JP2015052994A - Feature selection device, learning device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To be capable of selecting a feature effective for identification while maintaining consistency in time-of-day order between each of features.SOLUTION: A feature extraction unit 32 extracts, for each of two-times-of-day of learning data, a feature defined by a value of two-times-of-day from each of a plurality of learning data to which a class indicated by time series data expressed by a value of each time of day is previously imparted. An initial feature selection unit 342 selects, for each class, a two-times-of-day of at least one feature effective for class identification on the basis of each of two-times-of-day features and the class imparted to each of learning data. A time-of-day order feature selection unit 344 selects, for each class as two-times-of-day of a feature used for identifying whether the time series data to be identified represents a class, a two-times-of-day of a feature from two-times-of-day of the feature selected for the class so that consistency in time-of-day order between each of features is maintained.

Description

  The present invention relates to a feature selection device, a learning device, a method, and a program.

  Conventionally, recognition of objects such as characters and gestures represented by time-series data such as online handwritten character recognition and gesture recognition has been performed. As a typical technique for recognizing time series data, for example, a standard time pattern and an input pattern are generated by combining all character strokes and motion trajectories in order of time. A method of recognizing by performing DP matching (matching by dynamic programming (dynamic programming)) between feature points of movement at the time of each sampling in a handwriting coordinate series or gesture has been proposed (Non-Patent Document 1). .

  In addition, as a feature used for DP matching, there is a method using not only the coordinates of feature points at the time of each sampling but also a local direction defined as a relative coordinate value as a difference between feature points at adjacent sampling times.

  Furthermore, as a relative coordinate value, a method using a global feature including a relative coordinate value as a difference between feature points not only at adjacent sampling times but also at further apart sampling times has been proposed (Non-Patent Document 2). ).

  In addition, in order to perform DP matching using global features, a plurality of features are arranged in time order so as to maintain consistency in time order using an index indicating the effectiveness of the feature from a preset start position. A method of selecting is proposed (Non-patent Document 3).

Yukio Sato, Hidetsuna Adachi, “Online Recognition of Scribbled Characters” IEICE Transactions (D), Vol. J68- (D), Nol. 12, pp. 2116-2122 Satoshi Mori, Seiichi Uchida, Sakano et al., “Online Number Recognition Using Global Structure Information” IEICE Technical Report Pattern Recognition / Media Understanding Study Group, vol. 111, no. 317, PRMU2011-102, pp. 19-24, Nov. 2011 Satoshi Mori, Seiichi Uchida, Sakano et al., “DP Matching for Global Features (J) -Vol.J96-D, No.7, pp.1655-1657

  However, as in the method described in Non-Patent Document 1, the DP matching method using feature points at each sampling time can absorb a partial coordinate shift by DP matching, but a large change occurs in the coordinates. In such a case, there is a problem that the deviation cannot be absorbed and may be erroneously recognized by another pattern.

  In addition, the DP matching method using the local direction improves the tolerance against the shift of the coordinate value, but conversely, by using no feature point at each sampling time, it has a similar shape existing at different coordinates. There is a problem that it may be misrecognized by another pattern.

  In addition, in the method using global features, such as the method described in Non-Patent Document 2, it is possible to improve tolerance to larger coordinate fluctuations by using global structure information. However, since the features are not arranged in time series, there is a problem that it is impossible to apply a recognition method that allows variation in the time series direction such as DP matching.

  In addition, by using the method described in Non-Patent Document 3, it is possible to apply a global feature to DP matching, but it is necessary to determine the DP matching start position from the top down, and from the start position. Since the features are selected so as to maintain the consistency in time order, there is a problem that the features to be selected are limited, and features that are effective for identification are not necessarily selected.

  The present invention has been made to solve the above-described problems, and a feature selection device and method capable of selecting features effective for identification while maintaining consistency in time order between the features. And to provide a program.

  It is another object of the present invention to provide a learning device that can maintain a time-order consistency between features and obtain features for accurate identification.

  In order to achieve the above object, the feature selection device according to the first aspect of the present invention provides a value between two times from each of a plurality of learning data to which a class represented by time-series data represented by values at each time is given in advance. And a feature extracting unit that extracts the features defined in (2) for each of two times of the learning data, and (2) each feature of the learning data for two times extracted for each of the learning data by the feature extracting unit; An initial feature selection unit that selects, for each of the plurality of classes, two time points of the plurality of features effective for identifying the class, based on the class assigned to each of the learning data, and the class For each of the two times of the feature used to identify whether the time-series data to be identified represents the class or not, the initial feature selection unit selects the feature selected for the class. From between the time, it is configured to include a, a time order feature selecting section that selects between two times of the feature to retain the chronological consistency in each between features.

  A feature selection method of a second invention is a feature selection method in a feature selection device including a feature extraction unit, an initial feature selection unit, and a time-order feature selection unit, wherein the feature extraction unit includes a value at each time A feature defined by a value between two times is extracted from each of a plurality of learning data to which a class represented by the time-series data represented by the above is assigned in advance, and the initial The feature selection unit is configured to determine the class based on each feature of the learning data extracted for each of the learning data by the feature extraction unit between two times and the class assigned to each of the learning data. For each class, it selects between two times of at least one feature effective for class identification, and the time-order feature selection unit determines whether the time-series data to be identified represents the class for each class. Identify As the two times of the feature used for the purpose, the time order consistency in each of the features is maintained from the two times of the feature selected for the class in the initial feature selection unit. Select between two times of features.

  According to a third aspect of the present invention, there is provided a learning device according to the first aspect, and for each class, the learning data in which the class is previously assigned to the class by the time-order feature selection unit. A selection feature extraction unit that extracts the feature of the learning data as a feature used for identifying whether or not the time-series data to be identified represents the class for two times of the selected feature; It is comprised including.

  Moreover, the program of this invention is a program for making a computer perform each part which comprises said feature selection apparatus.

  As described above, according to the feature selection device, method, and program, it is possible to select features that are effective for identification while maintaining consistency in time order between the features.

  Further, according to the learning device, it is possible to obtain features for accurate identification while maintaining consistency in time order between the features.

It is a block diagram which shows the functional structure of the feature selection apparatus which concerns on embodiment of this invention. It is the schematic which shows an example of a global characteristic. It is the schematic for demonstrating DP matching. It is the schematic which shows an example of the global feature selected while maintaining the time order consistency, and the global feature selected without maintaining the time order consistency. It is the schematic for demonstrating selection of the characteristic. It is a flowchart figure which shows the learning process routine in the feature selection apparatus concerning embodiment of this invention. It is a flowchart figure which shows the identification process routine in the feature selection apparatus concerning embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, an example in which the present invention is applied to a character identification device that identifies online characters will be described. The on-line character is a character handwriting represented by a handwriting coordinate series for each stroke, that is, time-series data in which a character pattern is represented by a handwriting coordinate value at each sampling time.

<Configuration of Feature Selection Device According to this Embodiment>
First, the configuration of the feature selection device according to the embodiment of the present invention will be described. As shown in FIG. 1, a feature selection apparatus 100 according to an embodiment of the present invention includes a CPU, a RAM, a ROM that stores a program and various data for executing a learning processing routine and an identification processing routine described later, and , Can be configured with a computer including. Functionally, the feature selection apparatus 100 includes an input unit 10, a calculation unit 20, and an output unit 50 as shown in FIG.

  The input unit 10 receives online character data for learning (hereinafter referred to as “learning data”) having a known character class from an input device such as a keyboard. The input unit 10 also receives identification target data that is time-series data whose character class is unknown from an input device such as a keyboard. Note that the input unit 10 may accept input from the outside via a network or the like. Moreover, in this Embodiment, since 10 numbers (0-9) of numbers are received as learning data, several learning data are received for every class.

  The calculation unit 20 includes a selection feature storage unit 22, a learning unit 30, and an identification unit 40.

  The selection feature storage unit 22 stores the selection feature extracted by the learning unit 30 for each learning data.

  The learning unit 30 includes a feature extraction unit 32 and a feature selection unit 34.

  The feature extraction unit 32 uses, as a feature, a relative vector connecting any two points on the stroke from the learning data for each of the learning data to which any of the ten classes received by the input unit 10 is assigned. Extract each one. Note that two points are an example between two times.

  Specifically, the stroke size represented by the learning data is normalized, and time-series data composed of N feature points is resampled. Here, resampling means sampling the feature points by determining the number of feature points set for one character stroke. When sampling feature points on a character stroke written in real time, if the feature points are taken at regular intervals, the character writing speed varies depending on the person, so the number of feature points per character varies. is there. For example, the number of feature points set for one character stroke is N, and resampling is performed so that the interval between the feature points becomes a constant distance. Note that the number of N needs to be a number that can express the character shape sufficiently.

Next, a relative vector (dx, dy) defined between two feature points is extracted as a feature. Here, the features of the first dimension,..., The features of the jth dimension,. J is the number N × (N−1) / 2 of combinations for selecting two from N feature points. As described above, in this embodiment, global features represented by a relative vector between any two points on the stroke including a relative vector between two adjacent points and a relative vector between two non-adjacent points are extracted. To do. FIG. 2 shows an example of a global feature at the feature point _Pn . Here, P _n represents the n-th feature point among the N feature points obtained by resampling. For example, _{assuming that} the coordinate value of the feature point P _n is (x _n , y _n ) and the coordinate value of the feature point P _N is (x _N , y _N ), the relative vector (P _n → P _N ) between the two points. the relative position of the x-axis direction of the feature point _{P n} for the feature point _{P n dx = x n -x n} , represented by the relative positions of the y-axis direction dy ₌ y n -y _n.

  The feature selection unit 34 performs an identification process that allows variation in the time-series direction for each class based on the features extracted from each of the learning data by the feature extraction unit 32 (details will be described later). A feature that is effective for identification and that maintains consistency in time order is selected so that it can be applied to.

  Here, the reason why the feature selection unit 34 selects a feature that maintains consistency in time order will be described. Here, an example will be described in which the identification processing unit 44 performs DP matching as the identification processing that allows variation in the time-series direction. Note that “identification processing that allows variation in the time-series direction” is an identification process that enables evaluation of matching and matching even when expansion or contraction in the time-series direction occurs at a certain point in the time-series data. It is. That is, when comparing two time-series data sampled at a certain time interval or a certain distance interval, one is time-series data consisting of M points and the other is N (N> M or N <M) This is an applicable identification process even for time-series data consisting of these points.

In DP matching using global features, as shown in FIG. 3, the correspondence destination in the input data of the global features on the reference data is sequentially optimized. In the algorithm for optimum matching of global features by DP matching, the variable set to be optimized is {(s (k), t (k)) | k = 1,..., K, as shown in FIG. }. s (k) and t (k) are the corresponding points on the input data for the two points S _k and T _k that give the k th global feature selected from the reference data, respectively. These variable sets are arranged in the order of (s (1), t (1)), ..., (s (k), t (k)), ..., (s (K), t (K)). Consider the process of sequential determination. The objective function is the sum of K Euclidean distances between corresponding global features. In order to minimize the objective function, the value of each variable, that is, the correspondence relationship is determined.

  Here, as shown in FIG. 4A, when the global feature is selected, that is, in the global feature column maintaining the total order, it corresponds to the start point and the end point of the relative vector which is the global feature. When the order (time) of feature points monotonously increases or decreases, only the values of (s (k-1), t (k-1)) affect the determination of (s (k), t (k)). do not do. Therefore, this determination process is a Markov determination process with consistency in time order, and the optimal solution is obtained by DP matching. On the other hand, as shown in FIG. 4B, when the selected features are arranged, the order of the feature points corresponding to the start point of the relative vector is reversed in the middle, or the order of the feature points corresponding to the end point is changed. If a global feature is selected that returns in the middle, or the order of the feature points corresponding to the start point and the order of the feature points corresponding to the end point is reversed, the consistency in time order Cannot be maintained, and it is impossible to optimize the association by the above DP matching. This is apparent from the fact that values other than k-1 affect the determination of k, which is not a Markov determination process and does not involve time order consistency.

For the reasons described above, the feature selection unit 34 needs to select a feature that maintains consistency in time order. However, in the technique of Non-Patent Document 3, feature selection is started in advance in order to satisfy the above condition. The position to be selected is determined in order based on the index indicating the effectiveness of the identification from the start position. Therefore, selectable features are limited by the start position, and which position is an appropriate start position. There is a problem that it is difficult (see Non-Patent Document 3).

In the present embodiment, unlike the method of Non-Patent Document 3, it is not necessary to set a start position in advance, and a method that enables effective feature selection by identification is adopted.

  The method in this embodiment can select features that are more effective for identification, but by using a feature selection method that does not maintain consistency in time order, by adding new restrictions to the selection conditions, Consistency is maintained, and it is not necessary to set the selection start position in advance, and an effective feature can be selected without depending on the start position.

  The feature selection unit 34 includes an initial feature selection unit 342, a time-order feature selection unit 344, and a selection feature extraction unit 346. FIG. 5 shows an example of a process in which features are selected by the feature selection unit 34.

  The initial feature selection unit 342 selects, for each class, two features that are effective for identifying the class as initial features from each of the features extracted by the feature extraction unit 32. In the present embodiment, as the initial feature selection processing, in the boosting processing, for each class, extraction is performed by the feature extraction unit 32 based on learning data of the class and learning data to which a class other than the class is assigned. The initial feature is selected by selecting between the two points of the feature effective for identifying the class from the feature between the two points. For this boosting process, for example, a known AdaBoost (Adaptive Boosting, Adaboost) method can be used (Non-Patent Document 4: Y. Freund and R. Shapire, “A decision-theoretical generation of on-line.” an application to boosting "Journal of Computer and System Sciences, 1997.). In the AdaBoost method, from the viewpoint of minimizing the misclassification rate, two points of features that complement well the weak points between the two points of already selected features are successively selected. It is possible to select between two points of a plurality of features as an initial feature in order from the two points.

  In addition, since boosting processing is performed by comparing two classes, as in the present embodiment, a device for identifying two or more classes needs to be devised for use. In the present embodiment, since 10 numbers are the identification target, the class “0” and other classes, and the class “1” and other classes are classified into 10 ways of 0 to 9, and the numbers of each class are used. The initial feature corresponding to is selected. In the boosting process, two points of one feature are selected as initial features for the class. Further, in the second boosting process, a point between two points having the second most effective feature is selected.

  However, as described above, in the feature selection method such as boosting processing, each feature is selected in order from features that are effective for identification based on the criterion for minimizing the misidentification rate. Therefore, consistency in time order is not maintained, and it is impossible to apply the method as it is to DP matching or the like.

  For each class, the time-order feature selection unit 344 determines, for each class, between the two points of the feature selected as the initial feature of the class by the initial feature selection unit 342 and between the two points of the feature selected for the class. It is determined whether or not the consistency is maintained. Specifically, for each class, consistency in time order between the initial feature selected for the class in the initial feature selection unit 342 and the two points of the features already selected for the class is maintained. It is determined whether or not it is hit. Then, in the initial feature selected for the class, if it is determined that the time order consistency between two points of the feature selected for the class is maintained, the initial feature is If the selected feature of the class is used between two points, and if the consistency in time order is not maintained, the initial feature is not selected and not adopted. In addition, between the two selected features that have been adopted is stored in a memory (not shown) for each class. In addition, the initial feature determined to be the most effective in the initial feature selection unit 342 is adopted and stored as two points of the selected feature without performing the time order consistency determination process. Further, the process of the initial feature selection unit 342 and the process of the time order feature selection unit 344 are repeated until all the features selected as the initial features by the boosting process are completed.

The determination whether or not the two points of the feature selected as the initial feature of the specific class are consistent in time order with the two points of the feature selected for the class is specifically as follows. The following equation (1) is regarded as maintaining consistency in order of time when two points of the initial feature to be determined are satisfied with respect to two points of all selected features. Note that S _p and T _p represent two points of the selected feature, and S _q and S _p represent two points of the initial feature to be determined.

  The selected feature extraction unit 346 uses, for each class, each of the learning data belonging to the class, using each of the two points of the selected features selected for the class by the time-order feature selection unit 344, A feature (relative vector) is extracted from the learning data between two points that are the same as the two points of the selected feature and stored in the selected feature storage unit 22 as a selected feature.

  The identification unit 40 includes a feature extraction unit 42 and an identification processing unit 44.

  Similar to the feature extraction unit 32, the feature extraction unit 42 extracts features between arbitrary two points from the identification target data that is time-series data whose character class is unknown, which is received by the input unit 10.

  The identification processing unit 44 includes a similarity value calculation unit 442 and a determination unit 444.

  For each learning data stored in the selected feature storage unit 22, the similarity value calculation unit 442 includes each feature extracted for the identification target data in the feature extraction unit 42 and the learning stored in the selected feature storage unit 22. Based on each selected feature of the data, a similarity value between the identification target data and the learning data is calculated using a technique such as DP matching.

  The determination unit 444 extracts learning data having the highest similarity value based on the similarity value for each learning data calculated by the similarity value calculation unit 442, and sets the class to which the extracted learning data belongs to the class to which the identification target data belongs. And the determined class is output to the output unit 50 as an identification result.

<Operation of Feature Selection Device According to this Embodiment>
Next, the operation of the feature selection device 100 according to the embodiment of the present invention will be described. When each of the learning data to which any of the ten classes (0 to 9) of numbers is input by the input unit 10, the feature selection device 100 executes a learning processing routine shown in FIG.

  First, in step S100, each of learning data for every 10 classes (0 to 9) of numbers input by the input unit 10 is accepted.

  Next, in step S101, the size of the character pattern is normalized for each of the learning data acquired by the input unit 10 and resampled into time-series data including N feature points.

  Next, in step S102, for each of the learning data resampled in step S101, each of the relative vectors (dx, dy) defined between all two feature points is extracted as a feature (global feature). .

  Next, in step S104, a class to be processed is selected.

  Next, in step S106, in each step S102 of each of the learning data belonging to each of the classes not selected in step S104 and each of the features extracted in step S102 of each of the learning data belonging to the class selected in step S104. Based on each of the extracted features and the initial feature of the same class selected in the previous step S106, the initial feature of the class to be processed is selected.

  Next, in step S112, it is determined whether or not the initial feature selected in S106 satisfies the consistency in time order between two points of the selected feature obtained in the previous processing for the class to be processed. judge. If it is determined that the time order consistency is satisfied, the process proceeds to step S114. If it is determined that the time order consistency is not satisfied, the process proceeds to step S116. When the process is performed for the class to be processed for the first time, the process proceeds to step S114.

  Next, in step S114, the initial feature selected in step S106 is stored in a memory (not shown) as a point between the selected features of the class to be processed.

  Next, in step S116, it is determined whether or not processing has been completed for all features that can be selected as initial features for the class to be processed. If all the features that can be selected as the initial feature are processed, the process proceeds to step S118. If all the features that can be selected as the initial feature are not processed, the process returns to step S106. The process from S106 to step S116 is repeated.

  In step S118, for each piece of learning data belonging to the class to be processed, using each of the two points of the selected feature acquired in step S114, the two points that are the same as the two points of the selected feature are used. The feature (relative vector) of the learning data is extracted and stored in the selected feature storage unit 22 as a selected feature.

  Next, in step S120, it is determined whether or not processing has been completed for all classes. If the process has been completed for all classes, the process is terminated. If the process has not been completed for all classes, the process proceeds to step S104, the class to be processed is changed, and step S104 is performed. -The process of step S120 is repeated.

  Next, an identification processing routine in the feature selection device 100 according to the embodiment of the present invention will be described. When identification target data that is online character data whose class is unknown is input by the input unit 10, the feature selection device 100 executes an identification processing routine shown in FIG. 7.

  First, in step S200, identification target data input by the input unit 10 is received.

  Next, in step S202, the selected feature for each learning data stored in the selected feature storage unit 22 is read.

  Next, in step S204, the features of the identification target data acquired in step S200 are extracted as in step S102.

  Next, in step S206, target learning data is selected.

  Next, in step S208, based on the selected feature acquired in step S202 for the learning data selected in step S206 and the feature extracted in step S204, the learning data and the learning data are selected using a technique such as DP matching. The similarity value of is calculated.

  Next, in step S210, it is determined whether or not the processing has been completed for all target learning data. If the process has been completed for all target learning data, the process proceeds to step S212. If the process has not been completed for all target learning data, the process proceeds to step S206. The process of S210 is repeated.

  Next, in step S212, learning data having the highest similarity value is extracted based on the similarity value with each of the learning data calculated in step S208, and the class to which the learning data belongs is changed to the class to which the identification target data belongs. Identify as.

  Next, in step S214, the result identified in step S212 is output to the output unit 50, and the process ends.

  As described above, according to the feature selection apparatus according to the embodiment of the present invention, it is possible to select features that are effective for identification while maintaining consistency in time order between the features.

  In addition, it is possible to obtain features for accurate identification while maintaining consistency in time order between the features.

  In addition, when applying an identification method that allows variation in the time-series direction such as DP matching using features including global features, it is possible to select features that are more effective for identification.

  In addition, by selecting features that are more effective for identification from a plurality of global features extracted from time-series data and using features that are selected to maintain time order consistency, the coordinate values of each feature point Even in the case where features including global features as well as changes in the local direction are used, an identification method that allows variation in the time-series direction can be applied in the identification processing. That is, even if expansion / contraction in the time series direction occurs in the time series data, it becomes possible to identify.

Further, since it is not necessary to set the selection start position in advance as in Non-Patent Document 3, it is possible to select features that are more effective for identification without being restricted by the start position.

  In addition, by repeating the process of selecting the selected feature from the initial feature, it is possible to select a feature that is effective for identification and that maintains consistency in time order from features that are extracted from the learning data and that are effective for identification. .

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, in the present embodiment, the boosting process, specifically, the process using the AdaBoost method has been described as the initial feature selection process. However, the present invention is not limited to this, and other boosting techniques, A feature selection method other than the boosting process may be applied.

  In the present embodiment, the case where online character data is input as time series data and the character pattern is identified has been described. However, the present invention is not limited to this. For example, other time series such as gesture identification is used. It can also be applied to data.

  In the present embodiment, the learning unit and the identification unit are configured by the same computer. However, the present invention is not limited to this, and the learning unit and the identification unit may be configured by separate computers.

  In the present embodiment, the case where all the characteristics of the identification target data are used when identifying the class of the identification target data has been described. However, the present invention is not limited to this and is extracted from the identification target data. Of the features, the class of the identification target data may be identified using only the same selected features as the learning data.

Further, in the present embodiment, a case has been described in which only global features represented by relative vectors between two points are used as feature values used for pattern identification. When a global feature represented by a relative vector is used, a coordinate shift can be absorbed. However, even when the coordinate position is greatly different, erroneous identification may occur when the relative vectors are similar. Therefore, coordinate values may be used together with global features as feature values. Even when the coordinate value is used as a feature value, the j-th coordinate value of the i-th sample is set as P _{i, j} = (P _{i, j, x} , P _{i, j, y} ). It can be treated in the same way as special features. Also in this case, when the coordinate value itself represents an important structure of the character by using the AdaBoost method as in the above embodiment, the feature value indicating the coordinate value is adopted as the weak classifier.

  Further, in the present embodiment, the embodiment has been described as the embodiment in which the program is installed in advance. However, the present invention is not limited to this, and the program is stored in a computer-readable recording medium and provided. Is also possible.

DESCRIPTION OF SYMBOLS 10 Input part 20 Calculation part 22 Selection feature memory | storage part 30 Learning part 32 Feature extraction part 34 Feature selection part 40 Identification part 42 Feature extraction part 44 Identification processing part 50 Output part 100 Feature selection apparatus 342 Initial feature selection part 344 Time order feature selection Unit 346 selection feature extraction unit 442 similarity value calculation unit 444 determination unit

Claims (4)

A feature defined by a value between two times is extracted from each of a plurality of learning data to which a class represented by time-series data represented by each time value is assigned in advance for each of the two times of the learning data. A feature extraction unit,
For each class, based on each feature of the learning data extracted for each of the learning data for two times by the feature extraction unit and the class assigned to each of the learning data, the class An initial feature selection unit that selects between two times of at least one of the features effective for identification of
For each of the classes, the feature selected for the class by the initial feature selection unit as two times of the feature used to identify whether the time-series data to be identified represents the class A time-order feature selecting unit that selects between the two times of the features so as to maintain the time-order consistency in each of the features between the two times;
A feature selection device. A feature selection device according to claim 1;
For each of the classes, the feature of the learning data is identified for two times of the feature selected for the class by the time-order feature selection unit from each of the learning data to which the class is assigned in advance. A selection feature extraction unit that extracts as features used to identify whether the target time-series data represents the class;
Including learning device. A feature selection method in a feature selection device including a feature extraction unit, an initial feature selection unit, and a time-order feature selection unit,
The feature extraction unit extracts a feature defined by a value between two times from each of a plurality of learning data to which a class represented by time-series data represented by a value at each time is given in advance. Extract for each time interval,
The initial feature selection unit is based on each feature of the learning data extracted for each of the learning data by the feature extraction unit for two times, and the class assigned to each of the learning data, For each class, select between two time points of at least one of the features effective for class identification;
For each class, the time-order feature selection unit sets the class in the initial feature selection unit as two time points of the feature used to identify whether the time-series data to be identified represents the class or not. A feature selection method that selects between two times of the feature so as to maintain consistency in time order between the features from the two times of the feature selected for the feature. Computer
A feature defined by a value between two times is extracted from each of a plurality of learning data to which a class represented by time-series data represented by each time value is assigned in advance for each of the two times of the learning data. A feature extraction unit,
For each class, based on each feature of the learning data extracted for each of the learning data for two times by the feature extraction unit and the class assigned to each of the learning data, the class An initial feature selection unit that selects between two time points of at least one of the features effective for identification, and, for each class, used to identify whether the time-series data to be identified represents the class Between the two time points of the feature so as to maintain the time order consistency between each of the features from the two time points of the feature selected for the class in the initial feature selection unit. A time-order feature selection unit for selecting
Program to function as.