Quantized affinity diagram implementation device and program
The quantified affinity diagramming method uses a matrix-based approach to standardize affinity diagramming results, providing consistent problem identification and solution derivation through eigenvalue and score vector analysis.
Patent Information
- Application Number
- JP2023079260
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-12
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2043-05-12
AI Technical Summary
Affinity diagramming results vary depending on the practitioner due to complex linguistic data interpretation, leading to inconsistent problem identification and solution derivation.
A quantified affinity diagramming method using an n-dimensional square symmetric matrix to calculate eigenvalues and score vectors, ensuring consistent grouping and correlation analysis of linguistic data through 2D and 3D graphical representations.
Ensures consistent results by quantifying affinity and correlation, facilitating standardized problem identification and solution derivation across different practitioners.
Smart Images

Figure 2024163543000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a quantitative affinity diagramming method execution device and a quantitative affinity diagramming method execution program. [Background technology]
[0002] Affinity diagramming is a method of taking facts, presumptions, predictions, ideas, opinions, etc. as linguistic data from a vague and unclear state, such as unknown or unexperienced fields or future problems, and consolidating the linguistic data through their mutual affinity to clarify the structure (see non-patent literature). Affinity diagramming is a method that aims to clarify a problem by setting a vague issue as a theme and creating an integrated diagram for the set theme, and is used to derive solutions to unclear issues such as unknown and unexperienced issues. The Quality Management Glossary (Editor-in-Chief Yoshizawa Tadashi, 2004, Japanese Standards Association) also defines affinity diagramming as "a method of clarifying the location and form of the problem to be solved by taking facts, opinions, ideas as linguistic data for vague and unclear issues such as future or unknown problems, and creating an integrated diagram through their mutual affinity."
[0003] Affinity diagramming generally proceeds as follows: (1) Setting the theme (2) Collection of linguistic data (text data) such as facts, opinions, and ideas related to the theme (3) Creating cards containing language data (4) Grouping cards containing related linguistic data together (creating primary groups) (card collection) (5) Grouping primary groups with affinity (creating secondary groups) (6) Arrangement of primary and secondary groups (e.g., placing related items close together) (7) Clarification of the relationships between groups (causation, correlation, contradiction) (solid lines, single-sided arrow lines, and double-sided arrow lines are generally used)
[0004] However, especially when the linguistic data contains complex content (i.e., content with two or more meanings), the groupings may differ depending on the person using the affinity diagram, which will naturally lead to different results (the location and form of the problem to be solved). [Prior art documents] [Patent documents]
[0005] [Non-Patent Document 1] Yoshinobu Naya (ed.), "New Seven QC Tools", 1st edition, Japan Standards Association, April 20, 1987, pp.29-47 Summary of the Invention [Problem to be solved by the invention]
[0006] The object of this invention is to make the results of affinity diagramming consistent depending on the practitioner.
[0007] The invention also aims to aid in the consideration of problems, including the consideration of potential events. [Means for solving the problem]
[0008] The quantification affinity diagramming execution device of the present invention comprises language data input means for receiving input of n (n is an integer of 3 or more) pieces of language data, display means for displaying all combinations of two mutually different language data i, j (i ≠ j) using the received n pieces of language data, affinity input means for receiving input of a numerical value representing an affinity k for each of the displayed combinations of two language data i, j, matrix creation means for creating data representing an n-th order square symmetric matrix with n rows and n columns, in which the affinity k of the combination of language data i, j input by the affinity input means is set as the element of corresponding element i, j of the n-th order square symmetric matrix, the diagonal elements are set as the maximum value of the affinity, and row items and column items are the same, and λ2 is the eigenvalue, P is the n-th order square symmetric matrix created by the above matrix creation means, P T Let P be the transpose of the n-th order square symmetric matrix, I be the unit matrix, and p x is a matrix whose diagonal elements are the elements of the row sum vector of the n-th order square symmetric matrix P, and x is a score vector, the following score vector calculation formula is used JPEG2024163543000002.jpg1074 Based on this, multiple eigenvalues λ calculated from the n-th order square symmetric matrix created by the matrix creation means 2 The score vector calculation means calculates a score vector x for each of the n language data for each of the n language data.
[0009] For each pair of multiple language data, the affinity k (which can also be called affinity, similarity, or relevance) is input as a numerical value. The affinity is input for each pair of multiple given language data, and for all combinations (regardless of order). As will be described later, it is not necessarily necessary to accept numerical input for combinations of the same language data (for language data i, j, i ≠ j). This is because the affinity between the same language data should be the highest (maximum) affinity.
[0010] The affinity value is most simply set to either 1 (affinity exists) or 0 (affinity does not exist). Of course, it is also possible to accept input of a value that indicates affinity using a value that can indicate the degree in addition to the presence or absence, such as 0, 1, 2, or 3.
[0011] The affinity k of the combination of input language data i and j is used to create an n-dimensional square matrix with the number of rows and columns equal to the number n of accepted language data. The created n-dimensional square matrix has the same column items as row items, and the affinity k of the combination of input language data i and j is set as the element of the corresponding component i, j of the n-th order square matrix, resulting in a symmetric matrix (n-th order square symmetric matrix). In addition, as described below, the diagonal elements (corresponding to i=j) can be set to the maximum value of the above affinity. This is because the diagonal elements correspond to the affinity between the same language data, and as described above, the affinity between the same language data should be the highest (maximum) affinity.
[0012] According to this invention, multiple eigenvalues λ calculated from the created n-th order square symmetric matrix are 2 is used to find the eigenvalue λ 2 For each of the n language data, a score vector x is calculated according to a score vector calculation formula.
[0013] From an n-th order square symmetric matrix, n eigenvalues λ 2 However, in the quantification affinity diagram, the eigenvalue λ 2 = 1, and the corresponding element of the score vector will have the same value, so it will be meaningless as a score vector and can be excluded. In this case, the eigenvalue λ 2 Multiple eigenvalues λ except for λ = 1 2 The score vector is calculated by substituting the score vector into the score vector calculation formula, and the n-1 eigenvalues λ 2 For each of the n-1 eigenvalues λ, a score vector x is calculated. 2 It is not necessary to use all of the above to find the score vector x.
[0014] The score vector x is a square symmetric matrix P of degree n (and its transpose P T) into the score vector calculation formula, the score vector x thus obtained contains n elements (components). These n elements constitute the score vector for each of the n language data described above.
[0015] Eigenvalue λ 2 n-1 eigenvalues λ excluding =1 2 The larger the value of , the larger the correlation coefficient of the n-th order square symmetric matrix. In other words, it becomes easier to grasp the position of each of the n language data to be processed in the whole (which language data it is correlated with, which language data it is not correlated with, and to what degree).
[0016] In one embodiment, the score vector calculation means calculates a plurality of eigenvalues λ calculated from the n-th order square symmetric matrix created by the matrix creation means. 2 Among them, the eigenvalue λ 2 = 1, and above multiple eigenvalues λ 2 The m (m≦n-1) eigenvalues λ in descending order of value 2 Using each of these, calculate the score vector x for each of the n language data. 2 The smaller the correlation coefficient λ, the smaller the degree to which the relationship between rows and columns of a two-way table is explained. 2 is also called the contribution ratio. Therefore, if you want to ensure a contribution ratio of about 70% when analyzing the relationship between rows and columns of a two-way table, you can use the scores belonging to eigenvalues whose cumulative contribution ratio (the cumulative sum of the contribution ratios corresponding to the largest eigenvalue) reaches 70%. For example, if the square symmetric matrix to be processed is of degree 31, then the 31 eigenvalues λ 2 The eigenvalue λ 2 30 eigenvalues λ excluding =1 2 It is only necessary to use the scores from the largest one until the cumulative contribution ratio reaches a desired value, and since the scores of higher-order eigenvalues are not used, calculations can be saved and processing speed can be improved.
[0017] The quantification affinity diagram implementation implements two or three eigenvalues λ in one embodiment. 2 The score vector x calculated for each piece of language data using each of the above is plotted on a two-dimensional graph having two mutually orthogonal axes or a three-dimensional graph having three mutually orthogonal axes. The position of the n pieces of language data in the whole, and further groups (clusters) of similar language data can be visually grasped by plotting.
[0018] In addition to two-dimensional and three-dimensional graphs, multiple eigenvalues λ 2 A tree diagram or a heat map may be created using the score vector x calculated using:
[0019] According to this invention, the relationships between given language data can be grasped from multiple perspectives (multiple eigenvalues), and groups of language data with affinity (similar) can be grasped using two-dimensional graphs, three-dimensional graphs, tree diagrams, heat maps, etc., making it easier to find trends in the problem to be solved and measures to solve the problem.
[0020] For two-dimensional graphs, three-dimensional graphs, etc., if the affinity k for each pair of multiple language data is the same, the results will be the same and will not differ depending on the person who performs the experiment. 2 By examining the score vector x calculated for each language, it is also possible to examine events (matters) that are latent in multiple language data (a matrix created from multiple language data).
[0021] The present invention also provides a program for causing a computer to execute the quantified affinity diagram method. [Brief description of the drawings]
[0022] [Figure 1] FIG. 1 is a block diagram of a computer system that implements a method for quantifying affinity diagramming. [Diagram 2]13 is a flowchart showing the process of a triplet data creation program. [Diagram 3] 13 is a flowchart showing the process of a triplet data creation program. [Figure 4] The following shows an example of sentences (text data) that express phenomena related to the theme of child poverty. [Diagram 5] The structure of a triplet data table file is shown below. [Figure 6] 1 shows an affinity input screen. [Figure 7] 13 is a flowchart showing the process of a two-element table conversion program. [Figure 8] The two-way table (matrix P) created by the two-way table conversion program is shown below. [Figure 9] 13 is a flowchart showing the processing of the quantification type 3 application program. [Figure 10] 13 is a flowchart showing the processing of the quantification type 3 application program. [Figure 11] 13A and 13B diagrammatically show a score vector calculated using the first eigenvalue and a score vector calculated using the second eigenvalue. [Figure 12] The eigenvalues, contribution ratios, and cumulative contribution ratios calculated from the two-way table shown in FIG. 8 are shown. [Figure 13] The scores for each language data item are calculated using the first to seventh eigenvalues. [Figure 14] A two-dimensional scatter plot is shown. [Figure 15A] A 3D scatter plot is shown. [Figure 15B] A 3D scatter plot is shown. [Figure 16A] A rotated 3D scatter plot is shown. [Figure 16B] A rotated 3D scatter plot is shown. [Figure 16C] A rotated 3D scatter plot is shown. [Figure 16D] A rotated 3D scatter plot is shown. [Figure 17] A tree diagram is shown. [Figure 18]A heat map is shown. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0023] FIG. 1 is a block diagram of a computer system that executes the quantification affinity diagramming method (hereinafter referred to as the "quantification affinity diagramming method execution device").
[0024] The quantification affinity diagramming execution device 1 includes a CPU (Central Processing Unit) 11 that performs overall control of the entire system. The CPU 11 includes a ROM (Read Only Memory) 12 consisting of a non-volatile memory that stores a BIOS (Basic Input / Output System) and the like, a RAM (Random Access Memory) 13 consisting of a volatile memory that temporarily stores programs and data, an input unit 14 including a keyboard, a mouse, and the like for inputting various instructions, a display unit 15 including a display that displays processing results, and a storage unit 16 consisting of a hard disk, an SSD (Solid State Drive), and the like, that stores data to be analyzed 16-1 and a quantification affinity diagramming execution program 16-2, which will be described later, all of which are connected via a bus.
[0025] The Quantitative Affinity Diagramming Execution Device 1 is used to organize various events (matters) that cause social problems and to consider the structure of social problems, including the extraction of potential events. As described below, the Quantitative Affinity Diagramming Execution Device 1 displays causal events as, for example, a scatter diagram. By examining this scatter diagram, multiple solutions can be derived.
[0026] The quantification affinity diagram execution program 16-2 stored in the storage unit 16 includes (1) a triplet data creation program (function), (2) a binary table conversion program (function), and (3) a quantification type 3 application program (function), which will be described later. These three programs are described in detail below. Each of these three programs can be executed in the CPU 11 by calling them with parameters specified.
[0027] (1) Triplet data creation program 2 and 3 are flow charts showing the processing of the triplet data creation program.
[0028] The triplet data creation program is a program that creates data (called "triple data (i, j, k)") that associates data k representing the degree of affinity between each pair of two language data (i, j) selected from multiple language data prepared in advance.
[0029] First, the language data to be processed is read into the RAM 13 (step S10). The language data is prepared in advance as text format data. Here, the total number of language data is n.
[0030] Figure 4 shows a specific example of language data (data expressed by language (text data)), with n=31 (31 pieces of language data) being shown as an example. The 31 pieces of language data2 shown in Figure 4 represent information obtained through interviews in a survey on child poverty. When trying to solve a social issue (in this example, "eliminating child poverty"), information from a variety of perspectives is accumulated, and by applying quantification affinity diagramming to this, the issue can be examined from multiple angles and clarified, supporting the derivation of solutions.
[0031] Each piece of language data is associated with an ID (data ID) for identifying the language data. In the example shown in Fig. 4, different numbers (integers) from 1 to 31 are used as data IDs. Each of the 31 pieces of language data is identified (specified) by the data ID.
[0032] Returning to FIG. 2, a decision is made as to whether to create new (first) triplet data or to resume creation of triplet data that is currently being created (step S11).
[0033] FIG. 5 shows an example of the structure of a triplet data table file 31 in which the triplet data created is stored.
[0034] As described in detail below, when triplet data is created, the created triplet data is saved in triplet data table file 31. Newly saved triplet data is saved in the bottom row of triplet data table file 31, and the number of rows in file 31 increases each time triplet data is created (saved) (row data is stacked downwards). Triplet data is stored in the data ID, as described above. i , and data ID j This is data that associates affinity k with a pair of two language data specified by
[0035] When new triplet data is to be created (YES in step S11 in FIG. 2), a triplet data table file 31 (FIG. 5) for saving triplet data is newly created and stored in the RAM 13. In addition, two counters i and j are stored in the RAM 13. The initial value i of counter i s The initial value of counter j is "1". s Let i be the initial value of counter i. s The number obtained by adding 1 to each of these, here "2", is set (step S12).
[0036] The process of resuming the creation of triplet data that is currently being created (NO in step S11, step S13) will be described later.
[0037] Initial value of counter i s The following process is repeated from (=1) to (n-1), which is the total number of language data n minus 1 (n=1) (hereinafter referred to as i loop) (step S14). Here, since the total number of language data n is 31, the numbers 1 to 30 are input to the counter i in sequence.
[0038] Furthermore, the following process is repeated from the initial value (=2) of counter j to the total number n of language data (hereinafter referred to as j loop) (step S15). Numbers 2 to 31 are input to counter j in sequence.
[0039] As explained below, first only counter j is incremented and when counter j reaches n (=31), counter i is then incremented, and in that state only counter j is incremented again, thereby associating affinity k for every pair of two language data selected from multiple language data, as explained below.
[0040] A pair (i, j) of two language data items selected from the plurality of language data items according to the values of the counters i and j (data IDs) is presented (displayed on the display screen) (step S16).
[0041] For example, in the case of a new process, immediately after the process starts, the initial value i of the counter i mentioned above is s is "1", and the initial value of counter j is j s is "2", so counter i=1 and counter j=2, and a pair of language data with a data ID of "1" and language data with a data ID of "2" are displayed on the display screen.
[0042] 6 shows an example of a language data display screen (affinity input screen) 40. On the language data display screen 40, two (a set) of language data 41, 42 and a sentence 43 explaining how to input the data are displayed.
[0043] The affinity for a pair of language data displayed on the language data display screen 40 is input (accepted) (steps S17 and S18). In the example shown in FIG. 6, language data with data ID "1" ("Parents do not open up") and language data with data ID "2" ("Parents do not know the type and content of support") are displayed vertically side by side. The presence or absence of affinity between these two language data displayed vertically side by side is input numerically using the input unit 14 ("Yes" in step S18, "Input affinity" in step S19). For example, if there is affinity between the two displayed language data, the number 1 is input from the input unit 14 ("Yes" in step S18, "Input affinity" in step S19), and if there is no affinity, a line feed (or the number 0) is input from the input unit 14 ("No" in step S18).
[0044] Instead of the presence or absence of affinity, the degree of affinity may be input. For example, if there is affinity, the number "3" may be input if the affinity is strong, the number "2" if the affinity is medium, and the number "1" if the affinity is weak. The affinity (number) of a set of language data to be displayed is generally input based on the judgment of an expert. Of course, the affinity of a set of language data may also be determined by discussion among multiple people who are knowledgeable about the target social issue (child poverty in this embodiment).
[0045] In addition to the presence or absence of affinity, processing interruption is also input using the language data display screen 40. For example, processing interruption is input by inputting the numerical value "-1" ("Interrupt" in step S18, "Input interruption" in step S19). This function is provided because as the amount of language data increases, the number of affinity inputs increases exponentially, making it difficult to input all at once.
[0046] If affinity is input (step S18: "Yes", step S19: "Affinity input"), a triplet of data including the data ID of counter i, the data ID of counter j, and the input affinity k (in this embodiment, k=1; as mentioned above, k=3, 2, or 1 may also be used) is added to the bottom line of the data storage table file (step S20).
[0047] It is determined whether the j loop has ended, in this case whether the counter j has reached "31" (step S21). If the j loop has not ended (NO in step S21), the counter j is incremented (step S24), and then the j loop is repeated (step S15). That is, the counter j is incremented with the counter i fixed, and the next loop starts. In this way, when the processing of the pair of data ID "1" and data ID "2" is completed, the affinity k of the pair of data ID "1" and data ID "3" is input next, and in the next loop, the affinity k of the pair of data ID "1" and data ID "4" is input, and finally the affinity k of the pair of data ID "1" and data ID "31" is input.
[0048] When the j loop is finished (when counter j reaches "31") (YES in step S21), counter i (initial value i s ) is incremented, and the incremented value becomes the new initial value i s (step S22), and the process returns to step S14 (NO in step S23). s As the initial value of the counter i mentioned above, s For example, the initial value i of counter i is set to i+1 (step S22). s When is incremented to "2", the new initial value of counter j is j s becomes "3". The affinity k of the pair of data ID "2" and data ID "3" is input, then the affinity k of the pair of data ID "2" and data ID "4" is input, and finally the affinity k of the pair of data ID "2" and data ID "31" is input.
[0049] When the i loop ends (when counter i reaches "30"), it means that the affinity k has been input for all combinations of two mutually different language data (YES in step S23). A triplet data table file 31 (see FIG. 5) containing multiple triplet data is completed and stored in storage unit 16 (step S27).
[0050] If no affinity (line break) is entered when entering affinity, the process of creating triplet data and the process of saving to triplet data table file 31 are skipped ("None" in step S18).
[0051] When the affinity k is not input and a processing interruption is input (step S19: "interrupt input"), the value of counter i (data ID), the value of counter j (data ID), and a value (-1) indicating that processing has been interrupted are added as triplet data to the bottom row of triplet data table file 31 (step S25), and triplet data table file 31 is saved in storage unit 16 (step S26). When triplet data creation in the middle of creation is resumed, triplet data table file 31 saved in storage unit 16 is read (step S10), and the two data IDs (i, j) of the triplet data stored in the bottom row of the read triplet data table file 31 are given to counters i and j, respectively, and processing is resumed. The data in the bottom row (row data in which "-1" is stored) is deleted (step S11: NO, step S13).
[0052] (2) Two-dimensional table conversion program FIG. 7 is a flowchart showing the process of the two-element table conversion program.
[0053] The two-way table conversion program uses the above-mentioned language data (Figure 4) and triplet data table file 31 (Figure 5) to perform a process of creating matrix data to be processed by the quantification type 3 application program described next.
[0054] First, the language data 2 and the triplet data table file 31 are read from the storage unit 16 and loaded into the RAM 13 (step S30).
[0055] A square matrix P is created (step S31) having the number of rows and columns according to the number n of language data 2 loaded into the RAM 13, with all components set to 0. If the number of language data 2 is 31, a square matrix with 31 rows and 31 columns, with all components set to 0, is created.
[0056] As described above, the triplet data table file 31 (FIG. 5) includes a plurality of triplet data, and each triplet data includes, for a pair of two language data, a data ID (i, j) that identifies the language data, and data (triple data) that indicates affinity k (here, k=1 for ease of understanding). The following process is performed for each triplet data (i, j, k) included in the triplet data table file 31 (NO in steps S32 and S34).
[0057] In the created matrix P, the read triplet data (i, j, k) is used to replace the component in the i-th row and j-th column of matrix P from "0" to affinity "k", and the component in the j-th row and i-th column is also replaced from "0" to affinity "k" (step S33).
[0058] For example, referring to the triplet data table file 31 in FIG. 5, the triplet data (i, j, k) that is processed first is (1, 2, 1). Since the affinity k is "1", in the 31-row, 31-column square matrix P, the component in the first row and second column is replaced from "0" to "1", and the component in the second row and first column is replaced from "0" to "1". Since the next triplet data (i, j, k) is (1, 3, 1), in the 31-row, 31-column square matrix P, the component in the first row and third column is replaced from "0" to "1", and the component in the third row and first column is replaced from "0" to "1".
[0059] In the triplet data table file 31 shown in Figure 5, affinity k for i=1, j=4 is not stored (no data exists). This is because, in the processing by the triplet data creation program described above, it was input that there is no affinity for the language data with data ID "1" and the language data with data ID "4". For this reason, the components of the first row and fourth column of the square matrix P remain "0", and the components of the fourth row and first column also remain "0".
[0060] By going through the process of step 33, the same affinity k value is stored in the matrix P for a pair of two language data sets, even when the order (row and column) of i and j is swapped. For example, the affinity k value between the language data set with data ID "1" and the language data set with data ID "2" is 1,2 (element in the first row and second column of matrix P), the affinity k between the language data with data ID "2" and the language data with data ID "1" 2,1 (affinity k 1,2 (the element in the second row and first column of matrix P) is stored.
[0061] After the above-mentioned process is performed on all triplet data contained in triplet data table file 31 (YES in step S34), the diagonal elements of matrix P are replaced from "0" to the maximum value of affinity k (here, "1") (step S35). This is because the diagonal elements of matrix P correspond to pairs of data of the same language, and affinity k should be maximum. In matrix P, the elements of the 1st row, 1st column, 2nd row, 2nd column, 3rd row, 3rd column, ..., 31st row, 31st column are replaced from "0" to "1". As a result, matrix P becomes a 31st order square symmetric matrix in which the maximum value of affinity k is stored in all diagonal elements (diagonal components).
[0062] Finally, the language data is inserted (stored) into the row names (row titles) of each row and the column names (column titles) of each column of matrix P to complete the two-dimensional table (matrix P), and the completed matrix P is stored in storage unit 16 (step S36). According to the data IDs (numbers 1 to 31) (see FIG. 4) to which each piece of language data is associated, for example, the language data of data ID "1" ("My parents don't open up") is inserted into the row name (row title) of the first row, and the language data of data ID "1" ("My parents don't open up") is also inserted into the column name (column title) of the first column.
[0063] Figure 8 shows the completed two-way table (matrix P). For simplicity, in Figure 8, the row and column names are written as TEXT1, TEXT2, ..., TEXT31, rather than as language data. Also, the component "0" is represented by "·". For example, please understand that "TEXT1" is the language data with data ID "1" ("My parents don't open up"), and TEXT2 is the language data with data ID "2" ("My parents don't know the type or content of support"). Also, please understand that the row items and column items in the two-way table are the same.
[0064] (3) Quantification Type 3 Application Program The relationship between each pair of the 31 language data is understood by the above-mentioned input affinity, but the degree of correlation between each of the 31 language data is not certain. A quantification method type 3 application program is used to analyze the correlation of the given language data from various perspectives (dimensions).
[0065] Below, the processing of the quantification type 3 application program will be explained in detail.
[0066] 9 and 10 are flow charts showing the processing of the quantification method type 3 application program.
[0067] First, the matrix P (FIG. 8) representing a two-dimensional table including language data is read from the storage unit 16 and stored in the RAM 13 (step S40).
[0068] Based on the matrix P stored in the RAM 13, eigenvalues are calculated and the calculated eigenvalues are sorted in descending order (step S41). Furthermore, the calculated eigenvalues are used to calculate the contribution rate and cumulative contribution rate described below (step S42).
[0069] FIG. 11 is a schematic diagram for explaining the relationship between multiple eigenvalues calculated by the quantification method type 3 application program.
[0070] The eigenvalues of the matrix P can be found by solving the following characteristic equation for the unknown eigenvalue α, where E is the unit matrix.
[0071] det(P-αE)=0
[0072] det() represents the determinant.
[0073] Since matrix P is a 31-order square matrix, generally 31 eigenvalues are calculated from matrix P. The eigenvectors corresponding to each eigenvalue can be calculated by substituting each eigenvalue into the following simultaneous equations.
[0074] (P-αE)x=0
[0075] The set of eigenvalues and eigenvectors of a matrix can be calculated, for example, using the open source freeware statistical analysis software "R" (https: / / cran.r-project.org / ).
[0076] 8, 31 eigenvalues including the maximum eigenvalue of 1 are obtained. However, of the 31 eigenvalues, eigenvalues that are "1" are excluded in the following processing. This is because if an eigenvalue that is "1" is used, the score vectors (score vectors will be described later) calculated for each piece of language data will have the same value, and will have no meaning as score vectors representing the correlation between language data.
[0077] The larger the calculated eigenvalues are, the stronger the correlation between the row and column items of the two-way table is reflected. Conversely, the smaller the eigenvalue is, the weaker the correlation between the row and column items of the two-way table is, and the mutual relationships between multiple language data are difficult to understand. Figure 11 is a schematic plot of score vectors (Dim1 and Dim2) calculated for each piece of language data using two eigenvalues (assuming that the first eigenvalue is greater than the second eigenvalue). In Figure 11, the horizontal and vertical axes correspond to the score vectors belonging to the first and second eigenvalues, respectively. The plot numbers correspond to the numbers of the language data (text data). In this way, by plotting each piece of language data on a two-dimensional plane using the score vectors for the eigenvalues with large values, it becomes easier to understand the mutual relationships between them. It is possible to expand this to three dimensions and examine the relationships between language data. However, visualization is not possible in four or more dimensions, so some ingenuity is required, as will be described later.
[0078] FIG. 12 shows 30 eigenvalues (No. 1 to No. 30) (eigenvalues excluding "1") actually calculated from matrix P (FIG. 8), along with their contribution rates and cumulative contribution rates. The "contribution rate" is the value obtained by dividing each of the 30 eigenvalues by the sum of all eigenvalues. The "cumulative contribution rate" is the cumulative addition of the contribution rates calculated corresponding to each of the multiple eigenvalues, starting from the largest eigenvalue. Note that the eigenvalues, contribution rates, and cumulative contribution rates are actually calculated up to the seventh decimal place and beyond as shown in FIG. 12, and the values in FIG. 12 to the eighth decimal place are rounded off.
[0079] As mentioned above, the larger the eigenvalue, the stronger the correlation between the language data is expressed by the score vector, so it is not necessary to use all score vectors to examine the relationship between the language data. For this reason, it is realistic to use score vectors that can secure at least 70 to 80% of the cumulative contribution rate depending on the problem, to examine the relationship between the language data and try to grasp the overall picture (step S43).
[0080] Figure 13 shows score vectors calculated for each of the 31 language data (Figure 4) using the seven largest eigenvalues (No. 1 to No. 7 in Figure 12) among the eigenvalues (Figure 12) calculated from matrix P (Figure 8). In Figure 13, Dim1, Dim2, ... Dim7 represent the number of variables (axes) latent in matrix P, and correspond to the seven largest eigenvalues (their eigenvectors) among the 31 eigenvalues calculated from matrix P, excluding the eigenvalue that is "1".
[0081] Let eigenvalue be λ 2 Then, the score vector x for each of the 31 language data can be found by solving the following formula (called the score vector calculation formula).
[0082] JPEG2024163543000003.jpg1074
[0083] In the above score vector calculation formula, P is the square matrix to be processed, P T represents the transpose of a square matrix P, and I represents the identity matrix. x is a matrix whose diagonal elements are the elements of the row sum vector of the square matrix P. x is the score vector. Note that since the square matrix P is a symmetric matrix, P = P T The score vector x can be calculated for each language data and for each of the n-1 eigenvalues.
[0084] The score vector calculation formula mentioned above is based on the eigenvalue problem obtained from general quantification type 3, which deals with two mutually different items (row items and column items that are different from each other), and is expressed when the row items and column items are the same (not distinguished) as a condition, and is a formula for finding a score vector that maximizes the correlation coefficient of a two-way table.
[0085] The score vector x calculated for each of the 31 language data for each eigenvalue can also be calculated using the statistical analysis software "R" mentioned above. In this software, quantification type 3 is called correspondence analysis.
[0086] In the following explanation, the calculated eigenvalues (λ 2 = 1) are called the first eigenvalue, second eigenvalue, third eigenvalue, ..., in order from largest to smallest.
[0087] FIG. 14 is a graph in which the score vector of each language data calculated based on the score vector calculation formula using the first eigenvalue (hereinafter referred to as "Dim1 score") is on the horizontal axis, and the score vector of each language data calculated based on the score vector calculation formula using the second eigenvalue (hereinafter referred to as "Dim2 score"; the same applies to Dim3 score described below) is on the vertical axis, and shows a scatter diagram in which each language data is plotted according to the score vector calculated using the first eigenvalue and the second eigenvalue for each of the 31 language data. By using two axes corresponding to the first eigenvalue and the second eigenvalue, it is possible to grasp the correlation between each language data in the two-dimensional (plane) graph shown in FIG. 14. The quantification type 3 application program can display a two-dimensional scatter diagram as shown in FIG. 14 on the display unit 15 (step S44).
[0088] Referring to Figure 14, if we focus on the Dim1 score (horizontal axis), for example, out of the 31 language data, a high score is calculated for "25. The scholarship system needs to be improved," and a low score is calculated for "5. There is no place to work." As an example, it can be seen at a glance that the distance between these two language data on the Dim1 score axis is large.
[0089] On the other hand, if we look at the Dim2 score (vertical axis), for example, out of the 31 language data, a large score is calculated for "25. The scholarship system needs to be improved," while a small score is calculated for "29. It's good to be able to support food services." As an example, it is possible to see at a glance that there is a large distance between these two language data on the Dim2 score axis.
[0090] Looking at the scatter plot in Figure 14, there is an area where multiple pieces of language data (plots) with close Dim1 and Dim2 scores are concentrated. It is possible that this concentrated language data is important information that will lead to the resolution of the social issue of eliminating child poverty. If a different Dim score is used, the plots for the child poverty issue may be separated on the scatter plot into, for example, cause events and result events. If such a scatter plot is obtained, it may be possible to pick out only the cause events and repeat the above process.
[0091] Needless to say, the scatter plot shown in FIG. 14 can also be displayed with Dim1 scores on the horizontal axis and Dim3 scores on the vertical axis.
[0092] Figures 15A and 15B show scatter diagrams in which score vectors calculated based on the first, second, and third eigenvalues for each of the 31 language data are plotted in a three-dimensional space having mutually orthogonal axes (Dim1 score, Dim2 score, Dim3 score). As shown in Figures 15A and 15B, the correlation between each language data can also be grasped in a three-dimensional space. The quantification type 3 application program can also display three-dimensional scatter diagrams such as those shown in Figures 15A and 15B on the display unit 15 (step S44).
[0093] Figures 16A, 16B, 16C, and 16D show the three-dimensional scatter diagrams shown in Figures 15A and 15B rotated to view them from different viewpoints. The rotated displays shown in Figures 16A to 16D are also executed by the quantification type 3 application program, and the correlations of the 31 language data can be displayed on the display unit 15 in an easily understandable manner (step S44).
[0094] In FIG. 17, the Euclidean distance between each of the 31 language data in a seven-dimensional space is calculated using the scores from Dim1 to Dim7 for each of the language data, and the results are displayed as a tree diagram, which is a result of organizing the data into seven clusters using the Ward method of cluster analysis. More specifically, a combination of language data having a plot with the closest Euclidean distance is found in a two-dimensional plane including many plots shown in FIG. 14, and the combination is paired. The center of gravity of the paired combination is made into a new plot, and the above process is repeated including the new plot, so that a tree diagram as shown in FIG. 17 is finally created. In FIG. 17, the length of the straight line extending in the left and right directions is represented by the length corresponding to the Euclidean distance. Depending on the degree of correlation, the language data can be expressed as being grouped together (clustering, grouping). The quantification type 3 application program can also display a tree diagram as shown in FIG. 17 on the display unit 15 (step S45).
[0095] Fig. 18 shows a heat map in which the language data are arranged in descending order of the Dim1 score calculated for each of the 31 language data, the Dim1 score, Dim2 score, and Dim3 score are displayed corresponding to each of the language data, and the Dim1 score, Dim2 score, and Dim3 score are displayed in a color according to the magnitude of the score value. The quantification method type 3 application program can also display a heat map such as that shown in Fig. 18 on the display unit 15 (step S46). It is useful for examining the similarity between variables latent in matrix P such as Dim1, Dim2, and Dim3 (events (matters) latent in matrix P).
[0096] Finally, we will explain the eigenvalue problem obtained from general quantification type 3, which deals with two types of mutually different items (row items and column items that are mutually different), which is the basis for derivation of the score vector calculation formula mentioned above.
[0097] Quantification Method of Type 3 (QM3) is a method for analyzing the internal structure of data given in the form of a two-way table, such as Table 1 shown below. Here, X and Y are distinct items. X i is the i-th X item, Y j represents the j-th Y item. Each item is associated with a score (quantity), which is expressed as x i ,y j It is said that. ij is item X i and item Y j represents the relevance of i and p j respectively represent the sum of the i-th row and the j-th column. ij is item X i and Y j If there is a correlation between the two, it is marked as "1", and if not, it is marked as "0". The degree of correlation can be represented as 3, 2, or 1, corresponding to strong, medium, or weak. For simplicity, we will use 1 and 0 here. In Table 1, the element value is divided by the sum of the elements in the two-way table to convert it into a relative frequency form so that the sum is 1.
[0098] Table 1 Two-way table TIFF2024163543000004.tif5090
[0099] The above two-way table shows two types of items X that are not related to quantity. i and Y j The quantification type 3 is composed of item X i and Y j to quantity x i and y j It analyzes the internal structure by making the above correspond (quantification).
[0100] Without loss of generality, x i andy j The average of each of these can be set to 0 as follows:
[0101] JPEG2024163543000005.jpg8101 JPEG2024163543000006.jpg9100
[0102] Also, x i andy j The variance of each of can also be set to 1.
[0103] JPEG2024163543000007.jpg9101JPEG2024163543000008.jpg9101
[0104] The correlation coefficient ρ is expressed as follows: JPEG2024163543000009.jpg9105
[0105] In quantification type 3, we use equations (1.3) and (1.4) as constraints to determine the quantity x so that the correlation coefficient ρ is maximized. i and y j Therefore, we use the Lagrange method of undetermined multipliers λ1 and λ2. We need to maximize Q in the following equation.
[0106] JPEG2024163543000010.jpg13138
[0107] quantity x i and y j By partially differentiating with respect to and setting it to 0, we obtain the following equation.
[0108] JPEG2024163543000011.jpg11113JPEG2024163543000012.jpg11114
[0109] In equation (1.7), x i Multiplying by and summing over the index i gives us the following equation:
[0110] JPEG2024163543000013.jpg9116
[0111] Similarly, in equation (1.8), j Multiplying by and summing over the index j gives us the following equation:
[0112] JPEG2024163543000014.jpg9116
[0113] From equations (1.3) and (1.4), and from equations (1.9) and (1.10), λ1 = λ2, so we change λ1 and λ2 to λ. Then, from equation (1.5), λ = ρ.
[0114] Equation (1.7) can be transformed as follows, and from this transformation, the quantity (score) x i can be sought.
[0115] JPEG2024163543000015.jpg13100
[0116] Substituting equation (1.11) into equation (1.8) gives the following equation.
[0117] JPEG2024163543000016.jpg9116
[0118] Here, we introduce the following notation:
[0119] JPEG2024163543000017.jpg6108 JPEG2024163543000018.jpg6108 JPEG2024163543000019.jpg20112JPEG2024163543000020.jpg24139JPEG2024163543000021.jpg24139
[0120] Using these notations, equation (1.12) can be expressed as follows:
[0121] JPEG2024163543000022.jpg6110
[0122] Equation (1.18) can be rearranged as follows, which means that λ 2 This represents an eigenvalue problem with eigenvalues.
[0123] JPEG2024163543000023.jpg7123 therefore,
[0124] JPEG2024163543000024.jpg698 Then, we have the following formula:
[0125] JPEG2024163543000025.jpg7118
[0126] The score vector calculation formula described above can be derived by eliminating the distinction between x (row item) and y (column item) in the above formula (1.21) (that is, making item X the same as item Y). [Explanation of symbols]
[0127] 1 Quantitative Affinity Diagram Implementation Device 11 CPU 12 ROM 13 RAM 14 Input section 15 Display section 16 Memory section 16-1 Language Data 16-2 Quantitative Affinity Diagram Execution Program
Claims
1. a language data input means for receiving input of n (n is an integer equal to or greater than 3) pieces of language data; a display means for displaying all combinations of two mutually different language data i, j (i ≠ j) using the received n language data; an affinity input means for receiving an input of a numerical value representing an affinity k for each of the displayed combinations of two language data i and j; a matrix creation means for creating data representing an n-th order square symmetric matrix with n rows and n columns, in which the affinity k of the combination of language data i and j input by said affinity input means is set as the element of the corresponding element i, j of said n-th order square symmetric matrix, the diagonal elements are set as the maximum value of said affinity, and row items and column items are the same; λ 2 is the eigenvalue, P is the n-th order square symmetric matrix created by the matrix creation means, P T Let P be the transpose of the n-th order square symmetric matrix, I be the unit matrix, and p x When the elements of the row sum vector of the n-th order square symmetric matrix P are the diagonal elements, and x is the score vector, the following score vector calculation formula is used. Based on this, a plurality of eigenvalues λ calculated from the n-th order square symmetric matrix created by the matrix creation means 2 a score vector calculation means for calculating a score vector x for each of the n language data, Quantitative affinity diagram implementation device.
2. The score vector calculation means A plurality of eigenvalues λ calculated from the n-th order square symmetric matrix created by the matrix creation means 2 Among them, the eigenvalue λ 2 = Multiple eigenvalues λ excluding 1 2 For each of the n language data, a score vector x is calculated.
2. An apparatus for performing quantified affinity mapping according to claim 1.
3. The score vector calculation means A plurality of eigenvalues λ calculated from the n-th order square symmetric matrix created by the matrix creation means 2 Among them, the eigenvalue λ 2 = 1 and multiple eigenvalues λ 2 The m (m≦n-1) eigenvalues λ in descending order of value 2 Using each of the n language data, a score vector x is calculated.
2. An apparatus for performing quantified affinity mapping according to claim 1.
4. Two or three eigenvalues λ 2 The method further includes a graph display data creation means for plotting the score vector x calculated for each of the language data using each of the above on a two-dimensional graph having two mutually orthogonal axes or a three-dimensional graph having three mutually orthogonal axes, 4. An apparatus for performing quantified affinity mapping according to claim 3.
5. a rotation display means for rotating and displaying a three-dimensional graph including the plot on a display screen, 5. An apparatus for performing quantified affinity mapping according to claim 4.
6. The graph display data creation means The language data is displayed in correspondence with each of the plots.
5. An apparatus for performing quantified affinity mapping according to claim 4.
7. an affinity input program that causes a computer to execute a process of displaying all combinations of two mutually different language data i, j (i ≠ j) on a display screen using the received n (n is an integer equal to or greater than 3) pieces of language data, and receiving, from an input device, an input of a numerical value representing an affinity k for each of the displayed combinations of two language data i, j; a matrix creation program that causes a computer to execute a process of creating data representing an n-th order square symmetric matrix with n rows and n columns, in which the affinity k of the combination of accepted language data i and j is set as the element of the corresponding element i, j of the n-th order square symmetric matrix, the diagonal elements are set as the maximum value of the affinity, and the row items and column items are the same; and λ 2 is the eigenvalue, P is the n-th order square symmetric matrix created by the matrix creation means, P T Let P be the transpose of the n-th order square symmetric matrix, I be the unit matrix, and p x When the elements of the row sum vector of the n-th order square symmetric matrix P are the diagonal elements, and x is the score vector, the following score vector calculation formula is used. Based on this, a plurality of eigenvalues λ calculated from the n-th order square symmetric matrix created by the matrix creation means 2 A score vector calculation program for causing a computer to execute a process of calculating a score vector x for each of the n language data, A program that performs quantification affinity mapping.
