Information processing devices, computer-readable recording media, and information processing methods

Abstract

The system includes: a distance matrix calculation unit (104) that generates a similarity matrix based on a dataset containing multiple samples, the similarity matrix having multiple columns and multiple rows arranged in the order of the multiple samples, and storing the similarity between samples corresponding to one column and one row in a column defined by one of the multiple columns and one of the multiple rows; an order adjustment unit (105) that, with reference to label information representing the result of labeling the multiple samples using multiple labels each having multiple categories, adjusts the order in the similarity matrix so that the multiple samples are arranged according to multiple categories in the label specified from the multiple labels, i.e., object labels, thereby generating an adjusted similarity matrix; a visualization unit (106) that generates an evaluation screen image showing each column of the adjusted similarity matrix with a brightness corresponding to the similarity; and a display unit (107) that displays the evaluation screen image.

Description

Technical Field

[0001] This invention relates to an information processing apparatus, a computer-readable recording medium, and an information processing method. Background Technology

[0002] In recent years, with the development of measurement technology, various types of data have been measured and utilized. For example, when diagnosing equipment malfunctions, vibration sensors or microphones are installed in the equipment to measure the vibrations and sounds received from it. The data obtained in this way is used to diagnose equipment malfunctions.

[0003] When processing such measurement data, the quality of the data can sometimes be degraded due to deviations in the measurement environment or the sensor itself. For example, consider the following situation: differences in sensor installation methods due to different measurers can alter the nature of the data. When data deviations are significant due to factors unrelated to the intended purpose, it becomes an obstacle to identifying tags that were intended to be identified as abnormal or normal.

[0004] Patent document 1 describes a correlation analysis device that divides time series data into data for each specified time unit, calculates the similarity of each time unit, and displays the correlation of each specified time based on the calculated similarity at a predetermined position on a two-dimensional plane.

[0005] Existing technical documents

[0006] Patent documents

[0007] Patent Document 1: Japanese Patent Application Publication No. 2015-225637 Summary of the Invention

[0008] The problem that the invention aims to solve

[0009] The correlation analysis device described in Patent Document 1 can display the correlation based on the similarity of each time unit. Therefore, data deviations caused by time variations are clearly defined. However, it does not consider factors other than time, such as the "measurer" or the "individual device."

[0010] Therefore, one or more methods of the present invention can identify data deviations caused by factors unrelated to the original purpose.

[0011] Methods for solving problems

[0012] The information processing apparatus of the first aspect of the present invention is characterized in that it comprises: a similarity matrix generation unit that generates a similarity matrix based on a dataset having multiple samples, the similarity matrix having multiple columns arranged horizontally in the order in which the multiple samples are arranged and multiple rows arranged vertically in the same order, wherein a column defined by one column of the multiple columns and one row of the multiple rows stores the similarity between samples corresponding to the one column and samples corresponding to the one row; an order adjustment unit that, with reference to label information representing the result of labeling the multiple samples using multiple labels, each having multiple categories, adjusts the order in the similarity matrix such that the multiple samples are arranged according to the multiple categories in the label specified from the multiple labels, i.e., object labels, thereby generating an adjusted similarity matrix; a visualization unit that generates an evaluation screen image displaying each column of the adjusted similarity matrix with a brightness corresponding to the similarity; and a display unit that displays the evaluation screen image.

[0013] The information processing apparatus of the second aspect of the present invention is characterized in that the information processing apparatus comprises: a deviation calculation unit that, with reference to label information representing the result of labeling multiple samples contained in a dataset using multiple labels, each having multiple categories, calculates the degree of deviation of the multiple samples caused by the multiple labels, and generates a deviation result image displaying at least one label included in the multiple labels and the degree of deviation corresponding to the at least one label; and a display unit that displays the deviation result image.

[0014] The information processing apparatus of the third aspect of the present invention is characterized in that the information processing apparatus comprises: a deviation calculation unit that, with reference to label information representing the result of labeling multiple samples contained in a dataset using multiple labels, each having multiple categories, calculates the degree of deviation of the multiple samples due to the multiple categories, and generates a deviation result screen image displaying the multiple categories and the degree of deviation corresponding to the multiple categories; and a display unit that displays the deviation result screen image.

[0015] The program of the first aspect of the present invention is characterized in that the program enables a computer to function as the following components: a similarity matrix generation unit that generates a similarity matrix based on a dataset having multiple samples, the similarity matrix having multiple columns arranged horizontally in the order in which the multiple samples are arranged and multiple rows arranged vertically in the same order, wherein a column defined by one column of the multiple columns and one row of the multiple rows stores the similarity between samples corresponding to the one column and samples corresponding to the one row; an order adjustment unit that, referring to label information representing the result of labeling the multiple samples using multiple labels, each having multiple categories, adjusts the order in the similarity matrix such that the multiple samples are arranged according to the multiple categories in the label specified from the multiple labels, i.e., the object label, thereby generating an adjusted similarity matrix; a visualization unit that generates an evaluation screen image displaying each column of the adjusted similarity matrix with a brightness corresponding to the similarity; and a display unit that displays the evaluation screen image.

[0016] The program of the second aspect of the present invention is characterized in that the program enables a computer to function as the following components: a bias calculation unit that, with reference to label information representing the results of labeling multiple samples contained in a dataset using multiple labels, each having multiple categories, calculates the degree of bias of the multiple samples caused by the multiple labels, generates a bias result image displaying at least one label included in the multiple labels and the degree of bias corresponding to the at least one label; and a display unit that displays the bias result image.

[0017] The program of the third aspect of the present invention is characterized in that the program enables a computer to function as the following components: a deviation calculation unit, which refers to label information representing the result of labeling multiple samples contained in the dataset using multiple labels, each having multiple categories, calculates the degree of deviation of the multiple samples due to the multiple categories, generates a deviation result screen image displaying the multiple categories and the degree of deviation corresponding to the multiple categories; and a display unit that displays the deviation result screen image.

[0018] The information processing method of the first aspect of the present invention is characterized in that a similarity matrix is ​​generated based on a dataset having multiple samples, the similarity matrix having multiple columns arranged horizontally in the order of the multiple samples and multiple rows arranged vertically in the same order, a column defined by one column of the multiple columns and one row of the multiple rows storing the similarity between the samples corresponding to the one column and the samples corresponding to the one row, referring to label information representing the result of labeling the multiple samples using multiple labels each having multiple categories, the order in the similarity matrix is ​​adjusted in such a way that the multiple samples are arranged according to the multiple categories in the labels specified from the multiple labels, i.e., object labels, thereby generating an adjusted similarity matrix, generating an evaluation screen image showing each column of the adjusted similarity matrix with a brightness corresponding to the similarity, and displaying the evaluation screen image.

[0019] The information processing method of the second aspect of the present invention is characterized in that, referring to label information representing the result of labeling multiple samples contained in the dataset using multiple labels having multiple categories respectively, the method calculates the degree of deviation of the multiple samples caused by the multiple labels respectively, generates a deviation result image displaying at least one label contained in the multiple labels and the degree of deviation corresponding to the at least one label, and displays the deviation result image.

[0020] The information processing method of the third aspect of the present invention is characterized in that, referring to label information representing the result of labeling multiple samples contained in the dataset using multiple labels each having multiple categories, the method calculates the degree of deviation of the multiple samples caused by the multiple categories, generates a deviation result image displaying the multiple categories and the degree of deviation corresponding to the multiple categories, and displays the deviation result image.

[0021] Invention Effects

[0022] According to one or more methods of the present invention, it is possible to identify data deviations caused by factors unrelated to the original purpose. Attached Figure Description

[0023] Figure 1 This is a block diagram that schematically illustrates the structure of the data visualization device according to Embodiment 1.

[0024] Figure 2 (A) and (B) are graphs showing the distance between the distributions of the two samples.

[0025] Figure 3 This is a schematic diagram showing an example of a distance matrix.

[0026] Figure 4 This is a schematic diagram showing the first example of adjusting the distance matrix.

[0027] Figure 5 (A) and (B) are schematic diagrams showing the second and third examples of the adjusted distance matrix.

[0028] Figure 6 (A) and (B) are schematic diagrams showing examples 4 and 5 of the adjusted distance matrix.

[0029] Figure 7 This is a schematic diagram showing the sixth example of adjusting the distance matrix.

[0030] Figure 8 This is a schematic diagram of the first example of an evaluation screen image that visualizes the adjusted distance matrix.

[0031] Figure 9 This is a schematic diagram of the second example of an evaluation screen image that visualizes the adjusted distance matrix.

[0032] Figure 10 This is a schematic diagram of the third example of an evaluation screen image that visualizes the adjusted distance matrix.

[0033] Figure 11 This is a schematic diagram of the fourth example of an evaluation screen image that visualizes the adjusted distance matrix.

[0034] Figure 12 This is a schematic diagram of the fifth example of an evaluation screen image that visualizes the adjusted distance matrix.

[0035] Figure 13 This is a block diagram illustrating an example of the hardware structure of a data visualization device.

[0036] Figure 14 This is a flowchart illustrating the processing in the data visualization device of Embodiment 1.

[0037] Figure 15 This is a block diagram that schematically illustrates the structure of the data visualization device in Embodiment 2.

[0038] Figure 16 This is a schematic diagram showing an example of a partial matrix.

[0039] Figure 17 It is a schematic diagram used to illustrate the calculation method.

[0040] Figure 18 This is a flowchart illustrating the first process in the data visualization device of Embodiment 2.

[0041] Figure 19 This is a flowchart illustrating the second process in the data visualization device of Embodiment 2.

[0042] Figure 20 This is a block diagram that schematically illustrates the structure of the data visualization device in Embodiment 3.

[0043] Figure 21 This is a flowchart illustrating the processing in the data visualization device of Embodiment 3. Detailed Implementation

[0044] Implementation Method 1

[0045] Figure 1 This is a block diagram that schematically illustrates the structure of the data visualization device 100, which is an information processing device according to Embodiment 1.

[0046] The data visualization device 100 includes an input unit 101, a storage unit 102, a data acquisition unit 103, a distance matrix calculation unit 104, an order adjustment unit 105, a visualization unit 106, and a display unit 107.

[0047] The input unit 101 accepts input from user instructions. For example, the input unit 101 accepts input that specifies one label as the object label from a plurality of labels.

[0048] Storage unit 102 stores the programs and data required for processing in the data visualization device 100. For example, storage unit 102 stores the dataset of the visualized object and its label information.

[0049] A dataset is, for example, a collection of time-series data obtained through measurement. A dataset consists of a collection of multiple samples.

[0050] The samples constituting the dataset are units of a distance matrix calculated by the distance matrix calculation unit 104. The samples are, for example, time-series data obtained from a vibration sensor or microphone.

[0051] Here, a sample is one-dimensional data showing a single value measured at various times, but it is not limited to one-dimensional data. A sample can also be multi-dimensional data showing multiple values ​​at various times.

[0052] Label information represents the attributes associated with each sample in the dataset. Label information represents the results of labeling multiple samples using multiple labels. Each label has multiple categories.

[0053] Here, as an example, we will explain the labeling information associated with samples of measurement data collected for the purpose of diagnosing equipment malfunctions. Furthermore, the purpose of the measurement data is to distinguish between abnormal and normal conditions based on the measurement data itself.

[0054] In this case, each sample is assigned label information containing a label representing the inspection result. The label representing the inspection result can consist of two categories: "Normal Category" and "Abnormal Category," or it can consist of three or more categories, such as "Normal Category," "First Abnormal Category," and "Second Abnormal Category," depending on the type of abnormality. Here, the category represents a state contained within the same label. For example, different categories represent different states contained within the same label.

[0055] Furthermore, measurement data sometimes contains labels that are unrelated to the intended purpose of the measurement data (in this case, normal or abnormal). Examples include "measurer," "location of measurement," or "date and time of measurement."

[0056] The data acquisition unit 103 acquires the dataset DS and its label information LI from the storage unit 102. The acquired dataset DS and its label information LI are provided to the distance matrix calculation unit 104.

[0057] The distance matrix calculation unit 104 is a similarity matrix generation unit that generates a distance matrix as a similarity matrix based on the dataset DS. This similarity matrix has multiple columns arranged horizontally in the order of multiple samples and multiple rows arranged vertically in the same order. In a column defined by one column and one row, the similarity between samples corresponding to that column and samples corresponding to that row is stored. Furthermore, the similarity contained in the distance matrix is ​​also referred to as the elements constituting the distance matrix.

[0058] For example, the distance matrix calculation unit 104 uses a combination of all samples contained in the dataset DS to calculate the distance between samples, and generates a distance matrix with the number of rows and columns set to the number of samples, based on the calculated distance between samples.

[0059] The distance between samples is an indicator of the similarity between two samples. The more similar the two samples are, the smaller the distance between them; when the samples are completely identical, the distance between them is 0. Therefore, the distance matrix is ​​also called the similarity matrix.

[0060] The vibration data collected for the purpose of anomaly diagnosis will be used as an example below.

[0061] Measurement data obtained under certain conditions and over a certain period of time will be processed as a single sample. In this example, a single sample is time-series data showing one-dimensional values ​​at various moments. Each sample is then assigned a "test result label" indicating whether the equipment inspection result is normal or abnormal, and a "measurement date label" indicating when the measurement was performed.

[0062] Furthermore, here, it is assumed that two normal samples and two abnormal samples were measured on June 12, which serves as the measurement date label, and two normal samples and two abnormal samples were measured on June 13, which also serves as the measurement date label.

[0063] In the above case, an example is shown where the distance matrix calculation unit 104 actually calculates the distance between samples and generates a distance matrix. In this example, the similarity between the probability distributions of two samples is set as the similarity between the two samples.

[0064] Specifically, the distance matrix calculation unit 104 calculates the distribution formed by each sample. At this point in time, a distribution is calculated for each sample. As a hypothetical distribution, a normal distribution, etc., can be considered.

[0065] Next, the distance matrix calculation unit 104 calculates the distance between the distributions formed by the two samples as an indicator of the similarity between the samples. Methods for calculating the distance between distributions include the Bad Zaria distance, the Kullback-Leibler divergence, and the Jensen-Shannon divergence.

[0066] Figure 2 (A) and (B) are graphs showing the distance between the distributions of the two samples.

[0067] Figure 2 (A) represents the case where the distance between the distributions of the two samples is large. Figure 2 (B) represents the case where the distance between the distributions of the two samples is small.

[0068] Next, the distance matrix calculation unit 104 generates a distance matrix based on the distances between distributions calculated using the combination of all samples. The horizontal columns and vertical rows of the distance matrix are arranged according to the order of the eight samples.

[0069] Figure 3 This is a schematic diagram showing an example of a distance matrix.

[0070] Here, as described above, there are eight samples within two days. Let each sample be assigned a sample number used to determine its identification information. Then, the elements in the N rows and M columns of the distance matrix represent the distance between the Nth sample and the Mth sample. Here, N and M are integers greater than 1 and less than 8.

[0071] Figure 3 The values ​​on the diagonal of the distance matrix shown are the distances between identical samples, therefore, all are 0.

[0072] When a symmetric method such as the Bad Zaria distance is used to calculate the distance between samples, the distance value remains unchanged even if the samples are swapped; therefore, the distance matrix becomes a symmetric matrix.

[0073] The order adjustment unit 105 refers to the label information LI and adjusts the order of multiple samples in the similarity matrix according to the multiple categories specified from multiple labels, i.e., object labels, thereby generating an adjusted similarity matrix.

[0074] Specifically, the order adjustment unit 105 adjusts the order of the samples constituting the distance matrix generated by the distance matrix calculation unit 104 according to the specified tags. Specifically, as described above, when "inspection result tag" and "measurement date tag" are included as tag information, the order adjustment unit 105 adjusts the order of the samples of the distance matrix according to the "inspection result tag" and "measurement date tag" respectively.

[0075] For example, when rearranging the distance matrix according to the inspection result labels, the inspection result labels include two categories: "normal category" and "abnormal category." Therefore, the order adjustment unit 105 adjusts the order of the samples in a manner that arranges them according to each category, such that samples that initially belong to the normal category (normal samples) and then become samples that belong to the abnormal category (abnormal samples). For example, Figure 3 As shown, when the samples with odd numbers are in the normal category and the samples with even numbers are in the abnormal category among the eight samples, the order adjustment unit 105 adjusts the order of the samples to make them so that... Figure 3 The distance matrix shown becomes Figure 4 The adjusted distance matrix is ​​shown below. Additionally, the adjusted distance matrix is ​​also called the adjusted similarity matrix.

[0076] Furthermore, when rearranging the distance matrix according to the measurement date labels, which include two categories, "June 12th Category" and "June 13th Category," the order adjustment unit 105 changes the order of the samples so that samples initially belong to the June 12th category and then become samples belonging to the June 13th category. For example, Figure 3 As shown, when the sample numbers of samples 1-4 are from the June 12th category and samples 5-8 are from the June 13th category, the order adjustment unit 105 directly adjusts the order of samples. Figure 3 The distance matrix shown is set as an adjusted distance matrix.

[0077] The visualization unit 106 generates an evaluation screen image that makes the adjustment distance matrix visible, and displays the evaluation screen image on the display unit 107, thereby providing a prompt to the user.

[0078] For example, the visualization unit 106 determines the brightness of the column color corresponding to the element based on the values ​​of the elements contained in the adjustment distance matrix after the order adjustment unit 105 has adjusted the order, thereby generating an evaluation screen image. In this way, by changing the brightness of the column corresponding to the element of the adjustment distance matrix according to the value of the column, it is possible to visually represent the quality problems of the dataset caused by the labels used in the order change.

[0079] Specifically, the visualization unit 106 compares the value of the column in the distance matrix with a predetermined threshold, and if the value is less than the predetermined threshold, makes the color of that column darker than the color of the column with a value above the predetermined threshold.

[0080] Before explaining the effects obtained by evaluating the images, let's first explain what constitutes a "good dataset".

[0081] Let's assume this dataset is used for anomaly diagnosis. Therefore, it's preferable that anomalous and normal samples have properties that are as different as possible. On the other hand, regarding the label—the measurement date label—which is expected to be irrelevant to the diagnosis of anomalies and normalities, it's expected that the characteristics of the samples will not change regardless of the category. Based on the above, the "good dataset" in this example can be considered a dataset that satisfies the following two conditions.

[0082] The first condition is that the distance between samples of different categories in the inspection result labels is large. In other words, the similarity between samples of different categories in the inspection result labels is low.

[0083] The second condition is that the distances between different samples in the measurement day label are not clustered. In other words, in the measurement day label, regardless of whether the categories are the same or different, the similarity between samples is either low or high.

[0084] This allows visualization of the adjusted distance matrix after reordering the categories contained in a specified label, thus enabling visual confirmation that the dataset meets the two conditions mentioned above.

[0085] For example, Figure 5 As shown in (A), in the inspection result label, the order of samples is adjusted by dividing them into normal category and abnormal category. Thus, when the distance between samples in the normal category and samples in the abnormal category is less than a predetermined threshold and the corresponding column color darkens, and the distance between samples in the normal category and samples in the abnormal category is greater than a predetermined threshold and the corresponding column color brightens, it can be confirmed that the above-mentioned first condition is met.

[0086] On the other hand, such as Figure 5As shown in (B), the order of the samples is adjusted by dividing them into categories for June 12th and June 13th in the measurement date label. Thus, when the brighter and darker columns are not concentrated among samples of the same category and samples of different categories, it can be confirmed that the second condition mentioned above is met.

[0087] In contrast, for example, Figure 6 As shown in (A), the order of samples is adjusted by dividing them into normal and abnormal categories in the inspection result label. As a result, the first condition above is not satisfied when the brighter and darker columns are not clustered between samples of the same category and between samples of different categories.

[0088] Moreover, such as Figure 6 As shown in (B), in the measurement date label, the order of the samples is adjusted by dividing them into June 12th and June 13th categories. As a result, if the distance between the samples in the June 12th category and the distance between the samples in the June 13th category are less than a predetermined threshold and the color of the corresponding column is darkened, and the distance between the samples in the June 12th category and the samples in the June 13th category are greater than the predetermined threshold and the color of the corresponding column is brightened, the second condition mentioned above is not satisfied.

[0089] In this case, it can be determined that the data deviation during the measurement day is the cause of the deterioration in the inspection results.

[0090] Furthermore, the above example illustrates how the order of samples can be adjusted according to the categories contained in a label, but this is not an isolated case.

[0091] For example, it can also be like Figure 6 As shown in (A), after adjusting the order of samples by dividing them into normal and abnormal categories in the inspection result label, the samples contained in the normal and abnormal categories are further divided into the June 12th and June 13th categories contained in the measurement date label, and the order of samples is adjusted.

[0092] In this case, for example, Figure 7 As shown, within the normal and abnormal categories, the samples were further classified into the June 12th category and the June 13th category, respectively.

[0093] At this time, as Figure 7 As shown, in each category of the normal or abnormal category, when the brightness of the matrix elements is concentrated due to the difference in the measurement date (the difference in the measurement date label), it can be determined that the data deviation of the measurement date is the cause of the deterioration of the inspection results.

[0094] Describe the methods for interpreting visual evaluation images.

[0095] Here, in Figures 8-12The text provides four examples of how a dataset with nine samples will be visualized as an evaluation image by adjusting the distance matrix after reordering the labels of three categories.

[0096] In addition, Figures 8-12 In this process, we assign numbers 1 to 3 to the three categories respectively, and arrange the samples contained in the distance matrix in ascending order to generate the adjusted distance matrix.

[0097] like Figure 8 As shown in the adjusted distance matrix 120, when only the diagonal partitions have low brightness and the other partitions have high brightness, samples of the same category are similar to each other, but dissimilar to samples of other categories. Therefore, data bias arises due to the corresponding labels. Furthermore, for example, in the case of two categories, such as... Figure 6 As shown in (B).

[0098] like Figure 9 As shown in the distance adjustment matrix 121, without any change in brightness due to category differences and with all columns in a darker color, it indicates that all data are similar, and therefore, there is no data bias caused by labels.

[0099] like Figure 10 As shown in the adjusted distance matrix 122, when all columns except those containing the same samples on the diagonal are a brighter color, it can be seen that all data are biased regardless of class differences. In this case, there is no data bias caused by labels; however, it indicates that the dataset is completely uncentralized, and therefore, there is a high possibility that the data quality is problematic.

[0100] like Figure 11 As shown in the distance adjustment matrix 123, except for the diagonal partitions (refer to...) Figure 8 In addition, when the brightness of the column representing the values ​​between samples in category 1 and samples in category 3 decreases, it can be seen that only the data in category 2 has properties that are different from the data in other categories.

[0101] For example, when the data measurers are categorized, it is visually possible to identify that the data of the measurer corresponding only to category 2 (let's call it measurer 2 here) differs from the data of other measurers (measurer 1 and measurer 3). This allows us to examine factors contributing to data quality degradation, such as potential problems with measurer 2's measurement method.

[0102] like Figure 12As shown in the adjusted distance matrix 124, it can be seen that when samples belonging to category 1 are similar to samples belonging to category 3, and samples belonging to category 2 are not similar to samples belonging to categories 1 and 3, and different samples belonging to category 2 are also not similar to each other, only the data of category 2 is different from the data of categories 1 and 3, and the data of category 2 also have deviations from each other.

[0103] Additionally, when performing visualization, such as Figures 5-7 As shown, information representing categories can also be displayed on the left and top sides of the adjusted distance matrix to determine the order of the categories of the labels used in the rearrangement.

[0104] Figure 13 This is a block diagram illustrating an example of the hardware structure of the data visualization device 100.

[0105] The data visualization device 100 can be composed of a computer 130 having an input device 131, a display device 132, a storage device 133, a memory 134, and a processor 135.

[0106] The input device 131 is a device that allows the user to input instructions, such as a mouse, touch panel, keyboard, gesture input device for HMD (Head Mounted Display), or eye-operated input device.

[0107] Display device 132 is a display application device such as a monitor that provides prompts to the user during visualization. Display device 132 may also include, for example, a perspective display such as an HMD.

[0108] Storage device 133 is a storage device, including HDD (Hard Disk Drive) or SSD (Solid State Drive), etc.

[0109] The memory 134 is a device for temporary storage, including RAM (Random Access Memory).

[0110] Processor 135 is a processing circuit such as CPU (Central Processing Unit).

[0111] For example, the input unit 101 can be implemented by the processor 135 using the input device 131.

[0112] The display unit 107 can be implemented using the display device 132 via the processor 135.

[0113] The storage unit 102 can be implemented by the processor 135 using the memory 134.

[0114] The processor 135 reads the program stored in the storage device 133 into the memory 134 and executes the program, thereby enabling the data acquisition unit 103, the distance matrix calculation unit 104, the order adjustment unit 105, and the visualization unit 106 to be implemented.

[0115] Such programs can be provided via a network, or they can be provided recorded on a recording medium. That is, such programs can be provided, for example, as program products.

[0116] Figure 14 This is a flowchart illustrating the processing in the data visualization device 100 of Embodiment 1.

[0117] First, the data acquisition unit 103 acquires the dataset DS and its label information LI from the storage unit 102 (S10). The acquired dataset DS and its label information LI are then provided to the distance matrix calculation unit 104.

[0118] The distance matrix calculation unit 104 generates a distance matrix composed of the similarity of each sample constituting the dataset DS (S11). For example, the distance matrix calculation unit 104 calculates the distance between samples using a combination of all samples contained in the dataset DS, and generates a distance matrix with the number of rows and columns set to the number of samples, respectively, for the calculated distance between samples.

[0119] Next, the input unit 101 adjusts the order of samples in the distance matrix based on the user's assigned label (S12). The user information UI indicating the assigned label is provided to the order adjustment unit 105. The label assigned here is also referred to as the object label.

[0120] The order adjustment unit 105 adjusts the distance matrix generated by the distance matrix calculation unit 104 in such a way that the samples are arranged according to the categories of the specified labels, thereby generating an adjusted distance matrix (S13).

[0121] The visualization unit 106 generates an evaluation screen image that visualizes the adjustment distance matrix, and the display unit 107 displays the evaluation screen image (S14). For example, the visualization unit 106 determines the brightness of the color of the column containing the value based on the value contained in the adjustment distance matrix after the order is changed by the order adjustment unit 105, thereby generating the evaluation screen image.

[0122] Then, the sequence adjustment unit 105 determines whether to continue the visualization process (S15). For example, if the sequence adjustment unit 105 receives a user information UI indicating a label from the input unit 101, it determines to continue the visualization process. If the sequence adjustment unit 105 receives a user information UI indicating the end of the process from the input unit 101, it determines not to continue the visualization process. If the visualization process continues (S14: Yes), the process returns to step S12; if the visualization process does not continue (S14: No), the process ends.

[0123] As described above, according to Implementation 1, the order of samples in the distance matrix is ​​adjusted and visualized using arbitrary labels, thereby visually representing the bias in the dataset caused by the specified labels. This clarifies the quality issues of the dataset caused by the labels used in the adjustment.

[0124] Implementation Method 2

[0125] In Implementation 1, the results obtained by adjusting the order of samples in the distance matrix by specifying labels are visualized, thereby allowing the user to explain the data bias caused by the labels.

[0126] In implementation method 2, the degree of deviation of the data is pre-calculated to assist in specifying the labels whose order needs to be adjusted. Alternatively, the labels whose order needs to be adjusted are automatically specified based on the calculated degree of deviation.

[0127] Figure 15 This is a block diagram that schematically illustrates the structure of the data visualization device 200 according to Embodiment 2.

[0128] The data visualization device 200 includes an input unit 101, a storage unit 102, a data acquisition unit 103, a distance matrix calculation unit 104, a sequence adjustment unit 105, a visualization unit 106, a display unit 107, a deviation calculation unit 208, and a sequence adjustment auxiliary unit 209.

[0129] The input unit 101, storage unit 102, data acquisition unit 103, distance matrix calculation unit 104, order adjustment unit 105, visualization unit 106, and display unit 107 of the data visualization device 200 in Embodiment 2 are the same as those of the data visualization device 100 in Embodiment 1.

[0130] However, the data acquisition unit 103 also provides the acquired dataset DS and label information LI to the deviation calculation unit 208, and the distance matrix calculation unit 104 also provides the generated distance matrix to the deviation calculation unit 208.

[0131] The deviation calculation unit 208 calculates the degree of deviation of multiple samples as data due to each label.

[0132] In implementation 1, the user interprets the visualized adjustment distance matrix, thereby clarifying the quality issues. For example, as Figure 8 As shown in the distance adjustment matrix 120, if the brightness of the diagonal partitions of the same category that intersect each other is low, and the brightness of other partitions is high, it can be determined that there is a bias caused by the label.

[0133] The function of the deviation calculation unit 208 is to quantify the distance matrix as a numerical value, without requiring the user to interpret its characteristics. The deviation calculation unit 208 can quantify the degree of data deviation caused by specific labels. The specific processing of the deviation calculation unit 208 will be explained below.

[0134] Here, we will illustrate an example of calculating the data bias caused by a certain label for a dataset with N samples (N is an integer greater than 2).

[0135] Set the number of categories contained in the object label to C (C is an integer greater than 2), and represent each category as 1, 2, ..., C.

[0136] Furthermore, the adjusted distance matrix, which adjusts the samples of the distance matrix calculated based on this dataset in the order of 1, 2, ..., C, is denoted as M. Also, in this example, the distance matrix is ​​assumed to consist of the Bad Zaria distances between samples.

[0137] The number of elements in M ​​is N×N. Here, the samples in the distance matrix are adjusted using labels of C categories. Therefore, within the adjusted distance matrix M, there exists a partial matrix that forms C×C regions. For example... Figure 16 As shown, these partial matrices are represented as D 11 D 12 ... D CC .

[0138] D ij It becomes a partial matrix consisting of samples belonging to category i (where i is an integer satisfying 1≤i≤C) and samples belonging to category j (where j is an integer satisfying 1≤j≤C).

[0139] Let N be the number of samples belonging to category i. i Let N be the number of samples belonging to category j. j At that time, D ij Become N i ×N j The matrix D. ij The elements are denoted as d kl(ij) .

[0140] Here, the average value μ of the elements of the partial matrix is ​​calculated. ij Let D be the matrix of each part. ij The representative values ​​for each corresponding partition.

[0141] Based on the symmetry of the adjusted distance matrix M, in calculating μ ij When using Figure 17 The partition that is filled in. When i = j, D ij Since it is a symmetric matrix, the average is calculated using only the elements located to the upper right of the diagonal. Furthermore, μ ij =μ ji Therefore, partitions where i>j do not need to be calculated.

[0142] Equation (1) below is μ ij The calculation formula.

[0143]

Mathematical Formula 1

[0144]

[0145] Additionally, here, the average value μ is used. ij As a partial matrix D ij The representative value is given, but implementation 2 is not limited to this example. For example, a partial matrix D can also be used. ij The central value of the elements contained in the matrix is ​​used as a partial matrix D. ij The representative value.

[0146] As the average value μ ij The qualitative meaning of "likelihood" is that, when i ≠ j, it can be considered to represent the similarity between data belonging to category i and data belonging to category j. For example, in μ... 12 When smaller values ​​are shown, the data for category 1 and category 2 are similar, and it can be considered that there is no data bias due to the difference in categories. Furthermore, when i = j, the average value μ... ij This indicates the degree of deviation between data belonging to the same category.

[0147] In situations where data bias arises due to labeling, such as Figure 6 (B) or Figure 8 As shown, D is located on the diagonal of the distance matrix. ii The values ​​of the elements in the adjusted distance matrix decrease, while the values ​​of all other elements increase. Thus, the degree of data bias V caused by a certain label can be formulated, for example, as shown in equation (2) below.

[0148]

Mathematical Formula 2

[0149]

[0150] The left-hand side of equation (2) is the μ calculated using all combinations when i ≠ j. ij And calculate their average. This indicates how the data deviates when the categories are different.

[0151] The right-hand side of equation (2) is the μ calculated when i = J using all combinations. ij And calculate their average. This indicates how the data within the same category are biased.

[0152] like Figure 6 (B) or Figure 8 As shown, the value of V increases when data bias caused by labels occurs. Ideally, V should have a higher value for the identified object labels in the dataset (here, the inspection result labels) and a lower value for labels that are expected to be unrelated to those labels (here, the measurement date labels).

[0153] As described above, the bias calculation unit 208 calculates the bias degree V using all pre-given labels. This bias degree V is calculated by subtracting the average of the representative values ​​of the similarity between samples classified into different categories from the average of the representative values ​​of the similarity between samples classified into the same category. For example, if the dataset is assigned 5 labels, this processing yields 5 V values ​​corresponding to each label.

[0154] The sequence adjustment assist unit 209 causes the display unit 107 to display a deviation result screen image, which is the result obtained by the deviation calculation unit 208. This provides the user with the result and assists in specifying the labels used for adjustment. The deviation result screen image, for example, displays the labels in descending order of deviation degree V. The user can refer to this screen to specify the object labels.

[0155] The order adjustment assist unit 209 may not display the deviation result screen image, but instead notify the order adjustment unit 105 of the result obtained by the deviation calculation unit 208, automatically adjusting the order of the samples included in the distance matrix, thereby generating an adjustment distance matrix. For example, the order adjustment assist unit 209 may specify the label with the highest deviation degree as the object label, and generate the adjustment distance matrix according to the object label. Furthermore, the order adjustment assist unit 209 may sequentially specify multiple labels as object labels starting from the side with the highest deviation degree. Moreover, the order adjustment assist unit 209 may specify the label with the highest deviation degree and the label with the lowest deviation degree as object labels.

[0156] The data visualization device 200 in embodiment 2 can also be derived from... Figure 13 The computer 130 shown is configured as follows.

[0157] For example, the processor 135 reads the program stored in the storage device 133 into the memory 134 and executes the program, thereby enabling the deviation calculation unit 208 and the sequence adjustment auxiliary unit 209 to be implemented.

[0158] Figure 18 This is a flowchart illustrating the first process in the data visualization device 200 of Embodiment 2.

[0159] The first processing is the process in which the sequence adjustment auxiliary unit 209 causes the display unit 107 to display a deviation result screen image as the result obtained by the deviation calculation unit 208.

[0160] First, the data acquisition unit 103 acquires the dataset DS and its label information LI from the storage unit 102 (S20). The acquired dataset DS and its label information LI are provided to the distance matrix calculation unit 104 and the deviation calculation unit 208.

[0161] The distance matrix calculation unit 104 calculates a distance matrix (S21) consisting of the similarities of the samples constituting the dataset DS. The generated distance matrix is ​​provided to the order adjustment unit 105 and the bias calculation unit 208.

[0162] Next, the deviation calculation unit 208 repeatedly performs steps S23 and S24 (S22 and S25) based on the number of tags shown in the tag information LI.

[0163] In step S23, the deviation calculation unit 208 determines the labels from the labels shown in the label information LI that have not yet had their deviation degree V calculated, and adjusts the order of the samples contained in the distance matrix according to the determined labels.

[0164] In step S24, the deviation calculation unit 208 calculates the degree of data deviation V caused by the determined label based on the adjusted distance matrix, i.e., the adjusted distance matrix.

[0165] Next, the sequential adjustment auxiliary unit 209 causes the display unit 107 to display a screen image showing the deviation result, which is calculated by the deviation calculation unit 208 as the degree of deviation V (S26).

[0166] Next, the input unit 101 adjusts the order of samples contained in the distance matrix based on the user's assigned label (S27). The user information UI indicating the assigned label is provided to the order adjustment unit 105. The user can specify the label by referring to the deviation result screen image displayed on the display unit 107.

[0167] The order adjustment unit 105 adjusts the distance matrix generated by the distance matrix calculation unit 104 in such a way that the samples are arranged according to the categories of the specified labels, thereby generating an adjusted distance matrix (S28).

[0168] The visualization unit 106 generates an evaluation screen image that makes the adjustment distance matrix visible, and the display unit 107 displays the evaluation screen image (S29).

[0169] Then, the sequence adjustment unit 105 determines whether to continue the visualization process (S30). For example, if the sequence adjustment unit 105 receives a user information UI indicating a label from the input unit 101, it determines to continue the visualization process. If the sequence adjustment unit 105 receives a user information UI indicating the end of the process from the input unit 101, it determines not to continue the visualization process. If the visualization process continues (S30: Yes), the process returns to step S27; if the visualization process does not continue (S30: No), the process ends.

[0170] Figure 19 This is a flowchart illustrating the second process in the data visualization device 200 of Embodiment 2.

[0171] The second process is the process where the sequence adjustment auxiliary unit 209 assigns a label based on the deviation result obtained by the deviation calculation unit 208.

[0172] First, the data acquisition unit 103 acquires the dataset DS and its label information LI from the storage unit 102 (S40). The acquired dataset DS and its label information LI are provided to the distance matrix calculation unit 104 and the deviation calculation unit 208.

[0173] The distance matrix calculation unit 104 calculates a distance matrix composed of the similarities of the samples constituting the dataset DS (S41). The generated distance matrix is ​​provided to the order adjustment unit 105 and the bias calculation unit 208.

[0174] Next, the deviation calculation unit 208 repeatedly performs steps S43 and S44 (S42 and S45) based on the number of tags shown in the tag information LI.

[0175] In step S43, the deviation calculation unit 208 determines the labels from the labels shown in the label information LI that have not yet had their deviation degree V calculated, and adjusts the order of the samples contained in the distance matrix according to the determined labels.

[0176] In step S44, the deviation calculation unit 208 calculates the degree of data deviation V caused by the determined label based on the adjusted distance matrix, i.e., the adjusted distance matrix.

[0177] Next, the sequence adjustment auxiliary unit 209 assigns a label based on the deviation degree V calculated by the deviation calculation unit 208 (S46). For example, the sequence adjustment auxiliary unit 209 can assign the label with the largest deviation degree V.

[0178] Next, the sequence adjustment unit 105 adjusts the distance matrix generated by the distance matrix calculation unit 104 so that the samples are arranged according to the categories of the specified labels, thereby generating the adjusted distance matrix (S47).

[0179] The visualization unit 106 generates an evaluation screen image that makes the adjustment distance matrix visible, and the display unit 107 displays the evaluation screen image (S48).

[0180] As described above, according to Embodiment 2, the deviation calculation unit 208 calculates the data deviation caused by each label in advance, thereby enabling the adjustment of the samples contained in the distance matrix to be assisted or automated.

[0181] In addition, Figure 19 In step S46 of the flowchart shown, the sequential adjustment auxiliary unit 209 may assign multiple labels, thereby displaying multiple evaluation screen images on the display unit 107 via steps S47 and S48. In this case, the sequential adjustment auxiliary unit 209 may assign multiple labels sequentially, starting with the label indicating the highest deviation level V.

[0182] Furthermore, in implementation 2, automated tag assignment and user-based tag assignment can also be combined. For example, it can initially be as follows: Figure 19 The image shown is an evaluation screen image based on automatically assigned labels. Then, for example, a deviation result screen image is displayed via instructions from the user, with labels assigned by the user.

[0183] Implementation Method 3

[0184] Figure 20 This is a block diagram that schematically illustrates the structure of the data visualization device 300 according to Embodiment 3.

[0185] The data visualization device 300 includes a storage unit 102, a data acquisition unit 103, a distance matrix calculation unit 104, a display unit 107, and a deviation calculation unit 308.

[0186] The storage unit 102, data acquisition unit 103, distance matrix calculation unit 104, and display unit 107 of the data visualization device 300 in Embodiment 3 are the same as those of the data visualization device 100 in Embodiment 1.

[0187] Furthermore, compared to the data visualization device 200 of Embodiment 2, the data visualization device 300 of Embodiment 3 does not include an input unit 101, a sequence adjustment unit 105, a visualization unit 106, and a sequence adjustment auxiliary unit 209.

[0188] The deviation calculation unit 308 is similar to the deviation calculation unit 208 in Embodiment 2, and calculates the degree of data deviation caused by each of the labels shown in the label information LI.

[0189] Then, the deviation calculation unit 308 causes the display unit 107 to display a deviation result image that represents the calculated degree of deviation.

[0190] The deviation result image can be, for example, an image displaying a predetermined number of labels in descending order of deviation degree V, or an image displaying the deviation degree V of all labels. In other words, the deviation calculation unit 308 can display at least one label included in all the labels shown in the label information LI and the deviation degree corresponding to that at least one label in the deviation result image.

[0191] Figure 21 This is a flowchart illustrating the processing in the data visualization device 300 of Embodiment 3.

[0192] First, the data acquisition unit 103 acquires the dataset DS and its label information LI from the storage unit 102 (S50). The acquired dataset DS and its label information LI are provided to the distance matrix calculation unit 104 and the deviation calculation unit 308.

[0193] The distance matrix calculation unit 104 calculates a distance matrix composed of the similarities of the samples constituting the dataset DS (S51). The generated distance matrix is ​​provided to the order adjustment unit 105 and the bias calculation unit 308.

[0194] Next, the deviation calculation unit 308 repeatedly performs steps S53 and S54 (S52 and S55) based on the number of tags shown in the tag information LI.

[0195] In step S53, the deviation calculation unit 308 determines the labels from the labels shown in the label information LI that have not yet had their deviation degree V calculated, and adjusts the order of the samples contained in the distance matrix according to the determined labels.

[0196] In step S54, the deviation calculation unit 308 calculates the degree of data deviation V caused by the determined label based on the adjusted distance matrix, i.e., the adjusted distance matrix.

[0197] Next, the deviation calculation unit 308 causes the display unit 107 to display a deviation result image representing the calculated deviation degree V (S56).

[0198] According to implementation method 3, by prompting the user with the degree of deviation V caused by each label, it is possible to clarify the labels caused by the quality problems of the dataset.

[0199] As an example of the dataset processed in the data visualization devices 100-300 of the embodiments 1-3 described above, processing sound data collected for the purpose of automatically detecting processing defects during the cutting process of a laser processing machine can be cited. This data can be collected by installing a microphone as an acoustic sensor on the processing head of the processing machine. Since the purpose is to detect processing defects, each data point is assigned an "inspection result label" indicating whether a defect exists. Ultimately, it is expected that machine learning will be used to detect whether a processing defect exists based on the characteristics of the processing sound data.

[0200] As a preliminary stage for data learning, data visualization devices 100-300 are used to visualize and quantify the quality of data. When the data visualization devices 100-300 can identify quality degradation and its causes, countermeasures such as "spacing the data used in learning (in other words, not using data that leads to quality degradation)" or "improving data collection methods" can be taken based on the results, thereby improving data quality.

[0201] In addition, during data collection, information such as "measurement time", "measurer", "body number" or "measurement location" that is not expected to be related to the inspection results is also recorded. The degree of data deviation caused by each label is visualized and quantified through the processing described in Implementation Methods 1 to 3.

[0202] In embodiments 1 to 3, time-series data showing one-dimensional values ​​at each time step are processed. However, a single sample does not necessarily need to be data showing one-dimensional values ​​at each time step. Time-series data showing multi-dimensional values ​​at each time step, or data converted into multi-dimensional data by performing feature extraction processing on one-dimensional signals, can also be used. Furthermore, it does not need to be time-series data.

[0203] In embodiments 2 and 3, the deviation calculation units 208 and 308 calculate the degree of deviation for each "label". However, when focusing on a certain label, the deviation calculation units 208 and 308 can also calculate the "degree of deviation due to each category" contained in that label.

[0204] When the degree of data bias caused by category c (c is an integer satisfying 1≤c≤C) is set as V(c), for example, V(c) can be formulated as in equation (3) below.

[0205]

Mathematical Expression 3

[0206]

[0207] As shown in equation (3), the deviation V between categories out(c) Deviation V within the category in(c) V(c) is calculated from the difference between the two.

[0208] V out(c) V represents how samples of class c differ from samples of other classes. in(c) This indicates how the samples of category c differ from each other.

[0209] For example, in response to Figure 11 When V(c) is calculated using the adjusted distance matrix 123 for all categories 1 to 3, the value of V(2) is larger compared to V(1) and V(3). Therefore, it can be quantitatively shown that only samples of category 2 possess properties that differ from samples of other categories.

[0210] The preferred value of V(c) is higher in the categories contained in the identification object label (here, the inspection result label) of the dataset, and lower in the categories contained in the label that is expected to be unrelated to the label (here, the measurement date label).

[0211] When only the categories contained in labels that are not related to the original object of identification (such as the measurer, measurement date or body number) are set as the calculation objects, V(c) can also be calculated using the following formula (4).

[0212]

Mathematical Expression 4

[0213] V(c)=V out (c)+V in (c) (4)

[0214] That is, when one of the categories in which the deviation degree V(c) is calculated is set as the object category, the deviation calculation units 208 and 308 can also calculate the deviation degree V(c) by subtracting the representative value of the similarity between the sample classified as the object category and the sample classified as the category other than the object category from the average of the representative values ​​of the similarity between the sample classified as the object category and the sample classified as the category other than the object category.

[0215] In addition, the deviation calculation units 208 and 308 can also calculate the degree of deviation V(c) by adding the representative value of the similarity between the sample classified as the object category and the sample classified as the category other than the object category to the representative value of the similarity between the samples classified as the object category.

[0216] The sequence adjustment auxiliary unit 209 or the deviation calculation unit 308 can also generate a deviation result screen image that displays multiple categories and the deviation degree V(c) corresponding to each of the multiple categories, and display it on the display unit 107.

[0217] In this case, Figure 12 The increased median of category 2 in the adjusted distance matrix 124, where "data deviates significantly from other categories" and "data within the same category also deviates greatly from each other," indicates that this category is a factor contributing to quality degradation.

[0218] As described above, the deviation calculation units 208 and 308 can also calculate the degree of deviation V caused by the label and the degree of deviation V(c) caused by each category of the label for all given labels. In addition, the degree of deviation V is also referred to as the label deviation degree, and the degree of deviation V(c) is also referred to as the category deviation degree.

[0219] In addition, the partial matrix D that has been adjusted based on a certain label can also be processed. ii The calculations of V and V(c) described above are performed. In this case, the deviation calculation units 208 and 308 further rearrange the portion of matrix D located on the diagonal of the adjusted distance matrix M after adjustment by a certain label (e.g., the inspection result label) using a label other than the one described above. ii Given the samples included, calculate the re-adjusted distance matrix M. i Then, following the same steps, adjust the distance matrix M. i Calculate the degree of deviation V i and V i (c). At this point, for example, it is possible to take V. i and V i The final V and V(c) are calculated by averaging (c). This method can be used to calculate the degree of deviation of other labels when the labels with large deviations are known in advance. Therefore, the degree of deviation caused by other labels can be calculated without being affected by the initially adjusted labels.

[0220] As described above, the deviation calculation units 208 and 308 can also calculate "the degree of data deviation caused by the categories contained in a specific label".

[0221] For example, when a user specifies a label via the input unit 101, the deviation calculation unit 208 can also calculate the deviation degree V(c) of each category included in that label. Furthermore, the deviation calculation units 208 and 308 can also calculate the deviation degree V of each label and the deviation degree V(c) of each category, and display the calculation results, namely V and V(c), in the deviation result screen image.

[0222] In embodiments 1 and 2 described above, the values ​​of the columns in the adjustment distance matrix are compared with predetermined thresholds. If the values ​​are less than the predetermined threshold, the color of that column is darker than the color of columns with values ​​above the predetermined threshold, thereby generating an evaluation image. However, embodiments 1 and 2 are not limited to this example. For example, the adjustment distance matrix can also be visualized as a heatmap, using "color intensity" to represent the values ​​of the columns in the adjustment distance matrix. For example, a heatmap can be used where smaller values ​​correspond to darker colors, and larger values ​​correspond to brighter colors. Furthermore, heatmap-based representation and threshold-based representation can be combined. For example, a representation method can be adopted where columns exceeding a predetermined upper limit value are all "the brightest color in the heatmap (e.g., white)," columns below a predetermined lower limit value are all "the darkest color in the heatmap (e.g., black)," and the remaining columns are visualized according to the heatmap.

[0223] Label Explanation

[0224] 100, 200, 300: Data visualization device; 101: Input unit; 102: Storage unit; 103: Data acquisition unit; 104: Distance matrix calculation unit; 105: Sequence adjustment unit; 106: Visualization unit; 107: Display unit; 208, 308: Deviation calculation unit; 209: Sequence adjustment auxiliary unit.

Claims

1. An information processing apparatus, characterized by, The information processing apparatus has: a similarity matrix generation section that generates a similarity matrix from a data set having a plurality of samples, the similarity matrix having a plurality of columns arranged horizontally in an order in which the plurality of samples are arranged and a plurality of rows arranged vertically in the order, a degree of similarity between a sample corresponding to one of the columns and a sample corresponding to one of the rows being stored in a cell determined by the one column and the one row; an order adjustment section that adjusts the order in the similarity matrix in such a manner that the plurality of samples are arranged in a plurality of categories of a label specified from a plurality of labels used for labeling the plurality of samples respectively, each of the plurality of labels having the plurality of categories, thereby generating an adjusted similarity matrix; a visualization section that generates an evaluation screen image that shows each cell of the adjusted similarity matrix in luminance corresponding to the degree of similarity; and a display section that displays the evaluation screen image.

2. The information processing apparatus according to claim 1, characterized in that the information processing apparatus further has an input section that accepts designation of the object label from the plurality of labels.

3. The information processing apparatus according to claim 1, characterized in that the information processing apparatus further has: a bias calculation section that calculates a degree of bias of the plurality of samples due to the plurality of labels respectively; an order adjustment assistance section that generates a bias result screen image that displays the plurality of labels and the degree of bias corresponding to the plurality of labels respectively, and causes the display section to display the bias result screen image; and an input section that accepts designation of the object label from the plurality of labels.

4. The information processing apparatus according to claim 1, characterized in that the information processing apparatus further has a bias calculation section that calculates a degree of bias of the plurality of samples due to the plurality of labels respectively, and designates a label having the highest degree of bias as the object label.

5. The information processing apparatus according to claim 3 or 4, characterized in that the degree of bias is calculated by subtracting an average of representative values of degrees of similarity between samples classified into the same category from an average of representative values of degrees of similarity between samples classified into different categories among the plurality of categories.

6. The information processing apparatus according to claim 1, characterized in that the information processing apparatus further has: a bias calculation section that calculates a degree of bias of the plurality of samples due to the plurality of categories respectively; and an order adjustment assistance section that generates a bias result screen image that displays the plurality of categories and the degree of bias corresponding to the plurality of categories respectively, and causes the display section to display the bias result screen image.

7. The information processing apparatus according to claim 6, characterized in that the degree of bias is calculated by subtracting an average of representative values of degrees of similarity between samples classified into the same category from an average of representative values of degrees of similarity between samples classified into different categories among the plurality of categories. In a case where one of the plurality of categories in which the degree of bias is calculated is set as an object category, the degree of bias is calculated from an average of representative values of similarity between samples classified as the object category and samples classified as the plurality of categories other than the object category, minus representative values of similarity between samples classified as the object category.

8. The information processing apparatus according to claim 6, wherein In a case where one of the plurality of categories in which the degree of bias is calculated is set as an object category, the degree of bias is calculated from an average of representative values of similarity between samples classified as the object category and samples classified as the plurality of categories other than the object category, minus representative values of similarity between samples classified as the object category.

9. The information processing apparatus according to any one of claims 1 to 8, wherein The evaluation screen image shows the plurality of categories in the object label in correspondence with the plurality of columns and the plurality of rows of the adjusted similarity matrix.

10. An information processing apparatus, characterized by comprising: The information processing apparatus has: a bias calculation section that, with reference to label information indicating results of labeling a plurality of samples included in a data set using a plurality of labels each having a plurality of categories, calculates degrees of bias of the plurality of samples due to the plurality of categories respectively, and generates a bias result screen image showing the plurality of categories and the degrees of bias corresponding to the plurality of categories respectively; and a display section that displays the bias result screen image, In a case where one of the plurality of categories in which the degree of bias is calculated is set as an object category, the degree of bias is calculated from an average of representative values of similarity between samples classified as the object category and samples classified as the plurality of categories other than the object category, minus representative values of similarity between samples classified as the object category.

11. An information processing apparatus comprising: The information processing apparatus has: a bias calculation section that, with reference to label information indicating results of labeling a plurality of samples included in a data set using a plurality of labels each having a plurality of categories, calculates degrees of bias of the plurality of samples due to the plurality of categories respectively, and generates a bias result screen image showing the plurality of categories and the degrees of bias corresponding to the plurality of categories respectively; and a display section that displays the bias result screen image, In a case where one of the plurality of categories in which the degree of bias is calculated is set as an object category, the degree of bias is calculated from an average of representative values of similarity between samples classified as the object category and samples classified as the plurality of categories other than the object category, minus representative values of similarity between samples classified as the object category.

12. A computer-readable recording medium storing a computer program, which, when executed by a processor, generating a similarity matrix from a data set having a plurality of samples, the similarity matrix having a plurality of columns arranged horizontally in an order of the plurality of samples and a plurality of rows arranged vertically in the order, a similarity between a sample corresponding to one of the columns and a sample corresponding to one of the rows being stored in a cell determined by the one column and the one row, adjusting the order in the similarity matrix in such a manner that the plurality of samples are arranged in a plurality of categories of a label specified from a plurality of labels each having the plurality of categories, with reference to label information indicating results of labeling the plurality of samples respectively using the plurality of labels each having the plurality of categories, thereby generating an adjusted similarity matrix, generating an evaluation screen image showing each cell of the adjusted similarity matrix in luminance corresponding to the similarity, displaying the evaluation screen image.

13. A computer-readable recording medium storing a computer program, which, when executed by a processor, with reference to label information indicating results of labeling a plurality of samples included in a data set respectively using a plurality of labels each having a plurality of categories, calculating a degree of bias of the plurality of samples due to the plurality of categories respectively, generating a bias result screen image showing the plurality of categories and the degree of bias corresponding to the plurality of categories respectively, displaying the bias result screen image, in a case where one of the plurality of categories in which the degree of bias is calculated is set as an object category, calculating the degree of bias by subtracting a representative value of similarity among samples classified as the object category from an average of representative values of similarity between samples classified as the object category and samples classified as the plurality of categories other than the object category.

14. A computer-readable recording medium storing a computer program, which, when executed by a processor, with reference to label information indicating results of labeling a plurality of samples included in a data set respectively using a plurality of labels each having a plurality of categories, calculating a degree of bias of the plurality of samples due to the plurality of categories respectively, generating a bias result screen image showing the plurality of categories and the degree of bias corresponding to the plurality of categories respectively, displaying the bias result screen image, in a case where one of the plurality of categories in which the degree of bias is calculated is set as an object category, calculating the degree of bias by adding a representative value of similarity among samples classified as the object category to representative values of similarity between samples classified as the object category and samples classified as the plurality of categories other than the object category.

15. An information processing method, characterized by A similarity matrix is generated from a data set having a plurality of samples, the similarity matrix having a plurality of columns arranged horizontally in an order arranged by the plurality of samples and a plurality of rows arranged vertically in the order, a similarity between a sample corresponding to one of the columns and a sample corresponding to one of the rows being stored in a cell determined by the one column and the one row, The order in the similarity matrix is adjusted in such a manner that the plurality of samples are arranged by a plurality of categories in a label specified from a plurality of labels each having the plurality of categories, namely, an object label, with reference to label information indicating a result of labeling the plurality of samples respectively using the plurality of labels each having the plurality of categories, thereby generating an adjusted similarity matrix, An evaluation screen image in which each cell of the adjusted similarity matrix is shown in luminance corresponding to the similarity is generated, The evaluation screen image is displayed.

16. An information processing method characterized by, With reference to label information indicating a result of labeling a plurality of samples included in a data set respectively using a plurality of labels each having a plurality of categories, a degree of bias of the plurality of samples due to the plurality of categories respectively is calculated, An bias result screen image in which the plurality of categories and the degree of bias corresponding to the plurality of categories respectively are displayed is generated, The bias result screen image is displayed, In a case where one of the plurality of categories in which the degree of bias is calculated is set as an object category, a representative value of similarity between samples classified as the object category is subtracted from an average of representative values of similarity between samples classified as the plurality of categories other than the object category, thereby calculating the degree of bias.

17. An information processing method characterized by, With reference to label information indicating a result of labeling a plurality of samples included in a data set respectively using a plurality of labels each having a plurality of categories, a degree of bias of the plurality of samples due to the plurality of categories respectively is calculated, An bias result screen image in which the plurality of categories and the degree of bias corresponding to the plurality of categories respectively are displayed is generated, The bias result screen image is displayed, In a case where one of the plurality of categories in which the degree of bias is calculated is set as an object category, a representative value of similarity between samples classified as the object category is added to an average of representative values of similarity between samples classified as the plurality of categories other than the object category, thereby calculating the degree of bias.

Images