to step S31 or step S32, and the same processing is performed again. When returning to step S31, the process starts by changing the similarity between the attribute values registered in the database, and when returning to step S32, the process starts by changing the second index w.

When the classification of data is completed as shown in FIG. 18, these are stored as data information in the classification data storage unit 108. FIG. 19 shows the stored classification data.
In the flowchart of FIG. 2, these are stored in the classification data in step S24.

By storing the classification result, the trouble of classifying existing input data can be saved, which is efficient. Since there is no classification information for new data, it is necessary to perform the classification from the data. However, if there is existing classification information for the input problem, the classification processing is skipped as shown in the flowchart of FIG. It is possible. In step S21, an existing classification result is checked from the classification data, and in step S22, it is determined whether to classify the data. If the user selects not to perform classification, the process can proceed to the next processing step S25 using the classification information stored by the classification data storage / storage unit 108. If the user selects to classify data in step S22, the flow of the processing after step S23 is the same as that described with reference to FIG.

Next, a target similarity generator 107 generates a target similarity. FIG. 2 corresponds to step S25. The generation of the target similarity will be described in detail with reference to the flowchart of FIG.

In step S41, the number of cases is checked for each similarity range r, that is, for each grouped r.

In step S42, a distribution display method for this result is selected. The distribution display method is roughly divided,
There is a range information display of the similarity and a case number display according to the range. As a default of the similarity range information display, a range of r is set on the horizontal axis so that the similarity ranges of the attributes can be compared in group units. As the default of the case number display, r is taken on the horizontal axis and the case number is taken on the vertical axis, and all r are displayed.

In step S43, distribution is displayed by the selected method and presented to the user. An example of a display format (here, a default) is shown in FIG. 21A for displaying similarity range information, and FIG. 21B for displaying the number of cases according to the similarity range.

FIG. 22 shows the result of summarizing the similarity range r and the number of cases belonging to the range. Although the number of cases is searched for each group shown in FIG. 18, the search is not the case contents, so that the search can be performed at high speed without disk access.

Here, knowing the group result in the similar range and the number of cases corresponding to the similar result has an advantage that it is possible to grasp the cases to be input in the future as a secondary effect. It is also effective information in knowledge acquisition.

Next, if the user is satisfied with the presented display method, this is indicated in step S44 by the input device 101.
Then, the process proceeds to the selection of the target similarity in step S45. On the other hand, if the result is not satisfied in step S44, the process returns to step S42, and a display result based on a satisfactory display method is executed again.

Next, the selection of the target similarity in step S45 will be described. Selection of the target similarity is performed by designating a selected one from a plurality of similarity ranges r as the target similarity.
A manual mode in which the user himself refers to the above-described presentation contents and an automatic mode in which the target similarity generation unit 107 performs in accordance with a predetermined method can be considered. However, basically, in any of the modes, the following can be considered as a method of selecting the similarity range.

First, it is desirable to select one without bias for each attribute. In the present embodiment, each attribute is weighted according to importance so that determination can be made on the same scale. Therefore, an attribute having many common parts between attributes may be selected. Then, in consideration of the number of cases, if there are some groups satisfying the number of cases required at the time of similarity search, in that case, the one with the larger range may be given priority. Further, a different determination method may be used for each domain.

Hereinafter, the manual mode will be described more specifically.

(I) First, referring to the similarity range information display as illustrated in FIG.
From 1 to r18, the one having a large (wide) common part in the range of the similarity that each attribute value can take is selected. As described above, the importance between the respective attribute values is considered and the weight is assigned according to the importance, so that the determination can be made on the same scale. In this case, r4, r6, r10, r12, r16,
r18 or the like is selected as a candidate.

(Ii) Next, referring to the case number display according to the similarity range as exemplified in FIG. 12A, an appropriate case number is selected from the candidates. As a selection method, a method of determining the range of the number of cases to be selected may be determined. For example, a method of determining the ratio (or range of ratio) of the number of cases to the total number of data, or a method of fixedly determining the allowable number of cases (or range of number of cases) regardless of the total number of data, etc. Can be considered.

Here, assuming that the standard of the number of cases is 50, r10 is selected as the target similarity from the candidates.

It is also effective to avoid a large single area from the common part of the similarity range i.

Next, the automatic mode will be described using a more specific example.

(I) First, a range of similarity taken by each r is calculated. The results are shown below.

R1 0.075 to 0.08 r2 0.075 to 0.08 r3 0.030 to 0.08 r4 0.025 to 0.08 r5 0.030 to 0.08 r6 0.015 to 0.05 08 r7 0.075 to 0.08 r8 0.075 to 0.08 r9 0.030 to 0.08 r10 0.025 to 0.08 r11 0.030 to 0.08 r12 0.015 to 0.08 r13 0.075 to 0.08 r14 0.075 to 0.08 r15 0.030 to 0.08 r16 0.025 to 0.08 r17 0.030 to 0.08 r18 0.015 to 0.08 Priority ranking. The results are shown below.

First rank (0.015 to 0.08) r6, r12, r18 Second rank (0.025 to 0.08) r4, r10, r16 Third rank (0.030 to 0.08) r3 , R5, r9 r11, r15, r17 (ii) Next, an appropriate number of cases is selected. The method for determining the appropriate number of cases is the same as in the manual mode described above.

This selection is determined for each group of the similarity. That is, of the ranking results, the first
Judgment is made from those belonging to the ranking. If there is no order having the appropriate number of cases in the determined order, the order belonging to the next order is determined.

For example, first rank r6, r12, r18
With regard to, the attribute value of time is too relaxed and the number of cases is too large, so the second rank is determined next. As for r4, the condition of the attribute value of the area was too relaxed, and the number of cases was too large. However, as for r10, the number of cases was 52, which was an appropriate number, so that r10 was finally selected. Is performed.

In any mode, in addition to the above-described methods, various selection methods can be considered.

In the automatic mode, the distribution display in step S43 described above may be omitted.

The target similarity is generated by the processing described above.

Then, in step S26 in FIG.
Based on the generated target similarity, the first index is determined as “Honshu” for the region, “Spring / Summer” for the time, and “root” for the food.

Next, the first index generation unit 105
An index to be passed to the first search unit 104 is generated.

The index is generated by expanding hierarchical data from the similarity evaluation knowledge 110, that is, generating Honshu as a list in which all attribute values lower than Honshu are expanded, such as “Tohoku, Aomori, Akita,. Or an index that includes Honshu in the form of "IS-A Honshu". Although these description formats will depend on the database storage format, step S26 in FIG.
Then, an index in a format to be passed to the first search unit 104 is set.

The root is the top of the tree structure hierarchy. This means that all foods will be searched.

As described above, according to the set target similarity, the range of the case group is obtained without removing similar cases, and
The most similar case is searched from among them. ro
If you go back to ot, the number of cases will increase,
It must be determined according to the range of the upper and lower limits. Thus, in the present embodiment, the target similarity can be set while grasping the distribution status and the number of stored cases.

Next, the first index is changed in step S27. The above procedure is repeated from the data classification and the generation of the target similarity.

If there is no need to change, the first index will be “Spring,
Summer is determined, and food is determined as “root”.

As a method of changing the target similarity, a function may be added so that the user can appropriately adjust the target similarity based on the already generated target similarity and input the numerical value.

Next, the first search unit 104 performs a strict search using the first index generated by the first index generation unit 105 as described above. That is, the first search unit 104 outputs the first indexes “Honshu”, “Spring,
Based on “summer” and “root”, a corresponding case is searched from the database 109. For example, among the case data shown in FIG. 5, case 1, case 3, case 5, case 6, case 7, case 100, and the like are retrieved.

Here, according to the prior art, the user can set the first index by himself, for example, “Kanto”, “Spring, Summer”, “ROO”.
When “t” is set, case 100 is “Kanto” when searching the case corresponding to the first index from the case data.
There is a problem that the search is not performed because the condition is not satisfied, and the search intended by the user cannot be performed. However, in the present invention, the first index is generated by reflecting the status of the accumulated data in the database. Also, data such as the case 100 that is likely to be missed in the past can be narrowed down without being missed.

Further, according to the present invention, in the manual mode, the index is presented to the user, so that it is easy to set the target similarity for generating and changing the first index, and to determine the target similarity from the stored data. It is now possible to narrow down to a sufficient number of data (cases).

Next, the similarity between the cases is calculated by the second search unit 111 for these cases, and similarity data is searched from the magnitude thereof.

For example, the similarity r between cases is r = Σ (w ×
It is expressed by the calculation formula of d). Here, w is the second index, and d is the similarity between attribute values.

When the similarity between the problem case and the case 1 is actually calculated, (the second index, that is, the area weight 0.2
× Similarity between Tokyo and Kanagawa 0.40) + (time weight 0.
5 x June and August similarity 0.15) + (food weight 0.3 x sukiyaki and pan similarity 0.4), that is, r
= 0.2 × 0.40 + 0.5 × 0.15 + 0.3 × 0.
4 = 0.275.

Similarly, the similarity between the problem case and case 3, case 5, case 6, case 7, and case 100 is calculated, and the results are summarized in FIG.

When these are rearranged in descending order of similarity, case 100, case 7, case 3, case 1, case 5, and case 6 are obtained in this order.

Assuming that the number of cases required in this search is the top one, case 100 is searched as the case most similar to the problem case.

As described above, according to the present invention, the first index is generated by reflecting the status of the stored data, and the target similarity designated by trial and error is stored in the stored data. By presenting the situation to the user, an index can be set in a well-balanced manner according to the data accumulation situation, and an effective similarity search can be performed.

In the present embodiment, an example has been described in which the first index is generated using the target similarity as a medium. However, when the first index generation unit 105 is operated in the automatic mode, the generation of the target similarity is performed. It is also possible to generate the first index directly from the data classification result in step 23 without performing the above.

The present invention is not limited to the above-described embodiment, but can be implemented in various modifications without departing from the scope of the invention.

[0088]

According to the present invention, since the first index is generated by reflecting the status of the stored data in the database, a well-balanced index can be obtained, and an effective similarity search can be performed. .

When the first index is generated in the manual mode, the target similarity designated by trial and error can be presented to the user from the data accumulation status, so that efficient index setting is possible. is there.

[Brief description of the drawings]

FIG. 1 is a diagram showing a schematic configuration of a similarity search device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a procedure for generating a first index in the embodiment.

FIG. 3 is a flowchart showing a procedure of data classification in FIG. 2;

FIG. 4 is a flowchart showing a procedure of generating a target similarity in FIG. 2;

FIG. 5 is a diagram showing case data in the embodiment.

FIG. 6 is a view showing a hierarchical relationship of regional attributes in the embodiment.

FIG. 7 is a diagram showing a hierarchical relationship of a time attribute in the embodiment.

FIG. 8 is a diagram showing a hierarchical relationship between attributes of food in the embodiment.

FIG. 9 is a diagram showing a similarity between attribute values of regions in the embodiment.

FIG. 10 is a diagram showing a degree of similarity between attribute values of time in the embodiment.

FIG. 11 is a diagram showing the similarity between attribute values of food in the embodiment.

FIG. 12 is a view showing a range of similarity between attribute values of regions in the embodiment.

FIG. 13 is a diagram showing a range of similarity between attribute values of time in the embodiment.

FIG. 14 is a diagram showing a range of similarity between attribute values of food in the embodiment.

FIG. 15 is a diagram showing a range of s for an area in the embodiment.

FIG. 16 is a diagram showing a range of s for timing in the embodiment.

FIG. 17 is a diagram showing a range of s for food in the embodiment.

FIG. 18 is a diagram showing a result of grouping for each similarity range r in the embodiment.

FIG. 19 is a diagram showing classification data in the embodiment.

FIG. 20 is a flowchart showing another generation procedure of the first index in the embodiment.

FIG. 21 is a view showing a result of a graph display for each similarity range r in the embodiment.

FIG. 22 is a view showing the result of the number calculation for each similarity range r in the embodiment.

FIG. 23 is a diagram showing a calculation result of the similarity of each case in the embodiment.

FIG. 24 is a diagram showing a schematic configuration of a conventional similarity search device.

FIG. 25 is a diagram showing a schematic configuration of another conventional similarity search device.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 101 ... Input device 102 ... Input / output control part 103 ... Inference processing part 104 ... First search part 105 ... First index generation part 106 ... Data classification part 107 ... Target similarity generation part 108 ... Classification data storage part 109 ... Database 110 … Similarity evaluation knowledge 111… Second search unit

────────────────────────────────────────────────── ─── Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G06F 17/30 G06F 9/44 JICST file (JOIS)

Claims (6)

(57) [Claims] 1. A data storage means for storing a plurality of data represented by a plurality of sets of attributes and attribute values, and a similarity storing a hierarchical structure of the attribute values and similarity evaluation knowledge indicating a similarity between the attribute values. Based on the evaluation knowledge storage means, based on the similarity evaluation knowledge, a plurality of data stored in the data storage means are used to determine the range of attribute values for each attribute and the importance of each attribute given from the outside. Data classifying means for classifying according to the second index shown; target similarity generating means for generating a target similarity from the data classification result by the data classifying means; First index generating means for generating a first index indicating a range, using the first index, for a plurality of data stored in the data storage means A first search means for performing a search to narrow down the data amount; and a second search means for extracting data similar to data having a specific attribute value from the narrowed data using the second index. A similarity search device characterized by comprising: 2. A data storage means storing a plurality of data represented by a plurality of sets of attributes and attribute values, and a similarity storing a hierarchical structure of the attribute values and similarity evaluation knowledge indicating a similarity between the attribute values. Based on the evaluation knowledge storage means, based on the similarity evaluation knowledge, a plurality of data stored in the data storage means are used to determine the range of attribute values for each attribute and the importance of each attribute given from the outside. Data classifying means for classifying according to the second index shown; classification result presenting means for presenting the data classification result to a user; First index generating means for generating a first index to be shown, using the first index, search for a plurality of data stored in the data storage means First search means for narrowing the data amount by using the second index, and second search means for extracting data similar to data having a specific attribute value from the narrowed data using the second index. A similarity search device characterized by the following. 3. The similarity search device according to claim 1, further comprising target similarity changing means for interactively changing the target similarity. 4. A data storage means for storing a plurality of data represented by a plurality of sets of attributes and attribute values, and a similarity storing a hierarchical structure of the attribute values and similarity evaluation knowledge indicating a similarity between the attribute values. Based on the evaluation knowledge storage means, based on the similarity evaluation knowledge, a plurality of data stored in the data storage means are used to determine the range of attribute values for each attribute and the importance of each attribute given from the outside. Data classification means for classifying according to the second index shown, first index generation means for generating a first index indicating the range of attribute values of each attribute from the data classification result by the data classification means, and using the first index A first search unit that searches a plurality of data stored in the data storage unit to narrow down the data amount; and There, the similarity search apparatus characterized by comprising a second search means for extracting data similar to data having a specific attribute value from among the narrowed-down data. 5. The similarity search apparatus according to claim 1, further comprising a first index changing means for interactively changing the first index. 6. The similarity search apparatus according to claim 1, further comprising means for storing the data classification result.