JP2016133899A - Information processor, information processing method, and program - Google Patents

Abstract

It is possible to generate a chart (graph) capable of easily performing comparative recognition in consideration of relevance and similarity between constituent axes.
A feature distance calculation unit that calculates a similarity between a plurality of feature values extracted from input multivariate data, and a dimension reduction that reduces the number of dimensions based on the calculated similarity between feature values. A chart shape determining unit that determines the arrangement of the axes corresponding to the feature amount based on the result of dimension reduction, and an output unit that outputs a chart in which multivariate data is drawn according to the determined axis arrangement, Even multivariate data that is difficult to grasp with a glance can be displayed in an easy-to-understand expression form.
[Selection] Figure 1

Description

  The present invention relates to an information processing apparatus, an information processing method, and a program.

  When people perform data analysis, they do not look at the enumeration of numbers as they are, but use various graphs or charts as visual representations to display data in a form that is easy for humans to understand. Is common. By making it easy to grasp the data with a visualization method appropriately adopted according to the target data and the nature of the information to be grasped, it becomes possible to read data trends and information that are not noticed by simply enumerating numbers. Providing new problem-solving methods and added value by utilizing big data is becoming the mainstream of recent technological fields. For example, it is important to effectively visualize multivariate data to support human eye analysis.

  Generally, when analyzing data with a plurality of items, a pie chart, a bar chart (Bar Chart, Bar Graph), or the like is used. In addition, a radar chart, a parallel coordinate plot, or the like is used. Of these, bar graphs, radar charts, and parallel coordinate plots are methods that allow the relationship between multivariate data to be known as they are. These three graphs are data visualization methods that can be used for various evaluations and are generally used to compare multivariates. In addition, a plurality of items in the graph can be compared with each other, and other graphs having evaluation values for the same item can be compared.

  Bar graphs, parallel coordinate plots, and radar charts are data visualization methods used for multivariate analysis, but as the number of variables increases, the number of axes increases and the shape of the graph becomes more complex. It is known that it is difficult to grasp. A well-known fact in cognitive neuroscience is the “Magic Number 7 ± 2 theory” (Non-Patent Document 1). This is a theory that when a person recognizes information, it can be recognized well if it is information contained in 7 ± 2 chunks, but it becomes difficult if it exceeds that. According to this theory, for example, even when a radar chart is excellent in intuitive comparison of graph shapes, when comparing a plurality of graph shapes, there is a general case where there is no semantic chunk between adjacent axes. In the case of a radar chart, each axis is one chunk. Therefore, the radar chart having the number of axes exceeding 7 ± 2 is highly likely to exceed the human shape recognition ability.

  In the following Patent Document 1, partial parallel coordinate plots are extracted and arranged in parallel coordinate plots that extend in the horizontal direction as the number of feature quantities representing the data is increased when multivariate data is visualized. A technique for creating a graph that is easy to recognize is proposed. Non-Patent Document 2 below describes that the order of variables assigned to the axes of the radar chart is important, and sets those close by PCA to be close to the axes on the radar chart. It has been explained.

JP2013-161226A

“The Magical Number Seven, Plus or Minus Two: Some Limits on Our Capacity for Processing Information”, George A. Miller, The Psychological Review, 1956 "Multivariate Analysis Case Collection Vol. 1" (Yoshizawa and Haga, Hikaru Girenren, 1992)

  Here, the problem that can be said in all radar charts, parallel coordinate plots, bar graphs and bar graphs that are displayed as lines is that the data visualization is easy to grasp as a shape, while the relative position and spacing of axes (width and angle) ) Has no information value. Although these graphs are formed by lines connecting multivariate items (in the case of a bar graph, by the arrangement of bars), the shapes are determined by connecting adjacent items by chance. Therefore, in the information obtained by visualization, for human recognition, the relationship between adjacent axes is recognized with priority. For example, in a bar graph, the height difference between distant bars is more difficult to grasp intuitively than the difference between nearby bars, and the parallel coordinate plot only tells whether there is a correlation between adjacent variables. . Therefore, it is necessary to devise methods for effectively visualizing multivariate data, such as separating items that do not need to compare score differences in detail, or making items that should be compared in detail close to each other. It becomes.

  In the method of partially extracting and arranging the parallel coordinate plots described in Patent Document 1, since a portion that should be noted by a person is narrowed down, it becomes easy to compare partial parallel coordinate plots. However, if you want to grasp the data from a bird's-eye view of the entire graph, it is not suitable for intuitively grasping information at a glance because multiple partial combinations of multivariate are visualized into one graph. May be easy to overlook, and oversight may occur. In addition, according to the method for determining the order of the variables in the radar chart of Non-Patent Document 2, only the variable components to be grouped together can be obtained, and a chart reflecting the similarity between the axes representing these variables can be obtained. I can't.

  The present invention has been made in view of such circumstances, and can generate a chart (graph) that can easily perform comparative recognition in consideration of the relevance and similarity between constituent axes. The purpose is to.

  The information processing apparatus according to the present invention includes a calculation unit that calculates a similarity between a plurality of feature amounts extracted from input multivariate data, a similarity between the feature amounts calculated by the calculation unit, and the multivariate Dimension reduction means for reducing the number of dimensions based on data, shape determination means for determining an arrangement of axes corresponding to each of the plurality of feature amounts based on a result of dimension reduction by the dimension reduction means, and the shape Output means for outputting a drawing in which the multivariate data is drawn according to the arrangement of the axes determined by the determining means.

  According to the present invention, even multivariate data that is difficult to grasp with a glance can be displayed in an easy-to-understand expression form, and an effect of facilitating data analysis can be obtained.

It is a figure which shows the structural example of the multivariate data visualization apparatus as an information processing apparatus in this embodiment. It is a flowchart which shows the example of the processing operation in this embodiment. It is a figure which shows the example from which the shape of a radar chart differs for every abnormal pattern. It is a figure which shows the example of the radar chart unsuitable for discrimination | determination of the data of a different property. FIG. 5 is a diagram showing an example in which the radar chart shown in FIG. 4 is converted so that data of different properties can be easily identified. It is a figure which shows the example of the production | generation procedure of the radar chart which concerns on 1st Embodiment. It is a figure which shows the example of the graph by 1st Embodiment. It is a figure which shows the other example of the production | generation procedure of the radar chart which concerns on 1st Embodiment. It is a figure which shows the other example of the production | generation procedure of the radar chart which concerns on 1st Embodiment. It is a figure which shows the example of GUI which can select the display format which concerns on 1st Embodiment. It is a figure which shows the computer function which can implement | achieve the information processing apparatus in a present Example.

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

(First embodiment)
A first embodiment of the present invention will be described. In the first embodiment, when there is a plurality of data expressed by a plurality of feature amounts and it is desired to analyze what kind of tendency each data has among them, it is intuitively understood by humans. This is a method of generating a radar chart or a similar chart in an easy form.

  As an example, a use case in an appearance inspection in which a plurality of feature amounts are extracted from an input image and automatically inspected will be described. In an appearance inspection that automatically determines whether an input image is normal or abnormal, a plurality of feature values necessary for determination are extracted from the input image, and the score (evaluation value) is comprehensively determined to determine whether the image is normal or defective. judge. There is a correlation between each score indicated by the extracted feature quantity and the type of abnormality detected. In addition to the normal / abnormal labels that are the judgment results at the production site, the measurement is based on the judgment score and the score on the original extracted feature axis. The tendency of difficult defects can be analyzed. Thereby, for example, feedback can be applied to the production line design.

  In particular, visual inspection equipment that automated sensory visual inspection with the human eye has complex patterns and various levels of abnormal data, so what is the reason for determining abnormalities based only on images that are determined to be abnormal? Difficult to analyze. Therefore, by visualizing the extracted features with a chart (graph) having a high listability, it is possible to easily grasp a tendency that is difficult to understand simply by looking at the image.

  As an example, FIG. 3 shows a state in which the difference in the shape of the radar chart differs for each abnormal pattern as an example in which the difference in the tendency of the score of the feature amount extracted from each input image can be intuitively known. A case where the number of extracted features is five is shown, and an ID is assigned to each extracted feature by a number. 3A shows an example of a normal image as an input image and a corresponding radar chart, and FIG. 3B shows an example of a mura defect image as an input image and a corresponding radar chart. ing. FIG. 3C shows an example of a flaw defect image as an input image and a corresponding radar chart. FIG. 3D shows a foreign object defect image as an input image and a corresponding radar chart. An example is shown. As shown in FIGS. 3A to 3D, the shapes of the respective radar charts are different.

  As described above, when an appropriate radar chart is set, it is useful for the user to intuitively grasp the defect tendency simply by looking at the shape of the radar chart. Here, for the sake of clarity, each defect case is shown as an input image that can be understood only by looking at the image. However, the defect sample that occurs at the production site has almost the same appearance as a normal sample. There are also things. In such a case, it is possible to intuitively understand only by looking at the shape of the radar chart when confirming or analyzing what kind of abnormality the sample has. In addition, since the visualization of such information enables the user to intuitively check how the algorithm determines the image without understanding the complicated determination processing algorithm itself, the reliability of the inspection device It helps to improve the degree and detect defects early.

  However, the number of extracted features as shown in FIG. 3 is as few as five, and it is sufficient if each of them is sufficient to grasp the tendency of each defect, but generally it can cope with various abnormal patterns. In many cases, the number of extracted features increases from several tens to several hundreds. In such a case, if a radar chart is created without any ingenuity, a chart different from a chart that can be grasped at a glance as shown in FIG. 3 is often generated.

  As an example, when the number of extracted features is 35, the difference in graph shape due to the tendency of defects by drawing a radar chart in which the extracted feature quantities are simply arranged clockwise in the order of the extracted features FIG. 4 shows a typical example in which it is difficult to grasp the above. The radar charts shown in FIGS. 4A to 4D are graphs generated when an image of an inspection target that is regarded as a defect for different reasons is input. As a large number of inspection objects flow along the line at the production site in a day, a person grasps the difference between these four shapes at a glance, and immediately understands what kind of defect each means. It is difficult.

  Therefore, in the present embodiment, the graph is converted into a graph that makes it easy to grasp the difference in shape intuitively by performing processing described below on the graph. 5A to 5D, the order of the axes is arranged so that each of the axes shown in FIGS. 4A to 4D can be easily recognized while keeping the score indicated by each axis as it is. The changed radar chart is shown. According to this, in FIG. 4 (A) to FIG. 4 (D), the unclear property can be clearly understood. For example, FIG. 5A, FIG. 5B, and FIG. 5C are charts corresponding to images showing different types of defects, and FIG. 5D is a chart corresponding to FIG. It is easy to immediately understand that there is a high possibility that the two defect types shown in FIG.

  That is, even if the number of extracted features increases, as shown in FIG. 5 (A) to FIG. 5 (D), as shown in FIGS. It becomes possible to create a radar chart that can grasp the tendency of defects just by looking. Hereinafter, an example of an easy-to-see radar chart generation algorithm will be described in detail.

  As described above, in order to generate an easy-to-read radar chart, it is preferable to arrange highly relevant axes close to each other. If there is a feature quantity that has a predetermined reaction to an input defect signal having a predetermined tendency, in the radar chart in which those axes are arranged close to each other, the corresponding portion is always deformed in conjunction with the shape of the graph. For this reason, by giving the same meaning to a group of areas, it is sufficient for a person to read the graph while paying attention to a few partial areas. On the other hand, it is difficult for a person to collectively recognize a set of feature values having the same meaning in a radar chart in which axes that react in conjunction are scattered.

  Here, the flow of processing will be described with reference to a block diagram showing a configuration example of a multivariate data visualization apparatus as an information processing apparatus in the present embodiment of FIG. 1 and a flowchart of FIG. In the multivariate data input unit 101, data obtained by extracting n feature values from a target sample such as an image is input (S201). Then, the inter-feature amount distance calculation unit 102 calculates the similarity (distance) between the respective feature amounts based on the inputted n feature amounts (S202), and calculates the similarity between the obtained feature amounts. N vectors representing each feature amount expressed in the feature space as a scale are acquired.

  The dimension reduction unit 103 performs dimension reduction on these n vectors in two dimensions (or three dimensions) so that points are arranged at equal distances from the origin (S203). At this time, if the user wants to specify parameters such as the number of dimension reductions required for data visualization and which feature quantities want to be moved closer (or moved away from each other), input the data visualization parameters to reflect these changes in the results. It is also possible to input from the unit 104.

  The chart shape determining unit 106 determines the arrangement of the chart axes based on the vectors representing n feature quantities arranged at equal distances from the obtained origin and the parameters input by the user using the data visualization parameter input unit 104. Determination (S204) is made (S204). In accordance with this arrangement, each of the data input from the multivariate data input unit 101 is plotted with an evaluation value for each axis to complete a chart, and is output (displayed) by the output unit 105 such as a display. .

  In addition, the chart axis storage determined by the chart shape determination unit 106 is stored in the chart shape storage unit 108. Then, the multivariate data that has not been used for the chart axis arrangement is received by the multivariate data addition input unit 107 and plotted on the chart axis held in the chart shape storage unit 108 to create a new chart. Create and output to the output unit 105.

Next, an example of calculating the similarity (distance) between specific feature amounts in step S202 shown in FIG. 2 will be described. Whether or not the feature amounts are highly related can be defined by, for example, the amount of information of the Cullback / Librer. When calculating the similarity between feature quantities based on the amount of information of the Cullback / Librer, the distribution of data indicated by each feature quantity is estimated by an algorithm such as kernel density estimation (KDE) and defined by the distance between the estimated distributions. . If the probability distributions estimated from the data variations indicated by the two feature quantities F _i and F _{j with} the feature quantity IDs i and j are P _i and P _j , the similarity between the two distributions is expressed by the following equation (1). Defined by

  However, in addition to the distance between the distributions shown in the equation (1), it is assumed that the two probability distributions are multi-dimensional normal distributions, the Bhattacharya distance is used, or the other information is defined by mutual information. Needless to say, it may be calculated using an index. Similar results can also be obtained by a method called density ratio estimation in which each probability distribution estimation and distance calculation between the distributions are performed simultaneously. In addition, as an index that can be used for calculating the distance between distributions, there is a method that uses the sum of square sums of differences in score ranking of data indicated by each feature amount. Also in this case, the larger the score, the more the feature amounts can be treated as not similar.

As described above, the similarity between all n input feature values F = {F _i } ⁿ _{i = 1} is calculated. As a result, a vector whose elements are similarities to n feature quantities including n itself is obtained for all n input feature quantities. The vector obtained at this time is expressed by equation (2).

  Next, an example of a specific calculation method for the process performed in step S203 subsequent to step S203 will be described. In step S203, the axis representing the n feature quantities as the axis of the two-dimensional radar chart performs dimension reduction to two dimensions, and the coordinates of the point for drawing the corresponding axis after dimension reduction are expressed by equation (3). Represent. Equation (3) shows a case where the dimension is reduced to m dimensions, but in an ordinary radar chart, m = 2 may be used. In Equation (3), B is an embedding matrix for generating a radar chart, and the definition is as shown in Equation (4).

Here, several methods for obtaining the embedding matrix B are shown below. The embedding matrix B to be obtained is referred to as B ^*, and the embedding matrix is defined by the equation (5) by the similarity matrix W describing the rules of the feature quantities to be brought close to each other.

Here, W _{i, j,} which is an element of the similarity matrix W, is a matrix describing a rule for bringing the feature quantities close to each other. When the i-th feature and the j-th feature are brought close to each other, 1 is used. Any function can be used as long as the function is set to 0. In other words, when the similarity W _{i, j} is 1, the distance after dimension reduction is to be minimized by Equation (5), and when it is 0, B ^* is determined so as to be ignored. Of course, when the priority to be approached is to be changed depending on the target, a value between 0 and 1 may be set as Wi _{, j} . Examples of expressions that can be used as examples of the similarity W _{i, j} are shown in Expressions (6) and (7).

Equation (6) is a method for determining whether G _i in the n-dimensional feature space is near k of G _j or not, and Equation (7) is a distance base defined by a constant γ. This is a method for setting the value calculated in (1). As another example of _{Wi, j} , a value reflecting prior knowledge about the extracted feature amount may be set. If there is a feature quantity that is not desired to be approached when the data is visualized due to the convenience of the user due to the nature of the extracted feature, the similarity W _{i, j} between the feature quantities after the user designates the corresponding feature quantity ID. By setting 0 to 0, the result desired by the user can be obtained. Conversely, if there is a feature that the user wants to approach, Wi _{, j} between the corresponding feature amounts may be set to 1 by the same input.

  Dimension reduction is performed based on these indices, but the point representing each feature ID is a circle with a distance of 1 from the origin due to the constraint that the distance from the origin after dimension reduction in Equation (5) is 1. Get the top results. FIG. 6 shows an image diagram before and after processing at this time, and an example of a chart obtained at that time.

In FIG. 6A, the vector G _i representing each of the ten feature quantities is distributed in a 10-dimensional space defined by the similarity-based distance of all the input feature quantities described above by the Cullback / railer information quantity. It shows how it is. In FIG. 6B, a vector g _i representing each feature quantity two-dimensionally reduced by the dimension reduction embedding matrix calculated by the method described above is distributed on a circle having a distance of 1 around the origin. It shows that. This is an index of the positional relationship in which each axis is drawn on the radar chart.

Since the point g _i on the circle is a position index for setting the i-th axis of the feature amount ID representing each feature amount, the axes are set and drawn from the origin toward all ten points. A chart as shown in FIG. 6C can be obtained. This chart does not arrange the axes at regular intervals like the radar chart, but the similarity of each axis is expressed as the angle between each axis, so it is easy to grasp the semantic unity, It may be easier to understand than a general radar chart. In this way, a new chart that compensates for the shortcomings of radar charts that have no information value in relative positions and angles of axes is referred to as “Clumsy Spider Chart (CSC)”.

  FIG. 6D shows an example of CSC drawn by plotting each axis score of arbitrary data on this chart. Although FIG. 6D is expressed as a CSC when the number of axes is small in order to simplify the explanation, the CSC is useful when dealing with higher-dimensional multivariate. Since the axes whose scores change in conjunction with each other are arranged close to each other even if it becomes higher, if you pay attention to the local area, you only need to observe the score fluctuations of the nearby axes together, and if you look at the overall graph shape You can get an overview of the balance of the movement of scores with local unity.

  As described above, by taking into account the relevance and similarity between the axes, the arrangement (arrangement and interval) of the axes is set and a chart (graph) is generated, so that when the whole is viewed, a plurality of axes are It is possible to intuitively grasp the semantic unity. This makes it easy to compare the shape of the entire graph when viewed from a macroscopic viewpoint, and from a microscopic viewpoint, it is possible to compare by focusing on the differences between closely related objects that have close axes. A chart (graph) that is easy on recognition can be generated.

  FIGS. 7A and 7B show examples of CSCs in which comparison of high-dimensional (108-dimensional) multivariate data is intuitively superior. According to FIG. 7A and FIG. 7B, it is intuitively understood which axis moves with a strong correlation in comparison with a general radar chart. Thus, it can be seen that the shape can be easily recognized even when the dimensions become high. When comparing the charts shown in FIG. 7A and FIG. 7B, the overall shape trend does not change, but there is a difference in the shape of the upper right part and the upper left part as a local shape. Is easy to understand. Needless to say, a general radar chart in which the axes are not arranged in an appropriate order makes it difficult to grasp the trend by looking at the graphs on the 108 axes.

  In addition, for users who prefer a display with the same angle between the axes as in a normal radar chart, the axis-to-axis angles are equally spaced while keeping the axes aligned so that only the appearance matches the general radar chart. You may do it. The concept of the conversion process in that case is shown in FIGS. 8A to 8D, and a graph corresponding to FIG. 6D is shown in FIG.

  In order to grasp the tendency of data intuitively in a shorter time rather than in a detailed analysis, it may be better to remove a portion that is likely to be redundant information as much as possible and present it to the user. Therefore, a simple chart display mode or the like can be selected, and axes whose similarity between feature amounts is higher than a predetermined value may be displayed together on one axis. The processing procedure at that time is shown in FIGS. For example, if the angle between the axes after dimension reduction is less than or equal to the set threshold, those axes are collectively expressed as one axis, and the score represented by that axis is the average of the scores of the original multiple axes There is a method of calculating by making a value. Alternatively, before the dimension reduction, the points in the neighborhood relation may be averaged to be combined into one point, and then the dimension reduction may be performed. FIG. 9E is an example in which the data shown in FIG. 6D and FIG. 8E is displayed on the chart of FIG. 9D, but FIG. 9E is more general. For a user who wants to view in the form of a simple radar chart, the angles between all axes may be adjusted so as to be equally spaced.

  The display switching of the various charts described above can be easily selected by simply displaying a chart corresponding to a plurality of images as shown in FIG. 10, for example, and the user clicking a button while viewing the chart. You may do it. This facilitates analysis suitable for the data.

  In the above description, since all display forms are assumed to be similar to a general radar chart, the number of dimensions at the time of dimension reduction is described as an example of two dimensions. However, when more complicated analysis is required, etc. The number of dimensions at the time of dimension reduction may be three dimensions. When the number of dimensions after the reduction is 3, it becomes a three-dimensional representation of the chart described above, and a graph that more intuitively represents the relationship between each feature amount can be created. However, when the display for confirming the three-dimensional chart is a two-dimensional display, it may be more difficult to grasp the shape than the two-dimensional chart. Therefore, in the case of a three-dimensional chart, it may be displayed in color while changing RGB (display color) values according to latitude and longitude. Although the present embodiment has been described by taking the appearance inspection as an example, it is needless to say that the present embodiment is a method that can be used for general data visualization when a plurality of feature quantities are handled.

(Second Embodiment)
Next, a second embodiment of the present invention will be described. In the second embodiment, dimension reduction is performed based on the degree of retention of the distance relationship between feature quantities before and after the dimension reduction for the dimension generation for chart generation described in the first embodiment. This is a method for generating a radar chart or the like in an intuitively easy-to-understand form. Hereinafter, differences of the second embodiment from the first embodiment will be described.

  When dimension reduction is performed on the basis of the degree of retention of the distance relationship between feature quantities before and after the dimension reduction, what should be minimized is an error caused by the distance relation changing before and after the dimension reduction. The sum of squared errors at this time may be defined, for example, by the method shown in Equation (8).

However, the feature amounts G and g before dimension reduction are the same as those shown in the first embodiment, Δ _ij is defined as in Expression (9), and the distance between G _i and G _j is defined as Represent.

Similarly, δ _ij represents the distance between g _i and g _j after dimension reduction. However, the restriction shown in Expression (10) is added to δ _ij so that the distance from the origin becomes equal after the dimension reduction to the two-dimensional plane.

  After doing the above, a solution can be obtained by the steepest descent method. The solution obtained in this way obtains a representative point of the feature quantity after dimension reduction as a point arranged on a concentric circle around the origin on the two-dimensional plane. The procedure for acquiring subsequent charts based on this point is the same as in the first embodiment. Examples of error square sum functions other than Expression (8) are shown in Expressions (11) and (12). These are all functions used in multidimensional scaling.

  In addition, the error sum-of-squares function to be minimized described in Expression (8), Expression (11), and Expression (12) is an error for maintaining all the relationships of the feature amount G before dimension reduction. Yes. If a parameter such as approaching to the vicinity of k is added, dimension reduction can be performed in the same procedure as ISOMAP. The procedure for generating a chart based on g, which is a representative point representing each feature quantity after dimension reduction obtained by the above procedure, can be realized by the same procedure as in the first embodiment, and an appropriate chart can be obtained. it can.

(Third embodiment)
Next, a third embodiment of the present invention will be described. In the third embodiment, there is a plurality of data expressed by high-dimensional feature amounts, and when it is desired to analyze how each data has a tendency, it is intuitively understood by humans. This is a method of displaying a parallel coordinate plot or a bar graph on an easy display. Using the algorithm described in the first embodiment or the second embodiment to determine the spacing and order of radar chart axes, the order and spacing of parallel coordinate plots and bar graph items are similarly determined. can do. In addition, since the axes can be arranged in a circular shape in the radar chart, when generating a graph with n items in parallel coordinate plots and bar graphs, 1 is placed next to the n-th item at the end. A graph from the first item may be further drawn.

(Other embodiments of the present invention)
The present invention can also be realized by executing the following processing. That is, software (program) that realizes the functions of the above-described embodiments is supplied to a system or apparatus via a network or various storage media, and a computer (or CPU, MPU, or the like) of the system or apparatus reads the program. It is a process to be executed.

  For example, the information processing apparatus in each embodiment described above has a computer function 1100 as shown in FIG. 11, and the CPU 1101 performs operations in each embodiment. As shown in FIG. 11, the computer function 1100 includes a CPU 1101, a ROM 1102, and a RAM 1103. Also, a controller (CONSC) 1105 of the operation unit (CONS) 1109 and a display controller (DISPC) 1106 of a display (DISP) 1110 as a display unit such as an LCD are provided. Furthermore, a hard disk (HD) 1111, a controller (DCONT) 1107 of a storage device (STD) 1112 such as a flexible disk, and a network interface card (NIC) 1108 are provided. The functional units 1101, 1102, 1103, 1105, 1106, 1107, and 1108 are configured to be communicably connected to each other via the system bus 1104.

  The CPU 1101 performs overall control of each component connected to the system bus 1104 by executing software stored in the ROM 1102 or the HD 1111 or software supplied from the STD 1112. That is, the CPU 1101 reads out and executes a processing program for performing the operation as described above from the ROM 1102, the HD 1111 or the STD 1112, thereby performing control for realizing the operation in each embodiment. The RAM 1103 functions as a main memory or work area for the CPU 1101.

  The CONSC 1105 controls an instruction input from the CONS 1109. The DISPC 1106 controls the display of the DISP 1110. The DCONT 1107 controls access to the HD 1111 and the STD 1112 that store a boot program, various applications, user files, a network management program, a processing program for realizing operations in the embodiments, and the like. The NIC 1108 exchanges data bidirectionally with other devices on the network 1113.

  The above-described embodiments are merely examples of implementation in carrying out the present invention, and the technical scope of the present invention should not be construed as being limited thereto. That is, the present invention can be implemented in various forms without departing from the technical idea or the main features thereof.

101: Multivariate data input unit 102: Distance between feature amount calculation unit 103: Dimension reduction unit 104: Data visualization parameter input unit 105: Output unit 106: Chart shape determination unit 107: Multivariate data addition input unit 108: Chart shape storage Part

Claims (7)

A calculation means for calculating a similarity between a plurality of feature amounts extracted from input multivariate data;
Dimension reduction means for reducing the number of dimensions based on the similarity between the feature quantities calculated by the calculation means and the multivariate data;
Based on the result of dimension reduction by the dimension reduction means, shape determining means for determining the arrangement of the axes corresponding to each of the plurality of feature amounts;
An information processing apparatus comprising: output means for outputting a drawing in which the multivariate data is drawn according to the arrangement of the axes determined by the shape determining means.   The information processing apparatus according to claim 1, wherein the calculation unit calculates a similarity between the feature amounts based on a distribution of data included in the multivariate data.   3. The information processing according to claim 1, wherein the shape determining unit arranges axes corresponding to the plurality of feature amounts so as to form an angle corresponding to the similarity between the feature amounts. apparatus.   The information processing apparatus according to claim 1, wherein an arrangement of axes corresponding to each of the plurality of feature amounts is changed according to an input from a user.   5. The information processing apparatus according to claim 1, wherein the axes corresponding to the feature amounts having a similarity between the feature amounts higher than a predetermined value are collectively arranged. A calculation step of calculating a similarity between a plurality of feature amounts extracted from input multivariate data;
A dimension reduction step of reducing the number of dimensions based on the calculated similarity between the feature quantities and the multivariate data;
Based on the result of dimension reduction, a shape determining step for determining the arrangement of the axes corresponding to each of the plurality of feature amounts;
And an output step of outputting a diagram in which the multivariate data is drawn according to the determined arrangement of the axes. A calculation step for calculating a similarity between a plurality of feature amounts extracted from input multivariate data;
A dimension reduction step of reducing the number of dimensions based on the calculated similarity between the feature quantities and the multivariate data;
A shape determining step for determining an arrangement of axes corresponding to each of the plurality of feature amounts based on a result of dimension reduction;
A program for causing a computer to execute an output step of outputting a diagram in which the multivariate data is drawn according to the determined arrangement of axes.

Images