JP6081146B2 - Data monitoring device - Google Patents

Description

  Embodiments described herein relate generally to a data monitoring apparatus that monitors multivariate data.

  Collect and utilize various types of data, such as man-hours, source code scale, and number of defects, in order to accurately grasp the project status and implement appropriate management in large and complex product development projects It is important to do.

  In recent years, for example, various development support tools and development management systems such as source code static analysis tools and defect management systems have become widespread, and it has become possible to sequentially collect and accumulate large amounts of multivariate data related to projects. It was.

  Conventionally, control charts and cumulative graphs have been used as project data monitoring techniques for grasping the situation, but they are originally suitable for capturing univariate or a small number of variable time series changes. It is not always suitable as a monitoring method for multivariate data in which variables are related to each other.

  Project data that automatically narrows down candidates for outliers and outliers from a large amount of multivariate data and can easily identify factors that indicate outliers and outliers A monitoring technique is desired.

JP 2010-21113B JP 2004-349852 A JP 2008-146591 A JP 2008-129011 A JP 2001-75642 A

  The problem to be solved by the present invention is to provide a data monitoring apparatus capable of specifying a main factor of a specific sample.

Data monitoring apparatus embodiment, performed principal component analysis on multivariate data, and outlier candidate detection unit for detecting a candidate of large out-off value from a population of data samples and a set of variables Outlier value candidate list obtained by ranking the principal components obtained by principal component analysis in correspondence with the outlier candidates, and the principal values corresponding to the outlier candidates and the values and factors for each principal component score. A result display unit that generates and displays a variable influence graph representing the load amount.

It is a block diagram which shows schematic structure of the data monitoring apparatus which concerns on one Embodiment of this invention. It is a figure for demonstrating a deviation degree. It is a figure which shows an example of calculation of a deviation degree. It is a figure which shows an example of an outlier candidate list. It is a figure which shows an example of a variable influence graph. It is a flowchart which shows the flow of the data monitoring process in the data monitoring apparatus which concerns on 1st Embodiment. It is a figure explaining the weight of a sample. It is a figure explaining the weight of the variable. It is a block diagram which shows schematic structure of the data monitoring apparatus which concerns on 2nd Embodiment. It is a figure which shows the example of a list of the weighted multivariate data memorize | stored in a memory | storage part. It is a flowchart which shows the flow of the monitoring process of the data in the data monitoring apparatus which concerns on 2nd Embodiment. It is a figure which shows an example of a weight setting. It is a figure which shows the example of calculation of the deviance after weight setting. It is a figure which shows an example of a weight change. It is a figure which shows the example of calculation of the deviation degree after a weight change.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the drawings, the same portions are denoted by the same reference numerals, and redundant description is omitted.

First, main terms used in the present embodiment will be described.
“Variables” are items that can take different values.
“Sample” refers to a set of variables (x11, x12,..., X1m).
“Multivariate data” refers to data composed of a plurality of variables.
An “outlier” is a value that deviates significantly from the data group.

The “principal component analysis” is a method of simplifying the multivariate data in such a way that the coordinate transformation is performed in consideration of the correlation between the variables and the amount of information is not reduced as much as possible. As a principal component, a new composite variable Z1 is defined so that the variance is maximized from the input data (x1, x2, ..., xm).

  A composite variable is called a principal component, of which Z1 having the largest variance is the first principal component (PC1), Z2 having the next largest variance is the second principal component (PC2), and the m-th principal in order. It is called component (PCm).

  Each principal component is determined to be orthogonal to each other so as not to correlate with each other. Since there is no correlation between each synthetic variable, it means that “individual synthetic variables may be evaluated independently”.

The “principal component score” is a value obtained by projecting original data onto a principal component vector.
“Principal component contribution ratio” represents how much each variable is explained by m principal components, where 0 ≦ contribution ratio ≦ 1 and (principal component contribution ratio) = (dispersion value of principal component) / It is expressed as (sum of variance of each variable).
“Factor loading (principal component loading)” refers to the correlation coefficient between the principal component and the original variable.

(First embodiment)
In this embodiment, the maximum absolute value of each normalized principal component score is defined as “outlier”, and based on “outlier”, detection of “outlier” candidates that are greatly out of the data group is detected. Is what you do.

  FIG. 1 is a block diagram showing a schematic configuration of a data monitoring apparatus according to the first embodiment of the present invention. This apparatus is realized using a general-purpose computer (for example, a personal computer (PC) or the like) and software operating on the computer. The computer includes an engineering workstation (EWS) suitable for CAD (Computer Aided Design) and CAE (Computer Aided Engineering). The present embodiment can also be implemented as a program that causes such a computer to execute a series of procedures relating to monitoring of multivariate data.

  As shown in FIG. 1, the data monitoring apparatus according to the first embodiment mainly includes an input unit 11, an outlier candidate detection unit 12, and a result display unit 13.

  The input unit 11 inputs multivariate data. The input multivariate data is sent to the outlier candidate detection unit 12.

  The outlier candidate detection unit 12 performs principal component analysis on the multivariate data received from the input unit 11 to detect outlier candidates. In order to detect outlier candidates, the present embodiment calculates outliers. The detection of outlier candidates will be described later. The calculated degree of detachment is sent to the result display unit 13.

  The result display unit 13 generates and displays an outlier candidate list and a variable influence graph for each outlier candidate based on the outlier received from the outlier candidate detection unit 12. The outlier candidate list and the variable influence graph will be described later.

  Next, multivariate data monitoring processing by the data monitoring apparatus 100 configured as described above will be described.

<Detection of outlier candidates>
Perform principal component analysis on multivariate data. For example, a technique using a variance-covariance matrix is suitable. Principal component scores obtained by principal component analysis are normalized (feature axis normalization) to make it easier to compare multidimensional variable data that is far from the center of gravity for a certain characteristic (principal component) and cannot be simply compared. ). For example, data that is coordinate-transformed by principal component analysis by linear transformation so that the mean is 0 and the variance is 1, and is compressed so that the variance is the same length, and is far away from the center of gravity It becomes easy to grasp accurately.

For each sample (a set of variables), the outlier is calculated. In the present embodiment, the maximum absolute value of each normalized principal component score is defined as a deviation degree Dj. The deviation degree Dj is expressed by the following equation.

  FIG. 2 is a diagram for explaining the degree of detachment. As shown in FIG. 2, the degree of deviation is a value projected onto the principal component 2 (the principal component score of the principal component 2).

  FIG. 3 is a diagram illustrating an example of calculation of the degree of deviation. As shown in FIG. 3, samples A to L are arranged in the vertical direction, and normalized principal component scores Sn_1 to Sn_5 are arranged in the horizontal direction. In the example shown in FIG. 3, the first principal component score Sn_1 of −1.553798 has the largest absolute value of the principal component score normalized by the sample A. Therefore, an absolute value of 1.53798 is set as the degree of detachment of sample A. Similarly, the shaded principal component score is the degree of detachment of each sample. The rightmost column D in FIG. 3 represents the degree of detachment of each sample. The bottom column of FIG. 3 represents the cumulative contribution rate. Cumulative contribution is the cumulative contribution that represents how much each variable is explained by m principal components. In the example shown in FIG. 3, it can be said that about 98% of the entire data can be explained by the first to third principal components PC1 to PC3.

<Outside value candidate list>
The outlier candidate list is a ranking of the samples and principal components corresponding to the outlier candidates based on the degree of outliers. FIG. 4 is a diagram illustrating an example of an outlier candidate list. In the example shown in FIG. 4, those having an outlier degree of 2.0 or more are outlier candidates, and the fourth principal component PC4 of the sample B has an outlier degree of 2.682162, which is the highest rank as an outlier candidate.

<Variable influence graph>
The variable influence graph visualizes the value of each principal component score and the factor loading for the principal component corresponding to the outlier candidate. The variable influence graph has information on the variables that are the main factors of outliers and information on the correlation between the variables. FIG. 5 is a diagram illustrating an example of a variable influence graph. The vertical axis represents the factor loading, and the horizontal axis represents each variable. The solid line connects the factor loadings of each variable. The broken line connects the values for each variable. In the example shown in FIG. 5, it can be seen that the sample B has a high contribution to the fourth principal component PC4, and the influence of the variable W is particularly large.

  FIG. 6 is a flowchart illustrating a flow of data monitoring processing in the data monitoring apparatus according to the first embodiment.

  First, principal component analysis is performed on the input multivariate data using, for example, a variance-covariance matrix (step S61).

  Next, normalization of the principal component score obtained by principal component analysis is performed (step S62).

  Next, the detachment degree Dj is calculated for each sample as shown in FIG. 3 (step S63). The deviation Dj is calculated for the number of samples.

  Next, an outlier candidate list as shown in FIG. 4 and a variable influence degree graph as shown in FIG. 5 are generated and displayed based on the calculated outliers (step S64).

  According to the first embodiment, in multivariate data having a plurality of related variables, it is possible to specify an abnormal value or an outlier while considering the relationship between the variables. In addition, with regard to the variable that is the main factor of the specified abnormal value or outlier, effective information for cause analysis and corrective action showing the abnormal value or outlier can be obtained.

(Second Embodiment)
Next, a second embodiment will be described. In the second embodiment, in data analysis / monitoring, it is possible to set “weight” indicating relative importance and change “weight” for a variable or sample of interest.

  FIG. 7 is a diagram for explaining sample weights. As shown in FIG. 7, the larger the sample value, the greater the influence of that sample in the group. A weight of 0 means that the sample is ignored. By weighting the sample of interest, the entire center of gravity moves, and an outlier from the sample of interest can be detected. For example, when the project data is a sample and its importance (rank: A, B, C,...) Is defined, it is preferable to set a weight according to the importance.

  FIG. 8 is a diagram for explaining variable weights. As shown in FIG. 8, the larger the value of the variable, the greater the influence of the difference in the variable. A weight of 0 means that the variable is ignored. By weighting the variable of interest, the variation of the sample can be expanded or reduced in the direction of the variable, and outliers can be detected in consideration of the influence of the variable. For example, it is preferable to perform data monitoring focusing on important variables (source code scale, defect density, etc.) while ignoring unnecessary variables (such as those that can be generally covered by other variables).

  FIG. 9 is a block diagram illustrating a schematic configuration of a data monitoring apparatus according to the second embodiment. As shown in FIG. 9, the data monitoring apparatus 200 according to the second embodiment includes a weight setting unit 14 and a storage unit 15 in addition to the configuration of the first embodiment.

  The weight setting unit 14 sets weights for samples and variables. In the second embodiment, data correction is performed by adjusting the weights of samples and variables. The adjustment of the weight can be performed by the following method by normalizing the multivariate data based on the set weight.

(1) Exclude samples with zero weight and variables.
(2) For the variable j, find the weighted average Ave_j and the weighted standard deviation Dev_j.
If the weight of sample i is Ws_i and the value of sample i for variable j is V_i_j,
Ave_j = Σ (Ws_i * V_i_j) / Σ (Ws_i), Dev_j = √ Σ (Ws_i * (V_i_j−Ave_j) ^ 2) / (Σ (Ws_i) -1)
(3) For each variable, normalize with mean 0 and standard deviation Wv_j (weight of variable j).
V_i_j ← (V_i_j−Ave_j) / Dev_j * Wv_j
The storage unit 15 is a storage device, and stores weighted multivariate data. For example, a hard disk device or a solid state drive (SSD) device using a flash memory which is a semiconductor memory is suitable. FIG. 10 is a diagram illustrating a list example of weighted multivariate data stored in the storage unit 15. As shown in FIG. 10, multivariate data with the number of samples: n and the number of variables: m is stored in the storage unit 15 as an n × m matrix. First, weight = 1 is set as an initial value for samples and variables.

  FIG. 11 is a flowchart showing the flow of source code processing in the source code analyzing apparatus according to the second embodiment.

  First, weighting is performed (step S111). FIG. 12 is a diagram illustrating an example of weight setting. In the example shown in FIG. 12, the weights are changed from the initial values of 1 to 0 for sample B and sample H, which are large in the lower principal components (PC4, PC5) in the example of calculation of deviation shown in FIG. doing.

  Next, principal component analysis is performed on the input multivariate data using, for example, a variance-covariance matrix (step S112).

  Next, normalization of the principal component score obtained by principal component analysis is performed (step S113).

  Next, the detachment degree Dj is calculated for each sample (step S114). The deviation Dj is calculated for the number of samples.

  Next, outlier candidates for each sample are ranked based on the calculated outlier Dj (step S115). FIG. 13 is a diagram illustrating a calculation example of the degree of detachment after the weight change. Reflecting the weight setting for sample B and sample H performed in FIG. 12, all principal component scores of sample B and sample H are zero. It can be seen that the cumulative contribution ratio of the first principal component to the third principal component (PC1 to PC3) increases after the weight change. Further, it can be seen that the samples C, E, and K having large outliers from the first main component to the third main component (PC1 to PC3) maintain the tendency with the ranks of outlier candidates remaining high.

  FIG. 14 is a diagram illustrating an example of the weight change. In the example illustrated in FIG. 14, the weight is changed from 1 to 5 for the sample E in which the outlier candidate rank is first and the sample C in which the outlier candidate rank is second. FIG. 15 is a diagram illustrating a calculation example of the degree of detachment after the weight change. As shown in FIG. 15, the principal components corresponding to the outlier candidates are almost the same as before the change of the sample weight, and the structure is saved. Sample C and sample E whose weights are changed have a small degree of detachment and are the center of the structure. As a result, the outlier candidate rank has also decreased. Reflecting these changes, sample F has the largest outlier, the outlier candidate rank is first, and is an outlier from the sample of interest (C, E).

  Next, based on the calculated outliers, an outlier candidate list as shown in FIG. 4 and a variable influence graph as shown in FIG. 5 are generated and displayed (step S116).

  According to the second embodiment, by changing the sample of interest and the weight of the variable, in data analysis / monitoring, “this sample is obviously an abnormal value and should be excluded”, “This variable is important. It is possible to reflect the experience and intention of the user such as “it is necessary to monitor with priority”.

  Although several embodiments of the present invention have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the scope of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are included in the invention described in the claims and the equivalents thereof.

100, 200 ... data monitoring device 11 ... input unit 12 ... outlier candidate detection unit 13 ... result display unit 14 ... weight setting unit 15 ... storage unit

Claims (10)

An outlier candidate detection unit that performs principal component analysis on multivariate data and detects outlier candidates that are significantly out of the data population;
Outlier candidate list in which the principal component obtained by the sample and principal component analysis that is a set of the variables and the principal component analysis is ranked in correspondence with the outlier candidate, and the principal component corresponding to the outlier candidate, the principal component A result display unit that generates and displays a variable impact graph indicating a factor load value and a value for each variable of the score,
A data monitoring device provided.   The outlier candidate detection unit calculates an outlier expressed by a maximum absolute value of each principal component score obtained by normalizing a principal component score obtained by principal component analysis in order to detect the outlier candidate. The data monitoring apparatus according to 1.   The data monitoring apparatus according to claim 1, further comprising an input unit that inputs the multivariate data, wherein the input multivariate data is sent to the outlier candidate detection unit. The data monitoring apparatus according to claim 2 , wherein the deviation degree is calculated for each sample. The outlier candidate, data monitoring apparatus according to claim 4, outlier candidate the sample or the main component of the out-degree less than a predetermined value.   The data monitoring apparatus according to claim 1, wherein the factor loading is a correlation coefficient between a principal component and an original variable. further,
A weight setting unit for setting a weight for the sample and the variable;
A storage unit for storing the weighted multivariate data,
The data monitoring apparatus according to claim 1, further comprising: adjusting a weight of the sample or variable. Samples with large deviations expressed by the maximum absolute value of each principal component score obtained by normalizing principal component scores obtained by principal component analysis by setting initial values for the samples and variables The data monitoring apparatus according to claim 7, wherein the weight is changed to a value smaller than an initial value.   The data monitoring apparatus according to claim 7 or 8, wherein an initial value is set for the sample and the variable, and the weight is changed to a value larger than the initial value for a sample having a higher outlier candidate rank. The data according to any one of claims 1 to 9, wherein the outlier candidate list also reflects a cumulative contribution rate obtained by accumulating a contribution rate indicating how much each of the variables is explained by each principal component. Monitoring device.

Images