JP2010152692A - Similarity calculation apparatus, similarity calculation method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a similarity calculation apparatus and a similarity calculation method that can prevent degradation of reliability of a similarity calculation result in calculating the similarity between data having missing values and other data. <P>SOLUTION: The similarity calculation apparatus includes a reception means for receiving first and second data having an area related to each attribute in which an attribute value for the attribute related to the area is or is not described, a storage means for storing distribution information about each attribute indicating the distribution of possible values of the attribute value for the attribute, a first calculation means for calculating the similarity between attribute values for attributes for which the attribute value is described in both first and second data, and as for attributes for which the attribute value is missing in at least either of the first and second data, calculating the similarity expectation between attribute values from the distribution information, and a second calculation means for calculating the similarity between the first and second data from the similarities and similarity expectations calculated by the first calculation means. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a similarity calculation device, a similarity calculation method, and a program.

  Data having a plurality of attribute data (hereinafter referred to as “attribute values”) one-to-one corresponding to each of a plurality of attributes (also referred to as items) is used in various fields.

  For example, vehicle state data having a plurality of attribute values collected from a plurality of sensors of an automobile has a vehicle speed, an engine speed, a shift position, and the like as attributes. In addition, among these attributes, there is an attribute whose attribute value is not acquired (an attribute whose attribute value is missing) depending on the type, year, grade, etc. of the vehicle from which data is acquired, or a sensor Due to the above problem, some attribute values in the vehicle state data may be lost.

  As a method of calculating the similarity between data in which some of the attribute values are missing, the missing attribute value (hereinafter referred to as “missing value”) itself is a representative value, for example, There is a method of calculating the similarity between the data complemented with 0 or an average value and the missing value supplemented.

Patent Document 1 uses an attribute value corresponding to an attribute other than an attribute having a missing value (hereinafter referred to as a “specific attribute”) to resemble data having a missing value from data having no missing value. A method is described in which similar data is obtained, an attribute value corresponding to a specific attribute is specified from the attribute values of the similar data, and the missing value is complemented with the specified attribute value.
JP 2002-215646 A

  When the missing value is complemented with the representative value, there is a problem that the similarity is biased when calculating the similarity between the data.

  Moreover, if the similarity between the data which complemented the missing value using similar data and the similar data is calculated, the data will be unfairly similar. For this reason, the method of complementing a missing value using highly similar data is not a suitable method for obtaining the similarity between data having a missing value and other data.

  Therefore, in the calculation of the similarity between data having missing values and other data, there is a problem that the reliability of the calculation result of the similarity is lowered.

  The objective of this invention is providing the similarity calculation apparatus, similarity calculation method, and program which can solve the said subject.

  In the similarity calculation device of the present invention, an attribute value corresponding to an attribute associated with the region is described in each region associated with each predetermined attribute on a one-to-one basis, or the attribute value is described. Receiving means for receiving the first and second data that has not been performed, storage means for storing distribution information representing a distribution of values that can be taken by the attribute value corresponding to the attribute, for each attribute, and the attribute value For the attribute described in both the first and second data, the similarity between attribute values corresponding to the attribute is calculated, and the attribute value is at least one of the first and second data. About the attribute which does not exist, it calculated by the 1st calculation means which calculates the expected value of the similarity degree of the attribute value corresponding to the said attribute using the distribution information in the said storage means, and the said 1st calculation means Similarity and Based on the expected value of the similarity, and a second calculating means for calculating a similarity between the first and second data.

  The similarity calculation method of the present invention is a similarity calculation method performed by the similarity calculation device, and corresponds to each region associated with each predetermined attribute in a one-to-one correspondence with the attribute associated with the region. Distribution information representing a distribution of values that can be taken by the attribute value corresponding to the attribute, for each attribute, accepting first and second data in which the attribute value is described or the attribute value is not described Is stored in the storage means, and for the attribute whose attribute value is described in both the first and second data, the similarity between attribute values corresponding to the attribute is calculated, and the attribute value is at least For an attribute that does not exist in one of the first and second data, an expected value of similarity between attribute values corresponding to the attribute is calculated using the distribution information in the storage unit, and the calculated Similarity and Based on the expected value of similarity score to calculate the similarity between the first and second data.

  According to the program of the present invention, an attribute value corresponding to an attribute associated with the area is described in each area associated with each predetermined attribute on the computer, or the attribute value is described. Receiving processing for receiving the first and second data that has not been performed, storage processing for storing distribution information representing a distribution of values that can be taken by the attribute value corresponding to the attribute for each attribute in the storage means, For an attribute whose attribute value is described in both the first and second data, the similarity between attribute values corresponding to the attribute is calculated, and the attribute value is at least the first and second data. For the attribute that does not exist in one of the above, the first calculation process that calculates the expected value of the similarity between the attribute values corresponding to the attribute using the distribution information in the storage means, and the calculated similarity And based on the similarity of the expected value, a second calculation process of calculating a similarity between the first and second data, to the execution.

  According to the present invention, in the calculation of the similarity between data having missing values and other data, it is possible to prevent a decrease in reliability of the calculation result of the similarity.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(First embodiment)
FIG. 1 is a block diagram showing a similarity calculation apparatus 100 according to the first embodiment of this invention.

  Similarity calculation apparatus 100 includes an input processing unit 110, a data distribution storage unit 120, and calculation units 130 and 140. The calculation unit 130 includes an attribute similarity calculation processing unit 130a and an expected value calculation processing unit 130b. The calculation unit 140 includes a similarity integration processing unit 140a and an output processing unit 140b.

  Input processing unit 110 can generally be referred to as accepting means.

  The input processing unit 110 has a function of inputting data that is a target for calculating the similarity between data (hereinafter referred to as “calculation target data”) or distribution information representing a data distribution to be stored in the data distribution storage unit 120. Have.

  The input processing unit 110 receives first data and second data as calculation target data.

  Each of the first data and the second data has each area associated with each predetermined attribute on a one-to-one basis. In each area, an attribute value corresponding to an attribute associated with the area is described, or the attribute value is not described.

  FIG. 2 is an explanatory diagram showing an example of the first data, and FIG. 3 is an explanatory diagram showing an example of the second data. 2 and 3, the vehicle speed, engine speed, and shift position are used as attributes. The first and second data are not limited to those shown in FIGS. 2 and 3 and can be changed as appropriate.

  In the first data shown in FIG. 2, the area 2a associated with the attribute “speed” describes “AAAA” as the attribute value corresponding to “speed”. In the area 2b associated with the attribute “engine speed”, “XXXX” that is an attribute value corresponding to “engine speed” is described. Further, in the area 2c associated with the attribute “shift position”, the attribute value corresponding to the “shift position”, that is, no value is described.

  In the second data shown in FIG. 3, the attribute value “BBBB” corresponding to “speed” is described in the area 3a associated with the attribute “speed”. In addition, in the area 3b associated with the attribute “engine speed”, an attribute value corresponding to “engine speed”, that is, no value is described. In addition, in the area 3c associated with the attribute “shift position”, the attribute value corresponding to “shift position”, that is, no value is described.

  Further, the input processing unit 110 receives distribution information representing a distribution of values (for example, probability distribution or probability density function) that can be taken by the attribute value corresponding to the attribute for each attribute of the first and second data. .

  The input processing unit 110 is not limited to a device that directly receives data from a human such as a keyboard, but may be an interface connected to an external system or the like.

  Data distribution storage unit 120 can be generally referred to as storage means.

  The data distribution storage unit 120 has a function of storing distribution information for each attribute input from the input processing unit 110. However, the distribution information may be stored in advance without being input from the input processing unit 110.

  Calculation unit 130 can be generally referred to as first calculation means.

  For the attribute whose attribute value exists in both the first and second data (for example, “speed” in FIGS. 2 and 3), the calculation unit 130 calculates the similarity between the attribute values corresponding to the attribute. To do.

  Further, the calculation unit 130 stores data distribution for an attribute whose attribute value does not exist in at least one of the first and second data (for example, “engine speed” and “shift position” in FIGS. 2 and 3). Using the distribution information in the unit 120, the expected value of the similarity between attribute values corresponding to the attribute is calculated.

  For example, for each attribute, the calculation unit 130 determines whether an attribute value corresponding to the attribute exists in both the first data and the second data. When the attribute value exists in both, the calculation unit 130 calculates the similarity between the attribute values. When the attribute value does not exist in at least one of the first and second data, the calculation unit 130 stores the attribute value in the data distribution storage unit 120. Using the distribution information, an expected value of similarity between attribute values corresponding to the attribute is calculated.

  The inter-attribute similarity calculation processing unit 130a calculates the similarity between the attribute values corresponding to the attribute having the attribute value in both the first and second data.

  The expected value calculation processing unit 130b uses the distribution information in the data distribution storage unit 120 for attributes whose attribute values do not exist in at least one of the first and second data, and the similarity between attribute values corresponding to the attributes Calculate the expected value of degree.

  Calculation unit 140 can generally be referred to as second calculation means.

  The calculation unit 140 calculates the similarity between the first and second data based on the similarity calculated by the calculation unit 130 and the expected value of the similarity.

  For example, the calculation unit 140 adds the similarity calculated by the calculation unit 130 and the expected value of the similarity, and sets the addition result as the similarity between the first and second data.

  The similarity integration processing unit 140a integrates the calculation results for each attribute in the calculation unit 130, and calculates the similarity between the first and second data.

  The output processing unit 140b has a function of outputting the inter-data similarity calculated by the similarity integration processing unit 140a. The output processing unit 140b may be a display, audio output, an interface connected to an external system, or the like.

  Next, the operation will be described.

  FIG. 4 is a flowchart for explaining the operation of the similarity calculation apparatus 100. Hereinafter, the operation of the similarity calculation apparatus 100 will be described with reference to FIGS. 1 and 4.

  First, the input processing unit 110 receives first and second data for which similarity is to be calculated (step A1).

Hereinafter, for the sake of explanation, the first and second data are expressed by vectors x ₁ and x ₂ , respectively.

However, the j-th attribute of x ₁ and x ₂ represents the same attribute, and attribute values corresponding to the j-th attribute are expressed as x _1j and x _2j .

  If there is no attribute value corresponding to the jth attribute in the input data, that is, if there is no value in the area related to the jth attribute (for example, mull), the attribute value is Treated as missing values.

Then, calculation section 130, for each attribute of x ₁ and x _2, repeats the process of step A6 from step A2 shown below. In the following, processing regarding the jth attribute will be described.

First, the attribute similarity calculation processing unit 130a confirms whether at least one of x _1j and x _2j is missing (step A2).

When neither x _1j nor x _2j is missing (step A3), the attribute similarity calculation processing unit 130a is based on a calculation formula for calculating the attribute similarity stored in itself. , Similarity between attributes of x _1j and x _2j is calculated (step A4).

Here, the similarity between attributes is represented by an arbitrary function calculated with respect to x _1j and x _2j , and the similarity between attributes regarding the j-th attribute is expressed as g _j (x _1j , x _2j ).

  The similarity between attributes may take a larger value as the data is similar, or may take a smaller value as the data is similar.

As the similarity between attributes, for example, when the attribute value corresponding to the jth attribute takes a continuous value, the square distance (x _1j −x _2j ) ² or the absolute value distance | x _1j −x _2j | If the attribute value corresponding to the jth attribute takes a discrete value, the Hamming distance (0 if x _1j and x _2j match, 1 if not, etc.) It is possible to use.

When x _1j or x _2j is missing (step A3), the attribute similarity calculation processing unit 130a cannot directly calculate g _j (x _1j , x _2j ). Therefore, the attribute similarity calculation processing unit 130a operates the expected value calculation processing unit 130b. The expected value calculation processing unit 130b reads the distribution information related to the jth attribute from the data distribution storage unit 120, and uses the probability distribution (or probability density function) p _j represented by the distribution information to calculate the expected value of the similarity between attributes. calculate. The expected value calculation processing unit 130b provides the calculation result to the attribute similarity calculation processing unit 130a (step A5).

  For this reason, the attribute similarity calculation processing unit 130a has an attribute similarity or an expected value of the attribute similarity for each attribute.

Here, as p _j , the attribute value corresponding to the j-th attribute takes a discrete value or a symbol value (a value representing a variable value as a symbol such as a symbol instead of an actual value such as an integer). For example, a discrete value distribution such as a multinomial distribution can be used. When the jth attribute takes a real value, a continuous value distribution such as a normal distribution can be used. is there.

  Hereinafter, a specific procedure for calculating the expected value will be described.

First, when only one side of x _1j and x _2j is missing, the expected value calculation processing unit 130b uses the similarity between attributes according to the equation (1) in order to use the missing attribute value. Calculate the expected value of degree.

Note that equation (1) is a calculation equation used when only x _2j is missing, but when only x _1j is missing, the expected value calculation processing unit 130b determines that equation (1) By replacing x _1j with x _2j , the expected value of the similarity between attributes can be calculated by the same method.

  In Equation (1), as x, all values that can be taken by the attribute value corresponding to the j-th attribute are used.

Next, when both x _1j and x _2j are missing, the expected value calculation processing unit 130b cannot use the attribute value of the data. Calculate the expected value of similarity between attributes.

  In Expression (2), as x and y, all values that can be taken by the attribute value corresponding to the j-th attribute are used.

  Equations (1) and (2) include the case where the attribute value corresponding to the jth attribute takes a continuous value, and therefore integration is used as the expected value calculation method, but this corresponds to the jth attribute. When the attribute value to be taken takes a discrete value, the expected value calculation processing unit 130b may calculate the expected value of the similarity between attributes by replacing the integral with the sum.

  Note that the expected value according to the equation (1) can be calculated in advance when the attribute value corresponding to the jth attribute takes a finite number of discrete values. The calculation processing unit 130b or the data distribution storage unit 120 may store the value in advance.

  Since the expected value calculation according to the equation (2) can be calculated in advance regardless of the type of the attribute value corresponding to the jth attribute, the expected value calculation processing unit 130b or the data distribution storage unit 120 May be stored in advance.

Regarding the calculation of the expected value, the case where the attribute value corresponding to the jth attribute takes a discrete value, uses the Hamming distance as g _j (x _1j , x _2j ), and uses the multinomial distribution as p _j . The expression (3) can be used as the expression (1), and the expression (4) can be used as the expression (2).

In Equation (3), x _1j is the kth discrete value of the discrete values that can be taken by the attribute value corresponding to the jth attribute, and p ^k _j is the attribute value corresponding to the jth attribute. Let k denote the probability that the k-th discrete value will appear.

  When the inter-attribute similarity calculation processing unit 130a obtains the inter-attribute similarity or the expected value of the inter-attribute similarity for the j-th attribute, the inter-attribute similarity calculation processing unit 130a checks whether there is the remaining attribute. Updates the attribute to be calculated and returns the process to step A2. If there are no remaining attributes, the process proceeds to step A7 (step A6).

  In step A7, the similarity integration processing unit 140a integrates the similarity between attributes or the expected value for each attribute calculated in steps A2 to A6, thereby calculating the similarity between the first and second data. calculate.

In this embodiment, the similarity integration processing unit 140a calculates the first and first values according to equation (5), where s _j is the similarity between attributes or the expected value for each attribute calculated in steps A2 to A6. It is possible to define the similarity between two data as an arbitrary function regarding s _j .

Here, f is an arbitrary function, and d is the number of attributes that x ₁ and x ₂ have.

  As a specific example of f, Expression (6), which is the sum of similarity between attributes of each attribute, can be given.

It is also possible to use any nonlinear function related to s _j as f.

  The output processing unit 140b outputs the similarity between data calculated in step A7 (step A8).

  Next, the effect of this embodiment will be described.

  According to the present embodiment, for an attribute whose attribute value exists in both the first and second data, the calculation unit 130 calculates the similarity between attribute values corresponding to the attribute, and the attribute value is at least For an attribute that does not exist in one of the first and second data, the distribution information in the data distribution storage unit 120 is used to calculate the expected value of the similarity between the attribute values corresponding to the attribute.

  The calculation unit 140 calculates the similarity between the first and second data based on the similarity calculated by the calculation unit 130 and the expected value of the similarity.

  For this reason, even if one or both of the two data for which the similarity is calculated has a missing value, the distribution of possible values of the missing attribute value is used to determine the attribute An expected value of similarity is calculated, and the similarity between the first and second data is calculated using the expected value.

  Therefore, even when one or both of the two data whose similarity is calculated has a missing value, it is possible to calculate the similarity between data with high reliability.

  In the present embodiment, the calculation unit 140 adds the similarity calculated by the calculation unit 130 and the expected value of the similarity, and uses the addition result as the similarity between the first and second data.

  In this case, the similarity between the first and second data can be easily calculated.

(Second Embodiment)
Next, a similarity calculation apparatus 200 according to the second embodiment of this invention will be described.

  FIG. 5 is a block diagram showing the similarity calculation device 200. In FIG. 5, the same components as those shown in FIG.

  In addition to the configuration of the similarity calculation apparatus 100 shown in FIG. 1, the similarity calculation apparatus 200 includes an input processing unit 210 that inputs data for learning the distribution, and a data distribution learning processing unit for learning the data distribution. 220.

  Input processing unit 210 can generally be referred to as learning data receiving means.

  The input processing unit 210 has a plurality of learning objects in which attribute values corresponding to the attributes associated with each region are described in each region associated with each attribute on a one-to-one basis, or the attribute values are not described. Accept data.

  Data distribution learning processing unit 220 can be generally referred to as processing means.

  The data distribution learning processing unit 220 generates distribution information for each attribute based on the plurality of learning data received by the input processing unit 210 and stores the distribution information in the data distribution storage unit 120.

  Since the similarity calculation device 200 includes an input processing unit 210 and a data distribution learning processing unit 220, it is not always necessary to store distribution information in advance, and if the distribution of data can be calculated in advance, the distribution It is possible to input data for learning (learning data), calculate the distribution from the data, learn it, and store it.

  The operation of the similarity calculation device 200 is substantially the same as the operation of the similarity calculation device 100 shown in FIG. 4, but the expected value calculation processing unit 130b calculates the expected value of the similarity between attributes of a single attribute Step A5 Only the difference.

  In the similarity calculation device 100, the distribution information stored in advance was used when calculating the expected value of the similarity between attributes of a single attribute, but in the similarity calculation device 200, it was stored by the operation shown in FIG. The expected value of similarity between attributes of a single attribute can be calculated using the distribution information.

  Referring to FIG. 6, first, the input processing unit 210 receives a plurality of learning data (step B1).

  The data distribution learning processing unit 220 calculates distribution information for each attribute based on a plurality of learning data. At this time, the distribution to be calculated is, for example, a multinomial distribution with discrete values, but those that cannot be calculated with continuous values or the like are stored in advance in the data distribution storage unit 120 as in the similarity calculation device 100. This can be handled (step B2).

  The process of step B2 is performed for all attributes.

  The data distribution learning processing unit 220 stores the distribution information calculated for all attributes in the data distribution storage unit 120 (step B3).

  In the present embodiment, the data distribution learning processing unit 220 generates distribution information for each attribute based on a plurality of learning data received by the input processing unit 210 and stores the distribution information in the data distribution storage unit 120. .

  For this reason, it is not always necessary to store the distribution information in advance, and if the distribution information can be calculated in advance, the distribution information can be calculated, learned and stored using learning data. Become.

  In each of the above embodiments, when calculating the degree of similarity between data having a plurality of attribute values collected from a plurality of sensors of an automobile, or a market prediction or a financial related prediction system using various input attributes, It can be applied to various uses such as state similarity calculation of various computer-related devices.

  The similarity calculation devices 100 and 200 are not only realized by dedicated hardware, but also include an input processing unit 110, a data distribution storage unit 120, a calculation unit 130 and a calculation unit 140, an input processing unit 210, and data A program for realizing each function of the distribution learning processing unit 220 may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read and executed by a computer system. .

  The computer-readable recording medium refers to a recording medium such as a flexible disk, a magneto-optical disk, a CD-ROM (Compact Disk Read Only Memory), and a storage device such as a hard disk device built in the computer system.

  Furthermore, a computer-readable recording medium is a server that dynamically holds a program (transmission medium or transmission wave) for a short period of time, as in the case of transmitting a program via the Internet, and a server in that case. Some of them hold programs for a certain period of time, such as volatile memory inside computer systems.

  In each embodiment described above, the illustrated configuration is merely an example, and the present invention is not limited to the configuration.

It is the block diagram which showed the similarity calculation apparatus 100 of the 1st Embodiment of this invention. It is explanatory drawing which showed an example of 1st data. It is explanatory drawing which showed an example of 2nd data. 6 is a flowchart for explaining the operation of the similarity calculation device 100. 2 is a block diagram showing a similarity calculation device 200. FIG. 5 is a flowchart for explaining a part of the operation of the similarity calculation device 200.

Explanation of symbols

100 Similarity Calculation Device 110 Input Processing Device 120 Data Distribution Storage Device 130 Calculation Unit 130a Inter-attribute Similarity Calculation Processing Unit 130b Expected Value Calculation Processing Unit 140 Calculation Unit 140a Similarity Integration Processing Unit 140b Output Processing Unit 200 Similarity Calculation Device 210 Input processing unit 220 Data distribution learning processing unit

Claims (9)

First and second attribute values corresponding to the attribute associated with the region are described in each region associated with each predetermined attribute on a one-to-one basis, or the attribute value is not described. Receiving means for receiving data; and
Storage means for storing, for each attribute, distribution information representing a distribution of values that can be taken by the attribute value corresponding to the attribute;
For an attribute whose attribute value is described in both the first and second data, a similarity between attribute values corresponding to the attribute is calculated, and the attribute value is at least the first and second data. For attributes that do not exist in one of the data, using the distribution information in the storage means, a first calculation means for calculating the expected value of the similarity between the attribute values corresponding to the attribute;
And a second calculation means for calculating the similarity between the first and second data based on the similarity calculated by the first calculation means and the expected value of the similarity. The similarity calculation apparatus according to claim 1,
In each region, an attribute value corresponding to an attribute associated with the region is described, or learning data receiving means for receiving a plurality of learning data in which the attribute value is not described,
A similarity calculation device further comprising: processing means for generating the distribution information for each of the attributes based on the plurality of learning data and storing the distribution information in the storage means. In the similarity calculation apparatus according to claim 1 or 2,
The second calculation means adds the similarity calculated by the first calculation means and the expected value of the similarity, and sets the addition result as the similarity of the first and second data. Computing device. A similarity calculation method performed by the similarity calculation device,
First and second attribute values corresponding to the attribute associated with the region are described in each region associated with each predetermined attribute on a one-to-one basis, or the attribute value is not described. Accept data,
For each of the attributes, distribution information representing a distribution of values that can be taken by the attribute value corresponding to the attribute is stored in the storage unit,
For an attribute whose attribute value is described in both the first and second data, a similarity between attribute values corresponding to the attribute is calculated, and the attribute value is at least the first and second data. For attributes that do not exist in one of the data, the distribution information in the storage means is used to calculate the expected value of the similarity between the attribute values corresponding to the attribute,
A similarity calculation method for calculating a similarity between the first data and the second data based on the calculated similarity and an expected value of the similarity. The similarity calculation method according to claim 4,
In each of the areas, an attribute value corresponding to an attribute associated with the area is described, or a plurality of learning data in which the attribute value is not described is received,
Generating the distribution information for each of the attributes based on the plurality of learning data,
A similarity calculation method of storing the generated distribution information in the storage means when storing the distribution information. The similarity calculation method according to claim 4 or 5,
When calculating the similarity between the first and second data, the calculated similarity and the expected value of the similarity are added, and the addition result is calculated as the similarity between the first and second data. A similarity calculation method. On the computer,
First and second attribute values corresponding to the attribute associated with the region are described in each region associated with each predetermined attribute on a one-to-one basis, or the attribute value is not described. A reception process for receiving data;
For each attribute, a storage process for storing distribution information representing a distribution of values that can be taken by the attribute value corresponding to the attribute in a storage unit;
For an attribute whose attribute value is described in both the first and second data, a similarity between attribute values corresponding to the attribute is calculated, and the attribute value is at least the first and second data. For an attribute that does not exist in one of the data, using the distribution information in the storage means, a first calculation process that calculates an expected value of similarity between attribute values corresponding to the attribute;
A program for executing a second calculation process for calculating a similarity between the first and second data based on the calculated similarity and an expected value of the similarity. The program according to claim 7,
In the computer,
In each area, an attribute value corresponding to an attribute associated with the area is described or a learning data reception process for receiving a plurality of learning data in which the attribute value is not described;
A generation process for generating the distribution information for each of the attributes based on the plurality of learning data;
In the storage process, a program for storing the generated distribution information in the storage unit. In the program according to claim 7 or 8,
When calculating the similarity between the first and second data, the calculated similarity and the expected value of the similarity are added, and the addition result is calculated as the similarity between the first and second data. Program.

Images