Feature extraction device and program
The feature extraction device and program address the susceptibility of principal component analysis to outliers by using a bootstrap method to calculate eigenvalues and eigenvectors, adjusting data bias, and extracting principal component candidate vectors, thereby enabling robust analysis and derivation of representative components from multiple data sets.
Patent Information
- Application Number
- JP2021046277
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-03-19
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-03-19
AI Technical Summary
Conventional principal component analysis is susceptible to the influence of outliers, especially when dealing with small amounts of data, and lacks effective methods for detecting and mitigating their impact.
A feature extraction device and program that performs principal component analysis using a bootstrap method to calculate eigenvalues and eigenvectors, adjusts data bias, generates vector sets, and extracts principal component candidate vectors robustly, reducing the influence of outliers through repeated calculations and inner product analysis.
Enables robust principal component analysis that is less susceptible to outliers, allowing for the derivation of representative components from multiple data sets, even with small data amounts, by smoothing the influence of outliers and focusing on vectors with large inner products.
Smart Images

Figure 0007577002000003 
Figure 0007577002000004 
Figure 0007577002000005
Abstract
Description
[Technical field]
[0001] The present invention relates to a feature extraction device and a program for deriving representative components from a plurality of data while mitigating the effects of outliers in principal component analysis, which is performed in the field of statistical analysis. [Background technology]
[0002] Principal component analysis, a method of multivariate analysis, is sometimes used when extracting correlations and trends between numerical values from a large amount of numerical data (for example, see Non-Patent Document 1). Generally, when performing principal component analysis on a large amount of data, methods that are known include a method in which a single analysis is performed on one piece of data, the result is obtained, and then the analysis is terminated, and a bootstrap method in which a process of partial extraction in a bootstrap manner and principal component analysis is repeated. [Prior art documents] [Non-patent literature]
[0003] [Non-Patent Document 1] "Introduction to Multivariate Analysis - From Linear to Nonlinear" by Sadanori Konishi, Iwanami Shoten, January 2010 Summary of the Invention [Problem to be solved by the invention]
[0004] However, because the conventional principal component analysis is a method that focuses on the variance of data, it generally has the problem of being easily affected by outliers. Although there are many methods for detecting and removing outliers, there are many issues, such as the large amount of data, the setting of confidence intervals and thresholds, etc. In addition, there is also the problem that the influence of outliers is stronger, especially when the amount of data is small.
[0005] Therefore, the present invention has been made to solve the above-mentioned problems, and its object is to provide a feature extraction device and program that can perform principal component analysis robust to outliers, even for a small amount of data, regardless of the amount of data, in principal component analysis performed in the field of statistical analysis, that is, that can derive representative components from multiple data while mitigating the effects of outliers. [Means for solving the problem]
[0006] In order to achieve the above object, a feature extraction device of claim 1 is a feature extraction device for performing principal component analysis used in the field of statistical analysis, comprising: a data acquisition unit for acquiring a plurality of data serving as a basis for performing the principal component analysis; a bias adjustment unit for bias-adjusting the data acquired by the data acquisition unit to set all average values to 0 (zero); an eigenvalue / eigenvector calculation unit for repeatedly performing in a bootstrap manner a process of calculating and storing eigenvalues and eigenvectors from the data bias-adjusted by the bias adjustment unit; a vector set creation unit for generating a vector set of vectors obtained by multiplying the square root of each eigenvalue by a corresponding eigenvector, based on the eigenvalues and eigenvectors calculated and stored by the eigenvalue / eigenvector calculation unit; a principal component candidate vector extraction unit for extracting and storing a representative row vector from among row vectors included in the vector set generated by the vector set creation unit, as a principal component candidate vector; and a final vector output unit for outputting the principal component candidate vectors extracted and stored by the principal component candidate vector extraction unit as principal component data.
[0007] In addition, the feature extraction device of claim 2 is characterized in that the principal component candidate vector extraction unit extracts, from among the row vectors included in the vector set generated by the vector set creation unit, a vector having a large sum of absolute values of inner products between the vector itself and other vectors as the principal component candidate vector.
[0008] In addition, the feature extraction device of claim 3 is characterized in that, when the number of times (the number of loops) that the eigenvalue / eigenvector calculation unit repeats the process of calculating and storing the eigenvalues and eigenvectors in a bootstrap manner is K, the principal component candidate vector extraction unit repeats the process of extracting the principal component candidate vectors by extracting up to K vectors having a large inner product with the principal component candidate vector and setting all of them to zero vectors to update the vector set, until a predetermined number of the principal component candidate vectors are extracted.
[0009] The program of claim 4 is a feature extraction program for performing principal component analysis used in the field of statistical analysis, characterized in that it causes a computer to execute the following steps: a data acquisition unit acquires a plurality of data to be used as a basis for the principal component analysis; a bias adjustment unit bias-adjusts the data acquired by the data acquisition unit to set all average values to 0 (zero); an eigenvalue / eigenvector calculation unit calculates and stores eigenvalues and eigenvectors from the data bias-adjusted by the bias adjustment unit, repeating the above steps in a bootstrap manner; a vector set creation unit generates a vector set of vectors obtained by multiplying the square root of each eigenvalue by the corresponding eigenvector, based on the eigenvalues and eigenvectors calculated and stored by the eigenvalue / eigenvector calculation unit; a principal component candidate vector extraction unit extracts and stores a representative row vector from among the row vectors included in the vector set generated by the vector set creation unit, as a principal component candidate vector; and a final vector output unit outputs the principal component candidate vector extracted by the principal component candidate vector extraction unit as principal component data. Effect of the Invention
[0010] According to the present invention, in principal component analysis performed in the field of statistical analysis, it is possible to perform principal component analysis that is robust against outliers, regardless of the amount of data, even for a small amount of data, that is, it is possible to derive representative components from multiple data while mitigating the effects of outliers. [Brief description of the drawings]
[0011] [Figure 1] 1 is a block diagram showing an example of a functional configuration of a feature extraction device according to an embodiment of the present invention. [Diagram 2] 4 is a flowchart showing a process of the feature extraction device according to the embodiment of the present invention. [Diagram 3] FIG. 2 is a schematic diagram showing how principal component candidate vectors extracted by the processing of the feature extraction device according to the embodiment of the present invention are handled. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0012] The present invention relates to a feature extraction device and a program for deriving representative components from multiple data while mitigating the effects of outliers through multiple analyses and clustering in principal component analysis performed in the field of statistical analysis. Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.
[0013] Generally, principal component analysis, a method of multivariate analysis, is often used to extract correlations and trends between numbers from a large amount of numerical data. In what follows, the size of the data handled is taken as N samples, and the number of features (entities) as M. If principal component analysis is interpreted as a geometric problem in hyper-dimensional space, it can be rephrased as a problem of finding the major axis of the distribution function of data that is elliptically distributed in M-dimensional space. Mathematically, the problem is to create an M x M variance-covariance matrix based on the data, and find the eigenvalues and eigenvectors. In this case, the direction of the major axis of the distribution function corresponds to the eigenvector, and the square of the length of the major axis corresponds to the eigenvalue.
[0014] In this embodiment, the results of an experiment in which the size of the data handled was 3000 samples (N=3000) and the number of features (entities) was 88 (M=88) will be described. Note that N=3000 and M=88 are the values used when the experiment was conducted, and the present invention is not limited to these values.
[0015] Fig. 1 is a block diagram showing an example of a functional configuration of a feature extraction device 10 according to an embodiment of the present invention. As shown in Fig. 1, the feature extraction device 10 includes a data acquisition unit 11, a bias adjustment unit 12, an eigenvalue / eigenvector calculation unit 13, a vector set creation unit 14, a principal component candidate vector extraction unit 15, and a final vector output unit 16, and is a device that performs principal component analysis used in the field of statistical analysis.
[0016] The data acquisition unit 11 is an acquisition unit that acquires a plurality of data that are the source of the principal component analysis. If the source data is stored in an external device, the data may be acquired from the external device (received via a network, etc.), or if the source data is stored in advance in a data storage unit (not shown) of the feature extraction device 10, the data may be acquired from the data storage unit.
[0017] The bias adjustment unit 12 is an adjustment unit that bias-adjusts the data acquired by the data acquisition unit 11 to set all average values to 0 (zero). The eigenvalue / eigenvector calculation unit 13 is a calculation unit that repeatedly performs the process of calculating eigenvalues and eigenvectors from the data bias-adjusted by the bias adjustment unit 12 and storing them in a memory in a bootstrap manner.
[0018] The vector set creation unit 14 is a creation unit that reads out the eigenvalues and eigenvectors calculated and stored by the eigenvalue / eigenvector calculation unit 13 from a memory, and generates a vector set C of vectors obtained by multiplying the corresponding eigenvector by the square root of each eigenvalue, based on the eigenvalues and eigenvectors.
[0019] The principal component candidate vector extraction unit 15 extracts a representative row vector from among the row vectors included in the vector set C generated by the vector set creation unit 14, and designates the representative row vector as a principal component candidate vector V. t and stores the extracted information in a memory.
[0020] The final vector output unit 16 outputs the principal component candidate vectors (representative row vectors) Vt An output unit that reads from memory and outputs this as principal component data.
[0021] In this embodiment, the eigenvalue / eigenvector calculation unit 13 and the principal component candidate vector extraction unit 15 are each described as having a storage area (memory) inside. However, the data to be stored may be stored in a data storage unit (not shown), which is another storage area within the feature extraction device 10.
[0022] Next, the processing in this feature extraction device 10 will be described in detail with reference to FIG. 2. FIG. 2 is a flowchart showing the processing of the feature extraction device 10 in the embodiment of the present invention.
[0023] First, the data acquisition unit 11 acquires data of N samples (N = 3000) and M entities (M = 88) (step ST1). Next, the bias adjustment unit 12 adjusts the bias of the data of N samples (N = 3000) and M entities (M = 88) acquired by the data acquisition unit 11 in step ST1 so that all the average values are 0 (zero) (step ST2).
[0024] Then, after setting k = 1 (step ST3), the eigenvalue / eigenvector calculation unit 13 randomly extracts N' samples (N' < N; here, N' = 2400) from among the N samples to generate a data matrix D k (step ST4). Further, from the data included in this D k , an M×M covariance matrix Σ k is generated (step ST5), and the eigenvalues λ k , λ k1 , ···, λ k2 , ···, λ kt and the eigenvectors v k1 , v k2 , ···, v kt(t≦M) is calculated and stored (step ST6). At this time, all sets of eigenvectors and corresponding eigenvalues may be stored, or a threshold or upper limit may be set for the eigenvalues so that only sets of eigenvalues and eigenvectors that are equal to or greater than a certain value or a certain number of eigenvalues may be stored.
[0025] Here, if the number of loops (number of times) for repeating the above steps ST4 to ST6 is K (here, K=50), that is, if the number of times (number of loops) for the eigenvalue / eigenvector calculation unit 13 to repeat the process of calculating and storing eigenvalues and eigenvectors in a bootstrap manner is K (=50), it is determined whether or not the condition "k=K?" is satisfied (step ST7), and if k=K is not satisfied (NO in step ST7), k=k+1 is set, that is, k is incremented by one until the number of loops k reaches K (=50) (step ST8), and the processes of steps ST4 to ST6 are repeated.
[0026] In this way, by repeating the processes of steps ST4 to ST6, that is, by repeatedly performing the process of calculating and storing eigenvalues and eigenvectors in a bootstrap manner, the influence of outliers can be reduced. Outliers can be regarded as a kind of discrete data, and if the influence of outliers remains, the vector obtained by principal component analysis also behaves in a discrete manner, but by repeatedly performing the process in a bootstrap manner, the influence of outliers is smoothed, and the possibility of deriving appropriate principal components is increased.
[0027] On the other hand, if k=K in step ST7 (YES in step ST7), that is, if the processes in steps ST4 to ST6 have been repeated up to the K (=50)th loop, the process is handed over to the vector set creation unit .
[0028] At this time, if the eigenvalue and eigenvector calculated in step ST6 of each loop is stored in a memory area held by the eigenvalue / eigenvector calculation unit 13, the eigenvalue and eigenvector stored therein may be output (transferred) to the vector set creation unit 14. If the feature extraction device 10 is provided with a separate data storage unit (not shown) for storing the eigenvalues and eigenvectors, the vector set creation unit 14, upon receiving an instruction (notification) from the eigenvalue / eigenvector calculation unit 13, may acquire the eigenvalues and eigenvectors calculated in step ST6 from the data storage unit.
[0029] Then, the vector set creation unit 14 multiplies the corresponding eigenvector by the square root of each eigenvalue calculated and stored by the eigenvalue / eigenvector calculation unit 13 in step ST6 (step ST9), thereby obtaining the vector V 1 ,V 2 ,···,V T A vector set C (T=M·K) is generated (step ST10). Here, the vector V i (1≦i≦T) are each 1×M matrices (row vectors with M elements).
[0030] To explain this process in more detail, if the process of steps ST4 to ST6 described above is performed K times and a maximum of M eigenvalues are obtained in each loop, the total number of sets of eigenvalues and eigenvectors is M x K (= T), so T vector sets are generated by the process of step ST9, which multiplies the square root of the eigenvalue by the eigenvector. 1 ,V 2 ,···,V T There are V i (1≦i≦T) is a row vector with M elements. More specifically, V 1 =√(λ 11 ·v 11 ) (i.e., V i =√(λ i1 ·v i1)) However, this is the case when a maximum of M eigenvalues are obtained in each loop, but it is not necessarily the case that M eigenvalues are obtained, and it is possible that fewer than M eigenvalues are obtained.
[0031] When the vector set C is generated by the vector set creation unit 14, the principal component candidate vector extraction unit 15 receives it and extracts vectors V 1 ,V 2 ,···,V T A T×M matrix S is generated with each row having elements in ascending order (step ST11). i Since (1≦i≦T) is a row vector with M elements (1×M matrix), stacking it in ascending order by T rows will generate a T×M matrix S. The T×M matrix S is expressed by the following formula.
[0032] The T×M matrix S is expressed by the following formula.
number
[0033] Furthermore, the principal component candidate vector extraction unit 15 extracts a T×T symmetric matrix R=SS T (Step ST12). Since the matrix S is a T×M matrix, S T is an M×T matrix, and R=SS T is a T×T symmetric matrix.
[0034] The T×T matrix R is expressed by the following formula.
number
[0035] Next, the principal component candidate vector extraction unit 15 calculates the sum of the absolute values for each row of the symmetric matrix R, obtains the row number t (1≦t≦T) with the largest sum (step ST13), and extracts the vector V t are stored as candidate principal component vectors (step ST14). That is, among the T vectors, only one is selected as a representative vector, and the inner product of the vectors with the other vectors is taken. The vector with the largest size and the most representative vector with a similar direction, that is, the strongest component, is picked out and stored.
[0036] That is, the principal component candidate vector extraction unit 15 extracts the row vector V i Among (1≦i≦T), the vector with the largest sum of absolute values of the dot products of its own vector and other vectors is called a representative vector V t (Principal component candidate vector V t ) and store it.
[0037] These processes will be described in more detail with reference to the schematic diagram shown in Fig. 3. Fig. 3 is a schematic diagram showing how principal component candidate vectors extracted by the process of feature extraction device 10 in the embodiment of the present invention are handled. As shown in the box on the right side of Fig. 3, the arrows in Fig. 3 indicate the principal component candidate vectors of the kth loop, and those in the same loop are shown with the same pattern. As described above, the principal component candidate vectors are obtained by multiplying the square root of the eigenvalue by the eigenvector, and the eigenvalue corresponds to the variance, so multiplying the vector by the square root of the eigenvalue (think of it as a standard deviation or radius) creates a difference in length.
[0038] For example, the arrows 21 and 22 in the schematic diagram shown on the left side of Fig. 3 are vectors calculated in the first loop of steps ST4 to ST6 in Fig. 2, the arrows 31 and 32 are vectors calculated in the second loop, the arrows 41 and 42 are vectors calculated in the third loop, and the arrows 51 and 52 are vectors calculated in the fourth loop. In this case, there are four components that belong to group A surrounded by a solid line, and four components that belong to group B surrounded by a dashed line. Note that, for the sake of explanation, only four loops are shown here, but if steps ST4 to ST6 are repeated ten times, the result may be that, for particularly strong components, there are ten components that belong to group A and ten components that belong to group B.
[0039] Also, since vectors may be output in the opposite direction by 180 degrees, Fig. 3 is a diagram that takes such a case into consideration, but since eigenvalues are positive values, vectors can be considered as positive regardless of their direction. Therefore, in the above-mentioned steps ST11 to ST14, the directions of the principal component candidate vectors are not considered at all, and the inner product is taken all-to-all. Vectors that belong to the same group are almost parallel, so the absolute value of the inner product is large. On the other hand, vectors that belong to different groups in different loops are not necessarily orthogonal, so the inner product will not be zero even if it is taken, but it is likely to produce an absolutely small value.
[0040] For example, when the dot product of vector 21 and vector 32 shown in FIG. 3 is taken, the value always becomes small. Also, since a representative vector has a large eigenvalue and is long, that is, the standard deviation equivalent to the radius is long, the value tends to become large when the dot product is taken with other vectors. And since vectors with large eigenvalues and strong components tend to gather in a narrow range and have clear characteristics, if other vectors are also long and point in the same direction, focusing on the vector with the strong component and taking the sum of the dot products with other vectors will result in a large value. In this way, by taking the dot product of one vector with other vectors and then taking the sum, it can be confirmed that the vector can represent all pairs in the group.
[0041] That is, usually, the principal component vectors themselves should exist in a probability distribution according to the error of the data, but among them, the vector with the largest eigenvalue and the largest sum of the inner products with other vectors is considered to exist at the center of the distribution. Therefore, in steps ST11 to ST14 in FIG. 2, the vector with the largest eigenvalue, which has the largest sum of absolute values of the inner products in the matrix R obtained by taking the all-to-all inner products, is selected as the representative vector V t It is extracted as.
[0042] In addition, with regard to outliers in the data, some may not be picked up, there may be two or more instead of one, and the outliers may cancel each other out. Taking these factors into consideration, the distribution will inevitably be concentrated in patterns that do not have outliers. Therefore, by processing steps ST11 to ST14 described above, it has the effect of being less susceptible to the influence of outliers and enabling analysis that is robust against outliers.
[0043] Then, the principal component candidate vector V extracted and stored in step ST14 is t If the number of candidates for principal component V is equal to or less than a predetermined threshold (here, the predetermined threshold is 3) (if NO in step ST15), the candidate for principal component V is extracted from the vector set C as a preprocessing step for repeating the process from step ST11. t A maximum of K vectors having a large inner product with are detected (step ST16), all of the detected K vectors are set as zero vectors, and the vector set C is updated (step ST17), and the processes of steps ST11 to ST14 are repeated.
[0044] More specifically, in step ST16, a vector R obtained by extracting the t-th row of the symmetric matrix R is t =[r t1 ,r t2 ,···,r tT ], extract the K elements (K elements) with the largest absolute values, and assign their column indexes to u 1 ,u 2 ,···,u k (k≦K). In this case, u1 Then, in step ST17, among the vectors in the vector set C, V u1 ,V u2 ,···,V uK In this embodiment, the maximum number of vectors to be extracted is K, which is the same as the number of loops. However, the maximum number of vectors to be extracted may be about 80% of the K number.
[0045] That is, the principal component candidate vector extraction unit 15 extracts a predetermined number (a predetermined threshold value) of principal component candidate vectors V t Until the principal component candidate vector V t Then, up to K vectors that have a large inner product with are extracted, and vector set C is updated by making them all zero vectors. Then, the process of extracting the next representative vector, which is a principal component candidate vector, is repeated.
[0046] To explain this with reference to FIG. 3, the principal component candidate vector V extracted and stored in step ST14 is t For example, if the vectors are the striped vectors 21 and 22 in the schematic diagram shown in FIG. 3, in step ST16, the vectors 31 and 32, the vectors 41 and 42, and the vectors 51 and 52 in the group A surrounded by a solid line to which the vector 21 belongs, and the group B surrounded by a dashed line to which the vector 22 belongs, are determined to be principal component candidate vectors V t In step ST17, these vectors are all set as zero vectors and removed from vector set C, and the processes in the next steps ST11 to ST14 are then carried out.
[0047] Then, the processes in steps ST11 to ST14 are repeated to obtain the principal component candidate vector V t When the number of reaches a predetermined threshold value (=3), the principal component candidate vector V stored in step ST14 is t(Here, there are three) are output as the result of principal component analysis (step ST18). Note that the predetermined threshold value can be appropriately determined depending on how many representative components are desired to be extracted as principal components, but it is generally thought that three or four are often used.
[0048] In this way, three types (as many as the predetermined threshold) of representative vectors with the number of elements M (=88) are extracted. Also, in this embodiment, an explanation is given based on an experiment using data in which N' samples (N'=2400) are randomly extracted from data with N samples (N=3000) and M entities (M=88), but according to this method, regardless of the amount of data, the distribution is concentrated in patterns without outliers, so that even when there is a small amount of data, the processing of the above-mentioned steps ST11 to ST14 has the effect of making it difficult to be influenced by outliers and enabling robust analysis against outliers.
[0049] As described above, according to the feature extraction device 10 of the embodiment of the present invention, in principal component analysis performed in the field of statistical analysis, it is possible to perform principal component analysis that is robust against outliers, even for a small amount of data, regardless of the amount of data, that is, it is possible to derive representative components from a plurality of data while mitigating the effects of outliers.
[0050] (Application example 1) The feature extraction device 10 according to the embodiment of the present invention can be applied to a sensor with a mixture of multiple signals. That is, it can be applied to data derived from a sensor that is easily exposed to disturbances and outliers. For example, in the case of a keystroke inspection or an abnormal sound inspection for a structure, the location and type of an abnormality are estimated based on the data of the vibration returned from the inspection vibration. However, when multiple abnormalities exist, the multiple vibrations are mixed together, which may make detection difficult. Furthermore, if data belonging to an outlier is added, detection becomes even more difficult. However, even in such a case, by performing a principal component analysis using the feature extraction device 10 according to the present invention, it is possible to perform robust detection of outliers while exhibiting the function of abnormality detection even for a combination of multiple signals.
[0051] (Application example 2) The feature extraction device 10 according to the embodiment of the present invention can be applied to human attribute analysis. For example, data that depends on human responses, such as personality tests and questionnaires, is known to be noisy. However, even in such cases, the feature extraction device 10 of the present invention can classify many respondents into multiple attributes while reducing the influence of respondents who give abnormal responses by performing principal component analysis on the answer options to questions as numerical data (for example, by converting the options into 1 for YES and 0 for NO).
[0052] A typical computer can be used as the hardware configuration of the feature extraction device 10 according to the embodiment of the present invention. The feature extraction device 10 is configured by a computer equipped with a CPU, a volatile storage medium such as a RAM, a non-volatile storage medium such as a ROM, an interface, etc.
[0053] In addition, each function of the data acquisition unit 11, bias adjustment unit 12, eigenvalue / eigenvector calculation unit 13, vector set creation unit 14, principal component candidate vector extraction unit 15, and final vector output unit 16 provided in the feature extraction device 10 is realized by causing a CPU to execute a program describing these functions.
[0054] These programs are stored in the storage medium and are read and executed by the CPU. These programs can also be distributed by storing them in a storage medium such as a magnetic disk (floppy disk, hard disk, etc.), an optical disk (CD-ROM, DVD, etc.), or a semiconductor memory, and can also be transmitted and received via a network.
[0055] It should be noted that, within the scope of the present invention, any of the components of the embodiments may be modified or any of the components of the embodiments may be omitted. [Industrial Applicability]
[0056] The present invention can be widely applied to detecting principal components in the analysis of various noisy data available in the real world. [Explanation of symbols]
[0057] 10 Feature Extraction Device 11 Data Acquisition Section 12 Bias adjustment section 13 Eigenvalue and eigenvector calculation section 14 Vector Set Creation Section 15 Principal component candidate vector extraction section 16 Final vector output section 21,22 Principal component candidate vectors of a certain loop 31,32 Principal component candidate vectors of other loops 41,42 Principal component candidate vectors of other loops 51,52 Principal component candidate vectors of further loops
Claims
1. A feature extraction device for performing principal component analysis used in the field of statistical analysis, comprising: A data acquisition unit that acquires a plurality of data that are used as a basis for performing the principal component analysis; a bias adjustment unit that bias-adjusts the data acquired by the data acquisition unit to set all average values to 0 (zero); an eigenvalue and eigenvector calculation unit that repeatedly performs a process of calculating and storing eigenvalues and eigenvectors from the data bias-adjusted by the bias adjustment unit in a bootstrap manner; a vector set creation unit that generates a vector set of vectors by multiplying the corresponding eigenvector by the square root of each eigenvalue, based on the eigenvalues and eigenvectors calculated and stored by the eigenvalue / eigenvector calculation unit; a principal component candidate vector extraction unit that extracts and stores representative row vectors as principal component candidate vectors from among the row vectors included in the vector set generated by the vector set creation unit; a final vector output unit that outputs the principal component candidate vectors extracted and stored by the principal component candidate vector extraction unit as principal component data; A feature extraction device comprising:
2. The principal component candidate vector extraction unit extracts a vector having a large sum of absolute values of inner products of its own vector and another vector as the principal component candidate vector.
2. The feature extraction device according to claim 1.
3. When the number of times (the number of loops) that the eigenvalue / eigenvector calculation unit repeats the process of calculating and storing the eigenvalues and eigenvectors in a bootstrap manner is K, The principal component candidate vector extraction unit repeats the process of extracting the principal component candidate vectors after extracting up to K vectors that have a large inner product with the principal component candidate vector and setting all of the extracted vectors as zero vectors to update the vector set, until a predetermined number of the principal component candidate vectors are extracted.
3. The feature extraction device according to claim 2.
4. A feature extraction program for performing principal component analysis used in the field of statistical analysis, A data acquisition unit acquires a plurality of data serving as a basis for performing the principal component analysis; A bias adjustment unit bias-adjusts the data acquired by the data acquisition unit to set all average values to 0 (zero); a step of repeatedly performing a process in a bootstrap manner in which an eigenvalue / eigenvector calculation unit calculates and stores eigenvalues and eigenvectors from the data bias-adjusted by the bias adjustment unit; a vector set creation unit generating a vector set of vectors obtained by multiplying the corresponding eigenvector by the square root of each eigenvalue based on the eigenvalues and eigenvectors calculated and stored by the eigenvalue / eigenvector calculation unit; a principal component candidate vector extraction unit extracting and storing representative row vectors as principal component candidate vectors from among the row vectors included in the vector set generated by the vector set creation unit; a final vector output unit outputting the principal component candidate vectors extracted by the principal component candidate vector extraction unit as principal component data; A program characterized by causing a computer to execute the above.
Citation Information
Patent Citations
Device, method, and program for detecting pattern, computer readable recording medium, and apparatus for recording
JP2006202135A
Filter device, filtering method for filter device and filtering program
JP2009110262A
Parameter decision device, parameter decision method and program
JP2017076289A
Method, system, and software for identifying bio-molecule with interacting component
JP2017189176A
