JPH0823869B2 - Database similarity search method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a database similarity search method for constructing a case base search on a relational database system.

[0002]

2. Description of the Related Art Case-based reasoning
Reasoning (hereinafter abbreviated as CBR) is a new reasoning paradigm that has recently attracted attention in the field of artificial intelligence (Reference: "Case-based reasoning expected to be put to practical use as a new problem solving method" Shigenobu Kobayashi and Kotaro Nakamura. ,
Nikkei AI separate volume 1991 spring issue).

The structure of the CBR is shown in FIG. The case base stores a large number of cases composed of a set of problems solved in the past and a solution method thereof. When a new problem to be solved is given, the case that solves the problem most similar to the problem is retrieved from the case base (search phase). CB
In R, searching for similar cases is an important process. When searching the case base, the exact match (exa
Partial match without ct matching) (p
artistic matching). next,
In order to absorb the difference between the problem expression in the searched similar cases and the problem expression to be solved, a part of the similar cases should be adaptively modified (correction phase) using the transformation rules of the system, and the result should be solved. Present it as a solution to the problem. The solution created in this way is not always correct, so try applying it. If the result is correct, it is stored in the case base as a positive case showing "new case and its solution", and when it fails, the reason for failure is analyzed,
“New problems, erroneous diffraction methods for them, and criteria for avoiding repeated mistakes” are stored in the case base (evaluation / repair phase). In some cases, the CBR system itself may judge the success or failure of the applied result,
The user may make the decision. Thus, in CBR, when a new problem is given, whether the solution succeeds or fails, the result is always added to the case base of the system. If it fails, index it so that it will not make the same mistake twice. Based on this policy, tools for CBR have been proposed (reference: Anita Face for Cases-based Knowledge Acquisition, C. Leesbeck, An Inter.
face for Case-Based Knowl
edge Acquisition, Proc. of
Workshop on Case-Based R
easing, DARPA, 1988). A schematic diagram of this system is shown in FIG. In FIG. 8, 101 is a case attribute vector, 102 is a case base, 103 is a question storage attribute vector, 104 is a partial matching definition part, 10
5 is a weight definition unit, 107 is an attribute similarity calculation unit, and 109 is a total similarity calculation unit. FIG. 9 shows that in the above reference,
Although it does not have the same configuration as this embodiment, it is the same in principle. Many cases represented in the form of attribute vectors are stored in the case base 102, and when the problem to be solved is stored in the problem storage attribute vector 103,
The cases in the case base 102 are fetched one by one, loaded into the case attribute vector 101, and the degree of similarity is calculated. The definition of the partial matching between the corresponding element attributes is defined in the partial matching definition unit 104, and the weight of the element attribute is stored in the weight definition unit 105. When the attribute similarity calculation unit 107 corresponding to each attribute calculates the attribute similarity, the total similarity calculation unit calculates the sum.

As described above, in the conventional case-based system, the case is stored in the data storage unit incorporated in the case-based system itself. In addition, the basic method of case retrieval in the case-based shell is a method (called a Nearest Neighbor method) in which, when a problem is given, individual cases are compared one by one to find similar cases. When a case is represented by a set of attributes, the attribute value problem similarity is defined for each attribute. For example, if there is "language" as one attribute in the case, the degree of similarity is defined as shown in FIG. For example, the similarity between PROLOG and PROLOG is 1, and the similarity between PROLOG and LISP is that both belong to the group AI-LANGUAGE,
It can be read as 0.8. Conventional Nearest N
The algorithm of the eighbor method is as follows.

[0005] 1. Pick up one case.

2. For each attribute, use the similarity definition defined for each attribute to calculate the similarity between the problem and the case,
Then the overall similarity between the problem and its case is calculated. The formula for calculating the overall similarity is as follows.

[0007]

[Equation 1]

3. Repeat 1 and 2 for all cases.

4. The one with the highest overall similarity is selected as the similar name case for the problem.

[0010]

However, the method of storing the case base in the data storage unit built in the case base system itself has the following problems as actual problems.

In many cases, the case-based construction project tries to utilize the existing database more effectively.
In that case, it is useless to create a case base separately from the conventional database, maintenance is difficult, and the feasibility is low.

Even when a development project creates a completely new case base, the case base development, management, and search processes often overlap with those of the conventional database system, and the two are not incompatible. . In addition, various techniques for increasing the search speed that the existing database management system has should be effectively used in case-based search.

When a new case base is created,
If the case-based search / inference project fails or ends, the case-base created will be a waste. If a case base is constructed using a commercial database management system, there is a risk hedging effect that at least the database remains even if the case base project ends.

[0014] Even in a project that started out to build a case-based search system, users often do not want a similar search, but rather an accurate database search. It is best if the case-based or database can do fuzzy or exact searches.

An object of the present invention is to provide a method for efficiently constructing a case base search on a relational database system.

[0016]

The present invention comprises three inventions.

The first aspect of the present invention comprises claims 1, 2, and 3.

According to a first aspect, when an attribute and an attribute value are given, the attribute similarity definition description in which the similarity between the two values of the attribute is defined in advance is referred to, A search condition expression indicating that it must be one of the values of the set of attribute values having the highest similarity to the attribute value, and the attribute having the second highest similarity to the attribute value as the value taken by the attribute. A search condition expression indicating that it must be one of the values of the set of values, and one of the values of the set of attribute values that has the third highest similarity to the attribute value as the value that the attribute takes. It is characterized by having a similarity order attribute value search condition list creating means for creating a list of search condition formulas with similarities in descending order of similarity, such as a search condition formula indicating that it must be Database similarity search method

A second aspect of the present invention is a set of a plurality of attributes and attribute values.
When two cases are represented, a list of search condition expressions is created for each attribute by the similarity order attribute value search condition list creating means described in claim 1, and a list of search condition expressions created for each attribute is created. From the above, one conditional expression is arbitrarily extracted for each attribute, all brute force combinations of the search conditional expressions for each attribute are created, and the search conditional expressions for each attribute of each combination are combined by logical product. Generate a conditional expression,
Further, a value obtained by weighting and averaging the weights of the corresponding attributes defined in advance in the similarity added to the search condition expression in each combination is calculated as the total similarity of the combination, and the value of the total similarity and the above A database similarity search method characterized by having a case similarity condition list output means for arranging and outputting a pair of synthetic conditional expressions in the descending order of total similarity value.

According to a third aspect of the present invention, in a system having a similarity search function on a database in which one case corresponds to one record of the database, a problem represented by a plurality of attributes and attribute values is given, From the case similarity condition list output means described in claim 2, the top pair of the list of pairs of the total similarity and the combination condition search expression output corresponding to the case is extracted, and the combination condition of the top pair is extracted. Convert the search formula into a corresponding database search formula, issue it to the database, search records from the database, present the search result together with the value of the total similarity of the top pair, and do the same thing, The database similarity search method is characterized in that the first pair of the list is taken out one by one and repeated.

According to a second aspect of the present invention, in claim 2, not all the pairs of the value of the overall similarity and the combination conditional expression are output, but the value of the overall similarity has a lower limit of the overall similarity defined in advance. A database similarity search method comprising: a case similarity condition list output unit that selects only pairs that are larger than the boundary value and arranges and outputs them in descending order of total similarity value.

According to a third aspect of the present invention, in claim 3, instead of extracting and processing all the pairs in the list of pairs of the total similarity and the synthetic condition search expression, a database search is performed for each pair. The total number of records is calculated, and if the value of the total exceeds the maximum number of records set in advance, then a pair is taken out from the list of pairs of the total similarity and the composition condition search formula. The database similarity search method is characterized by terminating the process of searching the database.

When constructing a case-based search system on an existing database, records are searched one by one from the database and the above-mentioned Nearest Neighb
It is totally inefficient to implement the or algorithm, so
In the present invention, a completely different algorithm was devised.

Here, each case corresponds to a record in the relational database. The major flow of the algorithm of the present invention is as follows.

1. Generate a search condition expression to search for records similar to the given problem.

2. The retrieval condition formula having the highest degree of similarity is taken out of these retrieval condition formulas, and the database is searched using the retrieval condition formula.

3. When one or more records are retrieved,
It is presented with the corresponding similarity. If a sufficient number of records have been presented, this ends the process. If not, the process returns to 2 and the same operation is repeated for the remaining search condition expressions.

In the conventional method, the value of the similarity between the problem and the case is known only when the similarity between the problem and the individual comparison case is calculated, but in the method of the present invention, the similarity is calculated first. However, the order is greatly different, such as obtaining a condition to obtain the value of the degree of similarity, then performing a database search to obtain individual cases as entities.

[0029]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The first embodiment of the present invention will be described in claim 1,
It consists of a few. It is claim 3 that describes the overall flow of the three claims, but claim 2 is used as a part of the function. Claim 2 uses Claim 1 as a part of the functions. Therefore, as an embodiment, description will be made in the order of claims 1, 2, and 3.

FIG. 1 shows an embodiment of the portion of claim 1 of the first invention. As an example of the attribute similarity definition description, "language"
An example of the case of attributes is shown in FIG. It is assumed that such a definition is defined in advance for each individual attribute.

When an attribute and an attribute value are given, the attribute similarity definition description for the attribute is referred to, a node corresponding to the given attribute value is found from the attribute similarity definition description, and it is used as a parent node. . On the other hand, [empty] is entered as the current node. If there is no parent node, it ends,
If there is a parent node, go to processing A repeatedly.

The processing of the repeated processing A is as follows. Among the end nodes that are descendants of the parent node, including the parent node itself, all items excluding the end nodes that are the current node itself or descendants of the current node are counted. A database search conditional expression that the value taken by the attribute must be one of the values of the counted end nodes is generated. On the other hand, the similarity value of the parent node is checked. The conditional expression and the similarity value are output as a pair. Find the parent of the parent node. If there is no parent node, the iterative process A ends, but if there is a parent, the parent node becomes the current node, the new parent becomes the parent node, and the iterative process A
Do again.

An example will be described. Now PROLO
If G or a value given by the user, the similar conditional expression and the value of similarity are as follows.

The first similar set is PROLOG and the degree of similarity is 1. The second similar set is the AI-LAN excluding PROLOG.
Since it is an element under GUAGE, it is “LISP” and its similarity is 0.8.

The third similar set is PROLOG, LIS
Since it is an element under HL-LANGUAGE except P, "FORTRAN""PASCAL""C""C++"
And its similarity is 0.5.

The fourth similar set is PROLOG, LIS
Since it is an element under ANY-LANGUAGE except P, FORTRAN, PASCAL, C, and C ++, "AS
M ”, and the degree of similarity is 0. Becomes

An embodiment of claim 2 of the first invention is shown in FIG. For each individual attribute, the degree of similarity between two values is defined as shown in FIG. As an example, it is assumed that there are man-hour attribute and machine attribute in addition to the language attribute. It is assumed that values of language attribute, man-hour attribute, and machine attribute are given as input from the user. Then, using the similarity order attribute value search condition list creating means of claim 1, the language attribute, the man-hour attribute,
A list of search condition expressions for each machine attribute is generated. Next, one is taken out from each conditional expression list to create a combination. The element of the conditional expression list of language attribute is N
If there are M elements of the conditional expression list of the man-hour attribute and L elements of the conditional expression list of the machine attribute,
A total of (N × M × L) combinations are generated.

For example, as an example of one combination, "language = LISP (similarity 0.8)" is extracted from the search condition expression list of the language attribute, and "man-hour =" from the search condition expression list of the man-hour attribute. "10 person months (similarity 0.5)" is retrieved, and "machine =
EWS-RISC (similarity 0.9) "is retrieved,
"Language = LISP (similarity 0.8), man-hours = 10 man-months (similarity 0.5), machine = EWS-RISC (similarity 0.9)" is generated. Next, for each combination, a conditional expression that is a combination of the search condition expressions that are the elements within the combination is formed by combining them with a logical product. The above combination would be: "Language = LISP and man-hours = 10 man-months and
Machine = EWS-RISC ”At the same time, the total similarity is calculated for each combination from the similarities added to the search condition formulas that are the elements within the combination. An example of the formula for calculating the total similarity is to calculate a weighted average, and the formula is as follows.

[0039]

[Equation 2]

In the above example, if the language attribute weight is 0.5, the man-hour attribute weight is 0.3, and the machine attribute weight is 0.2, then (0.5 × 0.8) + (0.3 × 0.5) +
(0.2 × 0.9) ÷ (0.5 + 0.3 + 0.2) =
The value is 0.73, and the value of the total similarity is 0.73. When the total similarity is calculated for each combination, the total similarity is arranged and output in descending order of the total similarity value.

An embodiment of claim 3 of the first invention is shown in FIG. Given a problem represented by a plurality of attribute-attribute value pairs, use the case-similar condition list output means of claim 2. A list of pairs of the total similarity and the synthesis condition search expression output corresponding to the above-mentioned problem is generated and arranged in descending order of the value of the total similarity. The head pair of the list is taken out and a database search expression is generated from the synthesis condition search expression. this is,
It can be easily generated using SQL, which is a typical language for relational databases. The compound conditional expression is SQL SELECT-F.
Since it is the ROM-WHERE type WHERE clause itself,
A character string "SELECT * FROM target table" may be prepared in advance, and the combined condition retrieval formulas extracted therefrom may be combined at the character string level. For example, in the above example, "SELECT * FROM target table WHER
E language = LISP and man-hours = 10 man-months and
Machine = EWS-RISC ". As the search expression generated for each attribute, in addition to the expression expressed by an equation such as language = LISP, “language = LISP or PROL”
It may be expressed by a plurality of ORs such as “OG, smalltalk”. At this time, an SQL expression corresponding to this expression is generated. In this example, "language in
(LISP, PROLOG, smalltalk) ". In some cases, the range may be designated as “10 <man-hours <20”. Also at this time, S corresponding to this expression
Generate a QL expression. In this example, “man-hour betw
een 10 and 20 ". Since these conversions can be easily made one-to-one, the method does not require any new technique. Next, the generated SQL expression is issued to the database, and the record is retrieved from the database. The search result is presented together with the value of the total similarity. If the synthesis condition search expression list is empty, the process ends. If not, the next top element is extracted from the synthesis condition search expression list and the above process is repeated.

A second embodiment of the present invention will be described. The second aspect of the present invention is an extension of claim 2 of the first aspect of the present invention. Even if the present invention is replaced with the contents of claim 2, the function of the first present invention is not impaired, and the function expanded by the present invention is also reflected in the first present invention.

The extended portion in the second embodiment of the present invention is the total similarity defined in advance in the combination of the synthetic condition retrieval formula generated in claim 2 of the first present invention and the total similarity. Select only pairs that are large compared to the lower bound of degrees,
It means throwing away the things that are not so. By adding this function, it is not necessary to issue a SQL expression having a low overall similarity to search for a case where the similarity is not so high as to waste the calculation.

The third aspect of the present invention is the claim 3 of the first aspect of the present invention.
Is an extension of. The expanded part in the third embodiment of the present invention is to monitor the result of the database search in claim 3 of the first present invention and calculate the total number of the searched records, and define it in advance. If the maximum number of records is exceeded, search and display of results will be stopped there.

In many cases, the case-based project intends to utilize the existing database more effectively. In that case, the present invention can be used to share a case base with a conventional database, which has the effect of development cost, maintenance cost, and consistency.

Even if a development project creates a completely new case base, by using the present invention, various techniques for speeding up the search that a conventional database system has can be used for the case base development / management / search processing. There is an effect that it can be used as it is.

[Brief description of drawings]

1 an embodiment of the first claim part of the present invention, FIG.

FIG. 2 is an embodiment of the portion of claim 2 of the first invention.

FIG. 3 is an embodiment of the part of claim 3 of the first invention.

FIG. 4 is an embodiment of the second invention.

FIG. 5: Embodiment of the third invention

FIG. 6 is a definition diagram of similarity between attribute values.

FIG. 7 is a block diagram of case-based reasoning (CBR).

FIG. 8 is a schematic diagram of a CBR system.

[Explanation of symbols]

 101 Case Attribute Vector 102 Case Base 103 Problem Storage Attribute Vector 104 Partial Matching Definition Section 105 Weight Definition Section 107 Attribute Similarity Calculation Section 109 Total Similarity Calculation Section

 ─────────────────────────────────────────────────── ─── Continuation of the front page (56) References The Society for Artificial Intelligence Research Material SIG-KBS-9102-7 (1991) P. 57-97 AAAI-92 (1992) p. 843-849 Proceedings of the 45th Conference of the Information Processing Society of Japan (Late 1992) 2nd p. 175-178 Proceedings of the 44th Annual Conference of IPSJ (the first half of 1992) 2nd p. 49-52 Proceedings of Information Processing Society of Japan 44th Annual Meeting (Past 1992) 3P. 241-242 Proceedings of the 43rd IPSJ National Conference (Late 1991) 2 p. 95-96

Claims (5)

[Claims] 1. A Given certain attributes and attribute values, and previously referring to, the value is the attribute value of the attribute similarity is defined attribute similarity definition description between two of said attribute top
Is a search condition expression indicating that the attribute value belongs to a set of attribute values having a high degree of similarity, and the value of the attribute has the second highest degree of similarity to the attribute value.
Search condition expression indicating that the attribute belongs to a set of attribute values
A collection of attribute values whose sex values have the third highest degree of similarity to the attribute values.
Database, wherein the search condition indicating that they belong to the case, a high order of similarity as, by having a similarity order attribute value search list creating means for creating a list of search conditional expression are indicated by the similarity Similar search method. 2. When one case is represented by a set of a plurality of attributes and attribute values, a list of search condition expressions is created by the similarity-ordered attribute value search condition list creating means for each attribute, and the attribute is created. From the list of search condition expressions created for each attribute, one conditional expression is arbitrarily extracted for each attribute, all brute force combinations of the search condition expressions for each attribute are created, and the search condition expression for each attribute in the combination is created. A conditional expression that is composed by combining with the logical product is generated, and the weighted average value of the weights of the corresponding attributes defined in advance to the similarity added to the search conditional expression in each combination is calculated. A database having case similarity condition list output means for calculating a total similarity of a combination and arranging and outputting a pair of the value of the total similarity and the synthetic conditional expression in descending order of the value of the total similarity. Similar search method. 3. In a system having a similarity search function on a database in which one case corresponds to one record in the database, if a problem represented by a plurality of attributes and attribute values is given, the case similarity is obtained. From the condition list output means, the top pair of the list of pairs of the total similarity and the combined condition search expression output corresponding to the case is taken out, and the combined condition search expression of the top pair is converted into the corresponding database search expression. Convert it, publish it to the database, retrieve the record from the database, present the search result together with the value of the total similarity of the top pair, and do the same thing for one top pair in the list. A database similarity search method characterized in that it is repeated while fetching each. 4. Not outputting all the pairs of the total similarity value and the synthetic conditional expression, but only the pairs whose total similarity value is larger than the previously defined total similarity lower limit boundary value. 3. The database similarity search method according to claim 2, further comprising: case similarity condition list output means that selects and outputs the arranged case similarity condition lists in descending order of total similarity value. 5. The total number of records obtained by searching the database for each pair is calculated instead of extracting and processing all the pairs of the list of pairs of the total similarity and the synthetic condition search expression, If the value of the cumulative total exceeds a predetermined maximum number of records, a process for extracting a pair from the list of pairs of the total similarity and the synthetic condition search expression and terminating the database is terminated. The database similarity search method according to claim 3.