Detector, detection method and program

Abstract

To detect a changing heavenly body with high accuracy.SOLUTION: A detector comprises: a feature amount acquisition unit which acquires a feature amount for each heavenly body data which is time series data showing change of brightness of a heavenly body, by using a feature amount selection method based on sparse modeling; and a detection unit which sets the feature amount for each heavenly body data acquired by the feature amount acquisition unit, and the feature amount about brightness of the heavenly body obtained for each heavenly body data as input, and detects a changing heavenly body by using a classifier learned so as to classify the heavenly body indicated in the heavenly body data into either of the changing heavenly body or a non-changing heavenly body.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a detection device, a detection method, and a program. [Background technology]

[0002] There are two types of variable celestial bodies such as flashing stars, novae, and supernovae. The first type of celestial body is one whose position fluctuates. The second type of celestial body is one whose brightness fluctuates. In celestial bodies whose brightness fluctuates, there are some whose brightness fluctuates over the long term, some whose brightness fluctuates locally, and some whose brightness fluctuates in a mixture of both. Furthermore, the rate of brightness fluctuation is about 1%, and it is periodic. Discovering variable celestial bodies that have extremely rare celestial phenomena is closely related to obtaining highly accurate information and new knowledge in astronomy.

[0003] One of the representative methods for detecting variable objects is the statistical method. Here, we will explain the outline of the statistical method. In the statistical method, the intrinsic variability σ,χ 2 The average, standard deviation, and skewness of the magnitude measurements are calculated, and the records are narrowed down based on the obtained values ​​to obtain candidates for variable objects. This statistical method loses the time information of variable objects because the observation data is aggregated into statistical values. In other words, there is a problem in that it is not possible to utilize the periodic information of the light curve, which is known as a characteristic of variable objects.

[0004] As another example of a method for detecting variable celestial bodies, there is a method that focuses on the periodicity of variable celestial bodies (for example, see Non-Patent Document 1). However, the method shown in Non-Patent Document 1 cannot be executed except at specific frequencies. Therefore, there is a problem that the calculation cost is too high to apply it to celestial body data whose periods differ for each celestial body. Furthermore, when acquiring celestial body data, it is difficult to acquire accurate information due to the influence of weather and seasons, and the data set handled often contains many missing values ​​due to limitations on the observation range due to the diameter of the telescope and limitations on the observation time of the celestial body, and is further affected by noise. Therefore, it is very difficult to acquire frequency information by applying Fourier analysis to celestial body data in order to find the period of the celestial body. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] NR Lomb, “LEAST-SQUARES FREQUENCY ANALYSIS OF UNEQUALLY SPACED DATA”, Astrophysics and Space Science, 1976. Summary of the Invention [Problem to be solved by the invention]

[0006] As described above, the celestial body data used to detect a variable celestial body may contain many missing values ​​and noise. Therefore, it has been difficult to detect a variable celestial body with high accuracy using the celestial body data with conventional techniques.

[0007] In view of the above circumstances, an object of the present invention is to provide a technique capable of detecting a moving celestial body with high accuracy. [Means for solving the problem]

[0008] One aspect of the present invention is a detection device that includes a feature acquisition unit that acquires features for each celestial body data, which is time-series data representing changes in the brightness of celestial bodies, by using a feature selection method based on sparse modeling, and a detection unit that detects variable celestial bodies using a classifier that is trained to classify celestial bodies shown in the celestial body data as either variable celestial bodies or non-variable celestial bodies, using as input the features for each celestial body data acquired by the feature acquisition unit and the features related to the brightness of the celestial body obtained for each celestial body data.

[0009] One aspect of the present invention is a detection method that uses a feature selection technique based on sparse modeling to acquire features for each celestial body data, which is time-series data representing changes in the brightness of celestial bodies, and detects variable celestial bodies using a classifier that is trained to classify celestial bodies shown in the celestial body data into either variable or non-variable celestial bodies using as input the features for each acquired celestial body data and the features related to the brightness of the celestial bodies obtained for each celestial body data.

[0010] One aspect of the present invention is a program for causing a computer to execute a feature acquisition step of acquiring features for each celestial body data, which is time-series data representing changes in the brightness of celestial bodies, by using a feature selection method based on sparse modeling, and a detection step of detecting variable celestial bodies using a classifier trained to classify celestial bodies shown in the celestial body data as either variable celestial bodies or non-variable celestial bodies, using as input the features for each celestial body data acquired in the feature acquisition step and the features related to the brightness of celestial bodies obtained for each celestial body data. Effect of the Invention

[0011] According to the present invention, it is possible to detect a moving celestial body with high accuracy. [Brief description of the drawings]

[0012] [Figure 1] FIG. 2 is a diagram showing an example of celestial body data used in the present invention. [Diagram 2]FIG. 2 is a diagram illustrating an example of the configuration of a detection device according to an embodiment. [Diagram 3] FIG. 1 is a diagram showing algorithm 1 illustrating the procedure of a period analysis process performed by the detection device. [Figure 4] 13 is a flowchart showing the flow of a period analysis process performed by the detection device. [Diagram 5] FIG. 13 is a diagram showing algorithm 2 illustrating the procedure of the non-steady celestial object detection process performed by the detection device. [Figure 6] 10 is a flowchart showing the flow of a non-steady celestial object detection process performed by the detection device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the following description, symbols placed above letters in mathematical expressions or functions (hereinafter referred to as "mathematical expressions, etc.") are written before the letters. For example, the symbol "-" placed above letters in mathematical expressions, etc. will be written as "-y" before the letter "y" below.

[0014] Before going into a detailed explanation of the present invention, we will first explain the problem setting of the present invention. The objective of the present invention is to detect variable celestial objects from a large amount of celestial object data. The celestial object data used in the present invention is time-series data that represents changes in the brightness of celestial objects, as shown in Figure 1. In Figure 1, the horizontal axis represents time, and the vertical axis represents the magnitude of the celestial object. Each point P shown in Figure 1 represents a celestial object.

[0015] The problem addressed by the present invention can be expressed mathematically as shown in equation (1).

[0016]

number

[0017] In equation (1), the set Y represents the set (dataset) of time-series data (celestial data) y that represents the change in the brightness of a celestial object, m represents the number of frequency candidates used in the period analysis of the celestial object, and f nrepresents the Nyquist frequency, and F d represents a domain feature list including variables limited to domain-related features among the features shown in the feature list F, and L represents a list of variable celestial objects. Here, the feature list F is a proposed model in the present invention. In the present invention, the features shown in the feature list F include the domain feature list F d , the regression coefficient w, and the coefficient of determination reg_score. The list of unstable celestial objects L represents a list of celestial object data in which unstable celestial objects are detected.

[0018] [Terminology definition of features used in this invention] Here, the terms used as feature quantities below will be defined. Regression coefficient w: Represents the periodic information of each celestial object data calculated by periodic analysis. The magnitude of each coefficient represents the importance of the corresponding periodic length. The value of the regression coefficient w determines whether or not a celestial object has periodicity. Coefficient of determination reg_score: An index showing how well the regression curve obtained by periodic analysis expresses the objective variable. The accuracy of periodic analysis corresponds to the reliability of the regression coefficient w. Average m_ap40: Represents the feature value equivalent to the brightness measured by measuring the light entering a circle with a diameter of 4 arc seconds on the image for each exposure. The average m_ap40 value is the sum of all the brightnesses in the specified range, so it differs from the brightness of the surrounding target objects. z_apertureflux_40_mag: This represents the brightness measured by measuring the light entering a circle with a diameter of 4 arcseconds in the Coadd image, which is the average of the images for each exposure clipped for each pixel. z_apertureflux_40_mag has some noise removed compared to the average m_ap40, and represents a value closer to the original brightness of the celestial object. diff_ap40: The difference between the average m_ap40 and z_apertureflux_40_mag. In other words, diff_ap40 represents a feature that corresponds to the magnitude of noise that affects celestial objects. Since the period analysis uses fluctuation information on the brightness of celestial objects (average m_ap40), the magnitude of noise that affects the average m_ap40 affects the identification of fluctuating celestial objects. For this reason, this feature is used as the magnitude of noise that affects celestial objects. Label: Represents the identification information of a variable celestial object. Label represents the correct answer data that is associated with each celestial object data during learning. In the following explanation, label "1" represents a variable celestial object, and label "0" represents a non-variable celestial object.

[0019] Among the above features, the mean m_ap40, z_apertureflux_40_mag, diff_ap40, and label are domain knowledge for detecting variable objects. Therefore, these are collectively called the domain feature list F d In this way, the feature list F in the classifier used in the following description is the domain feature list F d In addition, it includes the regression coefficient w and the coefficient of determination reg_score.

[0020] [Technical Problems in the Invention] Next, the technical problems in the present invention will be described. There are mainly two types of technical problems in the present invention. The first problem is that the astronomical data to be analyzed contains many missing values, as shown in FIG. 1. Therefore, in order to directly use Fourier transform on astronomical data, it is necessary to apply missing value imputation. However, since the data content is about 0.01% on average, accurate periodic analysis is difficult if missing value imputation is applied to the remaining approximately 99.99%, and it is difficult to directly apply Fourier transform. To solve this problem, the present invention handles periodic analysis as a regression problem. However, it is necessary to be careful of overlearning at that time. To alleviate this problem, Lasso, a technology that increases the generalizability of the model by feature selection, is used.

[0021] The second issue is the design of features to be used when detecting variable celestial objects. Even if appropriate periodic information of celestial objects is obtained by periodic analysis, highly accurate results cannot be expected unless there is celestial object information that works well for classification. Therefore, it is necessary to design features that take domain knowledge into account. To solve this issue, the present invention uses the feature list F that takes domain knowledge into account, as described above.

[0022] [Explanation of prerequisite knowledge: Fourier transform] Next, we will explain the Fourier transform, Lasso, and Group Lasso, which are prerequisite knowledge for the present invention. The Fourier transform is a method for converting time-series data from the time domain to the frequency domain. In the present invention, the Fourier transform is used to investigate the periodicity of celestial data. By using the Fourier transform, the time-series data y that expresses the change in brightness can be expressed as the sum of a sine wave and a cosine wave. Specifically, when the time-series data y is expressed as the sum of a sine wave and a cosine wave, it can be expressed as in the following equation (2). In equation (2), ν j represents a candidate frequency.

[0023]

number

[0024] [Prerequisite knowledge: Lasso] Next, we will explain Lasso. Lasso is a type of sparse modeling (see Reference 1, for example). Sparse modeling is a method for learning significant properties from sparse data and extracting only information that is determined to be important. (Reference 1: ROBERT TIBSHIRANI, “Regression Shrinkage and Selection via the Lasso”, Journal of the ROYAL STATISTICAL SOCIETY, 1996.)

[0025] The loss function of Lasso is expressed as the following formula (3): In formula (3), n represents the length of the time series, X represents the explanatory variable (frequency), and λ represents the normalization parameter.

[0026]

number

[0027] In Lasso, by utilizing the L1 regularization term, regression coefficients corresponding to unnecessary features are maintained at 0 even after repeated updates. Therefore, feature selection can be performed using Lasso. Astronomical data contains many missing values, which makes it prone to overfitting. However, Lasso can set features that are deemed unnecessary to represent the core of the data to 0, which has the effect of suppressing overfitting. Using Lasso, conversion to the frequency domain can be treated as a regression problem.

[0028] [Prerequisite knowledge: Group Lasso] Next, we explain Group Lasso, which is an improved version of Lasso (see Reference 2, for example). (Reference 2: Ming Yuan and Yi Lin, “Model selection and estimation in regression with grouped variables”, Journal of the Royal Statistical Society, 2006.)

[0029] The loss function of Group Lasso is expressed as the following equation (4). In equation (4), G j represents the number of variables belonging to the jth group, and g represents the group set.

[0030]

number

[0031] Group Lasso has something in common with Lasso in that it can suppress overlearning by performing variable selection. However, unlike Lasso, Group Lasso performs variable selection on a group basis, rather than for each variable individually. When a group of features that are related to each other can be considered in advance, grouping them together can be expected to achieve higher accuracy than Lasso. In particular, in the case of the present invention, there are two features, a sine wave and a cosine wave, for a specific frequency, and grouping the sine wave and the cosine wave using Group Lasso can effectively perform period analysis.

[0032] [Method of the present invention] Next, the method of detecting variable celestial objects in the present invention will be described in detail. In the present invention, the following two models are used. In the following description, these two models will be referred to as the first model and the second model. The first model is a model for performing period analysis of each celestial object data. The second model is a model for detecting variable celestial objects.

[0033] The first model will be described. Prepare explanatory variable X and objective variable Y. The first model applies Group Lasso to each celestial data to calculate the regression coefficient w and the coefficient of determination reg_score.

[0034] The second model will now be described. The second model is based on important features such as the regression coefficient w (e.g., the domain feature list F d ) is used to classify the celestial objects included in the input celestial object data into whether they are evolving or non-evolving.

[0035] If the first model is F1 and the second model is F2, equation (1) can be written as equation (5) below.

[0036]

number

[0037] First, we will explain the period analysis of celestial data. To perform period analysis, explanatory variables must be prepared. As described below, in this invention, as in the existing method, m sine waves and m cosine waves with different period lengths are prepared as explanatory variables, totaling 2m (see Reference 3, for example). The sine waves and cosine waves can be expressed as the following equation (6). (Reference 3: Makoto Uemura, “How to Sparse Modeling for Astronomers 3 Application to Periodic Analysis of Time Series Data”)

[0038]

number

[0039] In equation (6), t i and ν j The maximum frequency that can be expressed is f n Then, it is defined as shown in the following equation (7).

[0040]

number

[0041] In the present invention, a group set g is defined to use the Group Lasso as described above. In the period analysis, in order to extract the period length of a celestial object including the phase, an arbitrary t i and a certain frequency ν j For the sine wave cos(2πt i ν j ) and cosine wave sin(2πt i ν j ) is selected as a set of features. As shown in equation (6), i and ν j A sine wave with cos(2πt i ν j ) and cosine wave sin(2πt i ν j ) are m sets of sine waves cos(2πt i ν j ) is x ij =cos(2πt i νj ), then the cosine wave sin(2πt i ν j ) is x i,m+j =sin(2πt i ν j Therefore, in the present invention, the group set g is defined as g={{j, m+j}} m Let us assume that.

[0042] [Device configuration] Next, an apparatus for performing the above-mentioned processing will be described. FIG. 2 is a diagram showing an example of the configuration of the detection apparatus 1 in the embodiment. The detection apparatus 1 is an apparatus for detecting a variable celestial object using celestial object data. The detection apparatus 1 includes a control unit 10. The control unit 10 includes a CPU (Central Processing Unit), a memory, an auxiliary storage device, and the like connected by a bus, and executes a program. By executing the program, the control unit 10 functions as an apparatus including an explanatory variable acquisition unit 110, a group set acquisition unit 120, a feature acquisition unit 130, a classifier learning unit 140, and a variable celestial object detection unit 150. Note that all or part of the functions of the control unit 10 may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The program may also be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM (Read Only Memory), or a CD-ROM, or a storage device such as a hard disk built into a computer system. The program may also be transmitted and received via an electric communication line.

[0043] The explanatory variable acquisition unit 110 acquires an explanatory variable X in the periodicity analysis.

[0044] The group set acquisition unit 120 uses the explanatory variables X acquired by the explanatory variable acquisition unit 110 to acquire a group set g of the explanatory variables X used in the learning performed by the classifier learning unit 140.

[0045] The feature amount acquiring unit 130 acquires the feature amount based on the explanatory variable X acquired by the explanatory variable acquiring unit 110 and the group set g acquired by the group set acquiring unit 120. More specifically, the feature amount acquiring unit 130 acquires the feature amount for each celestial body data by using a feature amount selection method based on sparse modeling.

[0046] The classifier learning unit 140 performs learning using the features acquired by the feature acquisition unit 130 and the labels as input. The labels input to the classifier learning unit 140 are label "1" representing a variable celestial object and label "0" representing a non-variable celestial object. Thus, the classifier learning unit 140 will be described on the assumption that LightGBM is used as a classifier, but other classifiers such as a support vector machine may also be used.

[0047] The variable celestial object detection unit 150 classifies celestial objects included in the celestial object data into either variable celestial objects or non-variable celestial objects using a trained classifier trained by the classifier training unit 140. More specifically, the variable celestial object detection unit 150 detects variable celestial objects using a classifier trained to classify celestial objects shown in the celestial object data into either variable celestial objects or non-variable celestial objects, using as input the feature amount for each celestial object data acquired by the feature amount acquisition unit 130 and the feature amount related to the brightness of the celestial object acquired for each celestial object data.

[0048] (Periodic analysis process by detection device 1) The flow of period analysis processing by the detection device 1 will be described. Fig. 3 is algorithm 1 showing the procedure of period analysis processing performed by the detection device 1, and Fig. 4 is a diagram showing the processing of algorithm 1 as a flowchart. Below, the flow of period analysis processing by the detection device 1 will be described with reference to Fig. 4. In algorithm 1, frequency candidate "ν" and normalization parameter "λ" are parameters in Group Lasso. Also, -y is the average vector of y.

[0049] The feature acquisition unit 130 receives as input the explanatory variable X, the objective variable y, and the group set g (step S101). Here, the explanatory variable X is acquired by the explanatory variable acquisition unit 110, and the group set g is acquired by the group set acquisition unit 120. Next, the detection device 1 initializes the regression coefficient w using a zero vector (step S102). The detection device 1 updates the value of the regression coefficient w according to the update formula for the Group Lasso. Specifically, first, the detection device 1 initializes X T By performing the calculation of (Xw-y) / n, the calculation result obtained is defined as ∇f(w) (step S103). Next, the detection device 1 multiplies ∇f(w) obtained in the process of step S103 by the step width η, and subtracts the multiplication result (η∇f(w)) from the regression coefficient w. The calculation result obtained in this way is defined as ^w (step S104).

[0050] After that, the detection device 1 executes the loop Ls1 to the loop Le1. In the loop Ls1 to the loop Le1, the maximum value w Gj (Step S105).

[0051]

number

[0052] The detection device 1 determines whether the regression coefficient w has converged (step S106). When the detection device 1 determines that the regression coefficient w has not converged (step S106-NO), the detection device 1 executes the process from step S103 onwards again. The detection device 1 repeats the process of updating the value of the regression coefficient w until the value of the regression coefficient w converges. On the other hand, when the detection device 1 determines that the regression coefficient w has converged (step S106-YES), the detection device 1 calculates the coefficient of determination reg_score based on the following formula (9) using the converged regression coefficient w (step S107). As a result, the detection device 1 calculates the coefficient of determination reg_score for the converged regression coefficient w.

[0053]

number

[0054] The detection device 1 uses the value of the regression coefficient w after convergence and the coefficient of determination reg_score obtained based on the regression coefficient w after convergence as feature quantities of the classifier.

[0055] (Detection process of moving celestial objects by detector 1) The flow of the variable celestial object detection process by the detection device 1 will be described. FIG. 5 is algorithm 2 showing the procedure of the variable celestial object detection process performed by the detection device 1, and FIG. 6 is a diagram showing the process of algorithm 2 as a flowchart. First, an outline of algorithm 2 shown in FIG. 5 will be described. As shown in lines 2-9 of algorithm 2 shown in FIG. 5, the detection device 1 creates necessary variables. Specifically, as shown in lines 2-6 of algorithm 2 shown in FIG. 5, the detection device 1 creates explanatory variables X, and as shown in lines 8-9, creates a group set g. After that, the detection device 1 executes period analysis based on algorithm 1, as shown in lines 11-13 of algorithm 2 shown in FIG. 5. As a result, the detection device 1 obtains the regression coefficient w and the coefficient of determination reg_score.

[0056] The detection device 1 combines the obtained regression coefficient w and the coefficient of determination reg_score with the domain feature list F d Based on the above, a feature list F required for binary classification is created. After that, the detection device 1 executes binary classification as shown in lines 15-18 of algorithm 2 shown in Fig. 5. In particular, the division into training data and test data executed in line 15 of algorithm 2 shown in Fig. 5 can be utilized for predicting variable celestial bodies by using a data set in which no labels are assigned to the test data.

[0057] The process described in lines 2-6 of algorithm 2 shown in Fig. 5 is executed by the explanatory variable acquisition unit 110. The process described in lines 8-9 of algorithm 2 shown in Fig. 5 is executed by the group set acquisition unit 120. The process described in lines 11-13 of algorithm 2 shown in Fig. 5 is executed by the feature acquisition unit 130. The process described in lines 15-16 of algorithm 2 shown in Fig. 5 is executed by the classifier learning unit 140. The process described in line 18 of algorithm 2 shown in Fig. 5 is executed by the varying celestial object detection unit 150.

[0058] Hereinafter, the flow of the process of detecting a moving celestial object by the detection device 1 will be described with reference to Fig. 6. The detection device 1 obtains a data set Y, a number m of candidates for frequencies used in period analysis of a celestial object, and a Nyquist frequency f n and the domain feature list F d The explanatory variable acquisition unit 110 executes a loop Ls2 to a loop Le2 until the value of the constant i changes from 1 to n. Here, n=2. First, the explanatory variable acquisition unit 110 sets the value of the constant i to "1". In this case, the explanatory variable acquisition unit 110 executes an arbitrary t i (Step S202). After that, the explanatory variable acquisition unit 110 executes the loop Ls3 to the loop Le3 until the value of the constant j changes from 1 to m. Here, m=2.

[0059] First, the explanatory variable acquisition unit 110 sets the value of the constant j to “1.” In this case, the explanatory variable acquisition unit 110 j "f n / m" is substituted for i=1, j=1 (step S203). Then, the explanatory variable acquisition unit 110 substitutes cos(2πf n / m), X[1][1+m] as sin(2πf n / m) is acquired (step S204). After that, the explanatory variable acquisition unit 110 adds "1" to the value of j. As a result, the value of the constant j after the addition becomes "2".

[0060] Next, the explanatory variable acquisition unit 110 executes the processes of steps S203 and S204 again with the value of the constant j after the addition being “2”. In this case, the explanatory variable acquisition unit 110 performs the processes of steps S203 and S204 again with the value of the constant j after the addition being “2”. j "2F n / m" is substituted for X[1][2] when i=1, j=2 (step S203). n / m), X[1][2+m] as sin(4πf n / m) is acquired (step S204). Since the value of the constant j has become m, the explanatory variable acquisition unit 110 ends the processing from loop Ls3 to loop Le3.

[0061] Next, the explanatory variable acquisition unit 110 adds "1" to the value of i. As a result, the value of the constant i after the addition becomes "2". In this case, the explanatory variable acquisition unit 110 i (step S202). After that, the explanatory variable acquisition unit 110 executes loops Ls3 to Le3 until the value of the constant j changes from 1 to m. First, the explanatory variable acquisition unit 110 sets the value of the constant j to "1". In this case, the explanatory variable acquisition unit 110 assigns "2" to a certain frequency v j "f n / m" is substituted for X[2][1] when i=2, j=1 (step S203). n / m), X[2][1+m] as sin(4πf n / m) is acquired (step S204). After that, the explanatory variable acquisition unit 110 adds "1" to the value of j. As a result, the value of the constant j after the addition becomes "2".

[0062] Next, the explanatory variable acquisition unit 110 executes the processes of steps S203 and S204 again with the value of the constant j after the addition being “2”. In this case, the explanatory variable acquisition unit 110 performs the processes of steps S203 and S204 again with the value of the constant j after the addition being “2”. j "2F n / m" is substituted for X[2][2] when i=2, j=2 (step S203). n / m), X[2][2+m] as sin(8πf n / m) is acquired (step S204). Since the value of the constant j has become m, the explanatory variable acquisition unit 110 ends the processing from loop Ls3 to loop Le3. Furthermore, since the value of the constant i has become n, the explanatory variable acquisition unit 110 ends the processing from loop Ls2 to loop Le2.

[0063] The group set acquisition unit 120 outputs the multiple explanatory variables X obtained by the above process to the group set acquisition unit 120 and the feature acquisition unit 130. The group set acquisition unit 120 receives the multiple explanatory variables X output from the explanatory variable acquisition unit 110. The group set acquisition unit 120 creates a group set of explanatory variables based on the multiple explanatory variables X input. Specifically, the group set acquisition unit 120 executes loops Ls4 to Le4 until the value of the constant j becomes 1 to m. Here, m=2. First, the group set acquisition unit 120 sets the value of the constant j to "1". In this case, the group set acquisition unit 120 sets a combination of a sine wave and a cosine wave corresponding to [j, j+m]=[1, 1+m] as one group g[1] (step S205). After that, the group set acquisition unit 120 adds "1" to the value of j. As a result, the value of the constant j after the addition becomes "2".

[0064] Similarly, the group set acquisition unit 120 sets a combination of a sine wave and a cosine wave corresponding to [j, j+m]=[2, 2+m] as one group g[2] (step S205). Because the value of the constant j has become m, the group set acquisition unit 120 ends the processing from loop Ls4 to loop Le4. The group set acquisition unit 120 outputs information about the created group set g to the feature acquisition unit 130. The feature acquisition unit 130 receives as input the multiple explanatory variables X output from the explanatory variable acquisition unit 110 and the information about the group set g output from the group set acquisition unit 120.

[0065] The feature acquisition unit 130 acquires the feature of each celestial body data based on the input multiple explanatory variables X, the information of the group set g, and the data set Y. Specifically, the feature acquisition unit 130 executes the process from the loop Ls5 to the loop Le5 until the value of the constant i becomes N. Here, N=2. Note that N represents the number of celestial body data used for learning. First, the feature acquisition unit 130 sets the value of the constant i to "1". In this case, the feature acquisition unit 130 acquires the regression coefficient w and the determination coefficient reg_score in the celestial body data y[1] by executing the process of the algorithm 1 using the input multiple explanatory variables X, the information of the group set g, and the celestial body data y[1] included in the data set Y (step S206). Here, the celestial body data y[1] represents the first celestial body data among the multiple celestial body data included in the data set Y.

[0066] Furthermore, the feature acquisition unit 130 acquires a domain feature list F d [1] is obtained. The domain feature list F of the astronomical data y[1] is obtained. d [1] includes the mean m_ap40, z_aptureflux_40_mag, diff_ap40, and label obtained based on the celestial object data y[1]. Then, the feature acquisition unit 130 combines the acquired regression coefficient w and coefficient of determination reg_score with the domain feature list F d Let [1] be the feature of the celestial object data y[1]. Then, the feature acquisition unit 130 adds "1" to the value of i. As a result, the value of the constant i after the addition becomes "2".

[0067] In this case, the feature acquisition unit 130 executes the process of algorithm 1 using the input multiple explanatory variables X, the information of the group set g, and the celestial body data y[2] included in the data set Y to acquire the regression coefficient w and the coefficient of determination reg_score in the celestial body data y[2] (step S206). Furthermore, the feature acquisition unit 130 acquires the domain feature list F of the celestial body data y[2]. d [2] is obtained. The domain feature list F of the astronomical data y[2] is obtained. d[2] includes the mean m_ap40, z_aptureflux_40_mag, diff_ap40, and label obtained based on the celestial object data y[2]. Then, the feature acquisition unit 130 combines the acquired regression coefficient w and coefficient of determination reg_score with the domain feature list F d [2] is the feature of astronomical data y[2].

[0068] Since the value of the constant i has become N, the feature acquisition unit 130 ends the processing from loop Ls5 to loop Le5. The feature acquisition unit 130 generates a feature list F that summarizes information on the features of the acquired N pieces of celestial body data (step S207). The feature acquisition unit 130 outputs the generated feature list F to the classifier training unit 140.

[0069] The classifier training unit 140 receives the feature list F output from the feature acquisition unit 130. The feature list F may include features obtained from the test celestial body data. In this case, the features obtained from the test celestial body data do not include a label. The classifier training unit 140 receives the features (e.g., the regression coefficient w and the coefficient of determination reg_score) for each celestial body data y[i] included in the input feature list F, and the domain feature list F. d The classifier learning unit 140 obtains the domain feature list F d The features whose labels are included in [i] are used as learning features, and the domain feature list F d The feature quantity that includes label in [i] is classified as a test feature quantity (step S208). After that, the classifier training unit 140 trains the classifier using the training feature quantity (step S209). At this time, the classifier training unit 140 trains so as to minimize the loss function using label as the objective variable and the regression coefficient w, the coefficient of determination reg_score, the mean m_ap40, z_apertureflux_40_mag, and diff_ap40 as explanatory variables.

[0070] When the learning end condition is satisfied, the variable object detection unit 150 inputs the test feature to the trained classifier (step S210). Here, the learning end condition is a condition for terminating the learning of the classifier, and may be, for example, that a predetermined number of learning processes have been performed, or that the error has become less than a threshold. The test feature is acquired by the explanatory variable acquisition unit 110, the group set acquisition unit 120, and the feature acquisition unit 130 according to the algorithms shown in Figs. 3 and 5. The variable object detection unit 150 outputs the result of classifying the celestial objects included in the test celestial object data into either variable objects or non-variable objects (step S211).

[0071] The detection device 1 configured as above includes a feature acquisition unit 130 that acquires features for each celestial body data, which is time-series data showing changes in the brightness of the celestial body, by using a feature selection method based on sparse modeling, and a variable celestial body detection unit 150 that detects variable celestial bodies using a classifier trained to classify celestial bodies shown in the celestial body data into variable celestial bodies or non-variable celestial bodies using the feature acquisition unit 130 for each celestial body data acquired by the feature acquisition unit 130 (e.g., regression coefficient w, coefficient of determination reg_score) and the feature related to the brightness of the celestial body acquired for each celestial body data (e.g., average m_ap40, z_apertureflux_40_mag, and diff_ap40). In this way, the detection device 1 uses a feature selection method based on sparse modeling for learning significant properties from sparse data and extracting only information determined to be important, thereby making it possible to suppress overlearning for data with missing values, noise, etc. Furthermore, the detection device 1 creates features taking domain knowledge into consideration. This allows the classifier to be trained efficiently even when celestial body data containing many missing values, noise, etc. is used. The detection device 1 then uses the trained classifier thus trained to classify the celestial objects shown in the celestial object data into either unstable or non-unsteady objects, using the feature values ​​based on the celestial object data as input. This makes it possible to detect unstable celestial objects with high accuracy.

[0072] <Modification> In the above embodiment, a configuration has been shown in which one device performs learning of a classifier and inference (detection of a variable celestial object) using the learned classifier. Learning of a classifier and inference using the learned classifier may be performed by different devices. For example, learning of a classifier may be performed by a first device, and inference using the learned classifier may be performed by a second device. In this configuration, the first device outputs the learned classifier obtained by learning the classifier to the second device. The second device performs inference using the learned classifier obtained from the first device. In the above configuration, the first device includes an explanatory variable acquisition unit 110, a group set acquisition unit 120, a feature acquisition unit 130, and a classifier learning unit 140. The second device includes an explanatory variable acquisition unit 110, a group set acquisition unit 120, a feature acquisition unit 130, and a variable celestial object detection unit 150.

[0073] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included. [Explanation of symbols]

[0074] 1...detection device, 10...control unit, 110...explanatory variable acquisition unit, 120...group set acquisition unit, 130...feature acquisition unit, 140...classifier learning unit, 150...changing celestial object detection unit

Claims

1. a feature acquisition unit that acquires features for each celestial body data, which is time-series data representing changes in the brightness of the celestial body, by using a feature selection method based on sparse modeling; a detection unit that detects variable celestial objects using a classifier trained to classify celestial objects shown in the celestial object data into either variable celestial objects or non-variable celestial objects, using as input the feature values ​​for each celestial object data acquired by the feature value acquisition unit and the feature values ​​related to the brightness of the celestial object acquired for each celestial object data; A detection device comprising:

2. an explanatory variable acquisition unit that acquires a plurality of sine waves and a plurality of cosine waves having different period lengths as explanatory variables; a group set acquisition unit that acquires a group set of explanatory variables based on the explanatory variables acquired by the explanatory variable acquisition unit; Further equipped with the feature acquisition unit acquires feature values ​​for each celestial object data based on the group set of explanatory variables, the explanatory variables, and the objective variable; The detection device according to claim 1 .

3. the feature acquisition unit acquires features for each celestial object data based on a Group Lasso feature selection method using the group set of explanatory variables, explanatory variables, and a target variable; 3. The detection device according to claim 2.

4. a learning unit that learns a classifier based on the feature amount for each celestial body data acquired by the feature amount acquisition unit, the feature amount related to the brightness of the celestial body acquired for each celestial body data, and ground truth data; 4. A detection device according to claim 1.

5. By using a feature selection method based on sparse modeling, features are obtained for each celestial body data, which is time-series data that shows changes in the brightness of celestial bodies, A detection method for detecting variable celestial objects using a classifier trained to classify celestial objects shown in the celestial object data as either variable or non-variable celestial objects, using as input feature values ​​for each acquired celestial object data and feature values ​​related to the brightness of the celestial object obtained for each said celestial object data.

6. On the computer, a feature acquisition step of acquiring features for each celestial body data, which is time-series data representing changes in the brightness of the celestial body, by using a feature selection method based on sparse modeling; a detection step of detecting a variable celestial object using a classifier trained to classify the celestial objects shown in the celestial object data into either a variable celestial object or a non-variable celestial object, using as input the feature values ​​for each celestial object data acquired in the feature value acquisition step and the feature values ​​related to the brightness of the celestial object acquired for each celestial object data; A program for executing the above.

Images