ANALYZING APPARATUS, ANALYSIS METHOD AND ANALYSIS PROGRAM

• US 20200134430A1
• Filed: 10/08/2019
• Published: 04/30/2020
• Est. Priority Date: 10/29/2018
• Status: Active Grant

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: YAMASHITA, Yasuho, SHIBAHARA, Takuma, SUZUKI, Mayumi
• Granted Patent: US 11,568,213 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 18/2148 characterised by the process organisation or structure, e.g. boosting cascade
• CPC Class Codes: G06F 18/2193 based on specific statistical tests
• CPC Class Codes: G06F 18/2415 based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
• CPC Class Codes: G06F 18/2453 non-linear, e.g. polynomial classifier
• CPC Class Codes: G06F 2218/12 Classification; Matching
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

The analyzing apparatus: generates first internal data; converts a position of first feature data in a feature space, based on the first internal data and a second learning parameter; reallocates, based on a result of first conversion and the first feature data, the first feature data to a position obtained through the conversion in the feature space; calculates a predicted value of a hazard function of analysis time in a case where the first feature data is given, based on a result of reallocation and a third learning parameter; optimizes the first to third learning parameters, based on a response variable and a first predicted value; generates second internal data, based on second feature data and the optimized first learning parameter; converts a position of the second feature data in the feature space, based on the second internal data and the optimized second learning parameter; and calculates importance data.

Claims

1. An analyzing apparatus accessible to a database, the analyzing apparatus including a processor that executes a program and a storage device that stores the program,

the database storing a training data set that includes pieces of training data by an amount equal to the number of learning targets, each piece of the training data including first feature data having a plurality of feature amounts of a learning target, a response variable indicating analysis time from a start of analysis to an end of the analysis about the learning target, and a variable indicating continuity of the analysis within the analysis time,

the processor executing:

a first generation process of generating first internal data on a basis of the first feature data and a first learning parameter;

a first conversion process of converting a position of the first feature data in a feature space on a basis of the first internal data generated in the first generation process, and a second learning parameter;

a reallocation process of, based on a result of first conversion in the first conversion process and the first feature data, reallocating the first feature data to a position obtained through the conversion in the feature space;

a first calculation process of calculating a first predicted value of a hazard function about the analysis time in a case where the first feature data is given, based on a result of reallocation in the reallocation process and a third learning parameter;

an optimization process of optimizing the first learning parameter, the second learning parameter, and the third learning parameter by a statistical gradient method on a basis of the response variable and the first predicted value calculated in the first calculation process;

a second generation process of generating second internal data on a basis of second feature data including a plurality of feature amounts of a prediction target and the first learning parameter optimized in the optimization process;

a second conversion process of converting a position of the second feature data in the feature space on a basis of second internal data generated in the second generation process and the second learning parameter optimized in the optimization process; and

an importance calculation process of calculating importance data including an importance of each feature amount of the second feature data on a basis of a result of second conversion in the second conversion process and the third learning parameter optimized in the optimization process.

2. The analyzing apparatus according to claim 1, wherein

the processor executes a second calculation process of calculating a second predicted value of the hazard function about the analysis time in a case where the second feature data is given, based on the importance data calculated in the importance calculation process and the second feature data.

3. The analyzing apparatus according to claim 1, wherein

the processor executes an output process of outputting the second feature data and the importance data in association with each other.

4. The analyzing apparatus according to claim 1, wherein

the processor executes the first generation process using a neuron group of a plurality of layers, each of the layers being allocated with the first learning parameter, and

in the first generation process, the processor:

uses neurons of a first layer among the neuron group to generate first internal data of the first layer on a basis of the first feature data and a first learning parameter of the first layer, and outputs the first internal data of the first layer to neurons of a second layer; and

uses neurons of a middle layer to generate first internal data of the middle layer on a basis of first internal data output from neurons of an adjacent higher layer and a first learning parameter of the middle layer, and outputs the first internal data of the middle layer to neurons of an adjacent lower layer.

5. The analyzing apparatus according to claim 4, wherein

the database stores the response variables and the variables corresponding to the first feature data, by an amount equal to the number of analysis targets,

the neuron group is constituted by a first neuron group from the first layer to a particular middle layer, and a latter neuron group of an amount equal to the number of the analysis targets from an adjacent lower layer of the particular middle layer to a lowermost layer,

in the first conversion process, for each combination of the first neuron group and the latter neuron group, the processor converts the position of the first feature data in the feature space, based on first internal data output from each neuron in the combination,

in the reallocation process, for the combination, the processor reallocates the first feature data to a position obtained through the conversion in the feature space,

in the first calculation process, for the combination, the processor calculates the first predicted value,

in the optimization process, for the combination, the processor optimizes the first learning parameter, the second learning parameter, and the third learning parameter by a statistical gradient method,

in the second generation process, the processor uses neurons of the first layer to generate second internal data of the first layer on a basis of the second feature data and the first learning parameter of the first layer, the first learning parameter being optimized in the optimization process, and outputs the second internal data of the first layer to neurons of the second layer;

and uses neurons of the middle layer to generate second internal data of the middle layer on a basis of second internal data output from neurons of an adjacent higher layer and a first learning parameter of the middle layer, and outputs the second internal data of the middle layer to neurons of an adjacent lower layer,

in the second conversion process, for the combination, the processor converts the position of the second feature data in the feature space, based on second internal data output from each neuron in the combination, and

in the importance calculation process, for the combination, the processor calculates the importance data.

6. The analyzing apparatus according to claim 1, wherein

in the first calculation process, the processor calculates the first predicted value by classification operation, and

in the optimization process, a DeepHit loss function is used as a loss function used in determination of the first learning parameter, the second learning parameter, and the third learning parameter.

7. The analyzing apparatus according to claim 1, wherein

in the first calculation process, the processor calculates the first predicted value by regression operation, and

in the optimization process, a likelihood function is used as a loss function used in determination of the first learning parameter, the second learning parameter, and the third learning parameter.

8. The analyzing apparatus according to claim 7, wherein

the regression operation is a Cox regression operation, and

the likelihood function is a Cox partial likelihood function based on an Efron method, a Breslow method, or an Exact method.

9. The analyzing apparatus according to claim 7, wherein

the regression operation is exponential regression operation, and

the likelihood function is a logarithmic likelihood function of an exponential regression model.

10. The analyzing apparatus according to claim 7, wherein

the regression operation is Weibull regression operation, and

the likelihood function is a logarithmic likelihood function of a Weibull regression model.

11. The analyzing apparatus according to claim 7, wherein

the regression operation is logarithmic logistic regression operation, and

the likelihood function is a logarithmic likelihood function of a logarithmic logistic regression model.

12. An analysis method used by an analyzing apparatus that is accessible to a database and includes a processor that executes a program and a storage device that stores the program,

the database storing a training data set that includes pieces of training data by an amount equal to the number of learning targets, each piece of the training data including:

first feature data having a plurality of feature amounts of a learning target, a response variable indicating analysis time from a start of analysis to an end of the analysis about the learning target, and a variable indicating continuity of the analysis within the analysis time, the analysis method comprising;

by the processor,

a first generation process of generating first internal data on a basis of the first feature data and a first learning parameter;

a first conversion process of converting a position of the first feature data in a feature space on a basis of the first internal data generated in the first generation process and a second learning parameter;

a reallocation process of, based on a result of first conversion in the first conversion process and the first feature data, reallocating the first feature data to a position obtained through the conversion in the feature space;

a first calculation process of calculating a first predicted value of a hazard function about the analysis time in a case where the first feature data is given, based on a result of reallocation in the reallocation process and a third learning parameter;

an optimization process of optimizing the first learning parameter, the second learning parameter, and the third learning parameter by a statistical gradient method on a basis of the response variable and the first predicted value calculated in the first calculation process;

a second generation process of generating second internal data on a basis of second feature data including a plurality of feature amounts of a prediction target and the first learning parameter optimized in the optimization process;

a second conversion process of converting a position of the second feature data in the feature space on a basis of second internal data generated in the second generation process and the second learning parameter optimized in the optimization process; and

an importance calculation process of calculating importance data including an importance of each feature amount of the second feature data on a basis of a result of second conversion in the second conversion process and the third learning parameter optimized in the optimization process.

13. An analysis program to be executed by a processor accessible to a database,

the database storing a training data set that includes pieces of training data by an amount equal to the number of learning targets, each piece of the training data including:

first feature data having a plurality of feature amounts of a learning target, a response variable indicating analysis time from a start of analysis to an end of the analysis about the learning target, and a variable indicating continuity of the analysis within the analysis time, the analysis program comprising;

by the processor,

a first generation process of generating first internal data on a basis of the first feature data and a first learning parameter;

a first conversion process of converting a position of the first feature data in a feature space on a basis of the first internal data generated in the first generation process and a second learning parameter;

a reallocation process of, based on a result of first conversion in the first conversion process and the first feature data, reallocating the first feature data to a position obtained through the conversion in the feature space;

a first calculation process of calculating a first predicted value of a hazard function about the analysis time in a case where the first feature data is given, based on a result of reallocation in the reallocation process and a third learning parameter;

an optimization process of optimizing the first learning parameter, the second learning parameter, and the third learning parameter by a statistical gradient method on a basis of the response variable and the first predicted value calculated in the first calculation process;

a second generation process of generating second internal data on a basis of second feature data including a plurality of feature amounts of a prediction target and the first learning parameter optimized in the optimization process;

a second conversion process of converting a position of the second feature data in the feature space on a basis of second internal data generated in the second generation process and the second learning parameter optimized in the optimization process; and

an importance calculation process of calculating importance data including an importance of each feature amount of the second feature data on a basis of a result of second conversion in the second conversion process and the third learning parameter optimized in the optimization process.