KR20230039902A - Method and apparatus for dynamic ultra-high-dimensional feature selection for time series prediction and causal factor analysis - Google Patents

Abstract

Disclosed are a method for selecting a dynamic ultra-high-dimensional feature for time series prediction and cause factor analysis and a device thereof. The method for selecting the dynamic ultra-high-dimensional feature performed by a computer device according to one embodiment comprises a step of selecting and learning a cause factor that is helpful in predicting a prediction value in the time series prediction, wherein the cause factor may comprise at least one or more among the meta-data directly connected to the prediction value, the time-series input data, and the input covariates. Therefore, the present invention is capable of reducing an input dimension.

Description

Dynamic hyper-dimensional feature selection method and apparatus for time series prediction and causal factor analysis

The following embodiments relate to a dynamic high-dimensional feature selection method and apparatus for time series prediction and causal factor analysis.

Time-series data refers to data expressed as a function of time over a period of time. Such time series data can be predicted through analysis of past time series data.

In predicting time series data, by clustering time series data generated during the past training period with time series data of similar patterns and determining a cluster corresponding to the prediction period, the time series data of the prediction period has a pattern similar to the determined cluster. A method of predicting what will be seen is provided.

For example, values of the form x(t) in which the value of x changes with time t are called time series data, and can be expressed as follows.

X = [x(0), x(1), x(2), ??]

Assuming that the current point in time is T, forecasting x(t) where t > T, with observations up to x(t) where t <= T, is called time-series forecasting.

In order to predict the future, information about the cause of the effect must be sufficiently modeled. It is impossible to reduce the uncertainty of future predictions without inputting causal factors. Therefore, it is necessary to input various types of time-series data, meta-data, and covariates that can be the cause (causative factor) of the phenomenon to be predicted. Here, covariates are values for which the future is known. For example, on a public holiday, you can know when it will be a public holiday in the future.

For example, in order to predict the future demand for fashion items of a certain fashion brand, when trying to enter information on all products sold by that fashion brand, the number of products actually on sale is 600,000 x the number of sales channels is 2,000 x sales 14 days = 5.6 billion levels of input dimensions are created. There is no way to model the joint relationships of input features on this scale all at once.

Korean Patent Registration No. 10-2134682 describes a system and method for generating a predictive model for such real-time time series data.

Korean Patent Registration No. 10-2134682

Embodiments describe a dynamic ultra-high-dimensional feature selection method and apparatus for time series prediction and causal factor analysis, and more specifically, a learning-based technology that selects information useful for predicting a problem to be solved in time series prediction. to provide.

Embodiments learn a combination of a plurality of input spatial features and a metric about how effective the combination of input spatial features is in predicting a predicted value, thereby providing an effective input for predicting each predicted value in a data-driven manner. An object of the present invention is to provide a dynamic super-high-dimensional feature selection method and apparatus for time series prediction and causal factor analysis capable of reducing input dimensions while minimizing loss of information about a problem to be solved by selecting spatial features.

A dynamic ultra-high-dimensional feature selection method performed by a computer device according to an embodiment includes the step of selecting and learning a causal factor that is helpful in predicting a predicted value in time series prediction, wherein the causal factor is the predicted value. It may include at least one or more of metadata directly connected to (meta-data), time-series input data, and input covariates.

The method may further include selecting and learning the causal factor and outputting a final prediction value through a time series prediction model.

In the step of selecting and learning the causal factor, a combination of a plurality of input spatial features and a metric for how effective the combination of the input spatial features is in predicting a prediction value are learned through an algorithm to achieve data driven (data driven) -driven) to select an effective feature of the input space for prediction of each prediction value.

The selecting and learning of the causal factor may include constructing an agent that proposes a combination of input spatial features; learning, by the agent, the combination of input spatial features and a metric about how effective the combination of input spatial features is in predicting a prediction value; and selecting, by the agent, a feature of the input space that is helpful in predicting a given instance prediction value based on the metadata as learning progresses.

The step of selecting and learning the causal factor may include constructing a decoder when the number of elements in the input spatial feature is not limited.

In the step of selecting and learning the causal factor, if the number of elements in the input space feature is limited, a module capable of making arbitrary feature instance proposals is used, and the number of feature instances as many as the corresponding number is selected. instance) may be included.

In the step of selecting and learning the causal factor, if based on reinforcement learning, an instance to be predicted may be fixed and a loss may be optimized.

In the step of outputting a final prediction value through the time series prediction model, the loss of the time series prediction model is applied based on the value derived during learning by selecting the causal factor, and end-to-end end) can be learned.

Outputting the final prediction value through the time-series prediction model may include additionally configuring an encoder when the number of elements in the input spatial feature is not limited.

In the step of outputting the final predicted value through the time series prediction model, since the shape of the input manifold is not consistent when the feature is changed for every predicted instance, after feature selection, the instance ) may include a step of converting so that the consistency of the manifold between them can occur.

An apparatus for dynamic high-dimensional feature selection according to another embodiment includes a feature selection unit that selects and learns a causal factor that is helpful in predicting a predicted value in time series prediction, wherein the causal factor is a meta directly connected to the predicted value. It may include at least one or more of data (meta-data), time-series input data, and input covariates.

After learning by selecting the causal factor, a time series prediction modeling unit outputting a final prediction value through a time series prediction model may be further included.

The feature selector learns, through an algorithm, a combination of a plurality of input spatial features and a metric about how effective the combination of the input spatial features is in predicting a predicted value, and makes data-driven every prediction. It is possible to select the valid input spatial feature for value prediction.

The feature selector configures an agent that proposes a combination of input spatial features, and the agent learns the combination of input spatial features and a metric about how effective the combination of input spatial features is in predicting a predicted value. And, as learning progresses, the agent that helps predict a given instance prediction value based on the metadata can select the input space feature.

The feature selector may configure a decoder when the number of elements in the input space feature is not limited.

According to the embodiments, a combination of a plurality of input spatial features and a metric for how effective a combination of input spatial features is in predicting a predicted value are learned to predict each predicted value in a data-driven manner. By selecting valid input space features, it is possible to provide a dynamic hyper-high-dimensional feature selection method and device for time series prediction and causal factor analysis that can reduce the input dimension while minimizing the loss of information about the problem to be solved.

1 is a diagram schematically illustrating a dynamic ultrahigh-dimensional feature selection method for time series prediction and causal factor analysis according to an embodiment.
2 is a flowchart illustrating a dynamic ultrahigh-dimensional feature selection method for time series prediction and causal factor analysis according to an embodiment.
3 is a block diagram illustrating a dynamic high-dimensional feature selection device for time series prediction and causal factor analysis according to an embodiment.

Hereinafter, embodiments will be described with reference to the accompanying drawings. However, the described embodiments may be modified in many different forms, and the scope of the present invention is not limited by the embodiments described below. In addition, several embodiments are provided to more completely explain the present invention to those skilled in the art. The shapes and sizes of elements in the drawings may be exaggerated for clarity.

In general, methods for reducing the dimensionality of input data include dimensionality reduction and feature selection.

Dimensionality reduction is a method of projecting high-dimensional data into low-dimensional data. For example, PCA is unsupervised feature selection, so it does not look at the target data.

Also, feature selection is a method of selecting only some data. For example, mutual information is easy to evaluate for greedy features (relationship between two variables X vs Y), but difficult to evaluate for joint relationships (X1, X2, …, Xn vs Y). That is, when the variable becomes large, there is a problem of having a level of complexity that is impossible to calculate.

The following embodiments attempt to reduce the dimensionality of input data while minimizing loss of information about a problem to be solved.

In order to reduce the input dimensionality while minimizing the loss of information about the problem to be solved, the exact problem to be solved must first be used (supervised). And, for each reasoning, you must be able to select and write the necessary information. Also, you should be able to see joint relationships as well as greedy relationships (bivariate). And, it should be able to be used for both continuous and discrete variables. Finally, it must be of computable complexity.

In addition, all predictions must be explainable for the cause of the prediction, and when the above "must be able to select and write the necessary information for each inference" is achievable, the information necessary for inference was naturally selected, so the cause of the inference ( causal factor) is achieved.

In this embodiment, we propose a learning-based technique for selecting information that is helpful in predicting a problem to be solved.

1 is a diagram schematically illustrating a dynamic ultrahigh-dimensional feature selection method for time series prediction and causal factor analysis according to an embodiment.

가 될 수 있다. 입력 차원이 커지면 분석적(analytic)으로 joint relationship을 찾는 것이 어려워지므로, 실시예에서는 경험적(empirical) 방법을 최대한 활용할 수 있다.Referring to FIG. 1, for example, the problem to be solved in the time series prediction problem is the variable to be predicted.

can be Since it becomes difficult to find a joint relationship analytically when the input dimension is large, in the embodiment, an empirical method can be maximally utilized.

First, according to an embodiment, an experiment setting may be performed as follows.

The predicted value can be expressed as the following equation.

[Equation 1]

이고, 예측 구간 길이는

(

터

지 예측)이다.Here, the prediction time is

, and the prediction interval length is

(

place

prediction).

와 직접적으로 연결된 메타데이터(meta data)를

으로 표현할 수 있다. 여기서,

를 명백하게(explicit) 설명할 수 있는 직접적으로 연관된 데이터이다.value to predict

Meta data directly linked to

can be expressed as here,

is directly related data that can explicitly describe

에 속하지 않는 입력 데이터는

where

이고,

에 속하지 않는 입력 공변량(covariates

where

,

이다.

Input data that does not belong to

where

ego,

Input covariates that do not belong to

where

,

am.

고 한다.Each of the data pools containing the above input data and covariates

say

이다.and,

am.

는 입력공간 특징(feature)들의 조합이고,

가

를 예측하는데 얼마나 유효한지에 대한 메트릭(metric)이다.here,

is a combination of input space features,

go

It is a metric for how effective it is in predicting .

The final prediction can be expressed as:

[Equation 2]

According to one embodiment, the final goal can be expressed as follows.

) 쌍을 관찰한 후, 이를 학습하여 데이터 드리븐(data-driven)으로 매

예측을 위한 유효한

를 골라내는 알고리즘

을 제공할 수 있으며, 다음과 같이 나타낼 수 있다. Countless

) pairs are observed, and then learned and matched in a data-driven manner.

valid for prediction

Algorithm to pick out

can be provided, and can be expressed as:

[Equation 3]

where

은 예측의 정확도를 측정할 수 있는 손실 함수(loss function)이다. 이를 통해 최선의 예측 정확도를 달성하는 것이 목표이다.Here, the condition of validity is,

where

is a loss function that can measure the accuracy of prediction. The goal is to achieve the best prediction accuracy through this.

를 구성하는 방법, 2)

의 구성을 제공하는 것이다.The detailed goals for each step are 1) Algorithm

How to configure, 2)

is to provide the configuration of

를 구성하는 방법에 대해 설명한다.Algorithm below

Describe how to configure.

를 구성하는 방법은 딥러닝 기반 학습을 통한 접근 방식으로 제공될 수 있다.algorithm

A method of configuring may be provided as an approach through deep learning-based learning.

를 제안하는 에이전트(agent)를 구성한다. first,

configures an agent that proposes

쌍들을 지속적으로 학습한다. the agent

Pairs are continuously learned.

을 기반으로 주어진 인스턴스(instance)

를 예측하는데 도움이 되는

를 골라낼 수 있어야 한다.As training progresses, metadata

given instance based on

to help predict

should be able to pick out

내의 요소(element) 개수에 제한을 두는 방법과 제한을 두지 않는 방법이 있다.in this process

There are methods that limit the number of elements in the element, and methods that do not limit the number of elements.

의 요소(element) 개수에 제한을 두지 않기 위해서는 디코더(decoder)가 필요하다. 현재 메타데이터를 기반으로, 임의의 개수만큼 요소(element)를 제안할 수 있는 디코더를 구성한다.

In order not to limit the number of elements of , a decoder is required. Based on the current metadata, a decoder capable of proposing an arbitrary number of elements is constructed.

의 요소(element) 개수에 제한을 둘 경우, 임의의 특징 인스턴스 제안(feature instance proposal)이 가능한 모듈을 써서 해당 개수만큼 특징 인스턴스(feature instance)를 출력하도록 한다.

If the number of elements of is limited, a module capable of random feature instance proposals is used to output as many feature instances as the corresponding number.

Here, when based on reinforcement learning, the instance to be predicted is fixed, and an agent capable of optimizing the loss can be trained within it. That is, the environment can be fixed.

For learning, any technique capable of minimizing loss can be used.

를 구성하는 방법은 학습 기반이 아닌 접근 방식으로 제공될 수도 있다. On the other hand, the algorithm

The method of constructing may be provided as a non-learning-based approach.

In greedy feature selection, the range of search is expanded by focusing on features with high scores first. In other words, you can start by removing meaningless features once.

의 구성에 대해 설명한다. below

The configuration of is described.

의 구성은 딥러닝 기반 학습을 통한 접근 방식으로 제공될 수 있다.

The configuration of can be provided by an approach through deep learning-based learning.

로 사용한다. 이 경우 전체 과정이 엔드 투 엔드(end-to-end)로 학습될 수 있다.First, by constructing a time-series forecasting model, the loss of the time-series forecasting model

use as In this case, the entire process can be learned end-to-end.

의 요소(element) 개수에 제한을 두지 않은 경우는 추가적으로 인코더(encoder)가 필요하다.in front

If the number of elements of is not limited, an additional encoder is required.

제안(proposal) 시에 각 요소(element)가 순열(permutation)이 되었을 때, 일반적인 딥러닝 모델들은 위치 불변(position-invariant)하지 않으므로 문제가 될 수 있다. 따라서, 시계열 예측 모델

의 아키텍처(architecture)는 위치와 관계없는 모델을 적용해야 한다. 예컨대, 각 key, value 간의 attention 기반 모델을 적용할 수 있다.

When each element is permuted at the time of proposal, it can be a problem because general deep learning models are not position-invariant. Therefore, the time series forecasting model

The architecture of should apply a position-independent model. For example, an attention-based model between each key and value can be applied.

If the feature is changed at every prediction instance, the shape of the input manifold is inconsistent, so general deep learning models cannot unfold the manifold. To this end, it may be necessary to add a function that transforms after feature selection so that manifold consistency between instances can occur.

의 구성은 학습 기반이 아닌 방식으로도 접근 가능하다.Meanwhile,

The composition of is also accessible in a non-learning-based way.

와 값이 유사한 그룹과

와 값이 유사하지 않은 그룹 간의 차이가 커지도록

를 구성한다. If not learning-based,

groups with values similar to

so that the difference between groups with dissimilar values of

make up

2 is a flowchart illustrating a dynamic ultrahigh-dimensional feature selection method for time series prediction and causal factor analysis according to an embodiment.

Referring to FIG. 2 , a method for selecting a dynamic ultra-high-dimensional feature performed by a computer device according to an embodiment includes selecting and learning a causal factor helpful for predicting a predicted value in time series prediction (S110). It can be done. Here, the causal factor may include at least one or more of metadata directly connected to the predicted value, time-series input data, and input covariates.

In addition, after learning by selecting a causal factor, outputting a final predicted value through a time series prediction model (S120) may be further included.

The dynamic ultra-high dimensional feature selection method for time series prediction and causal factor analysis according to an embodiment may be described in more detail by taking the dynamic ultra-high dimensional feature selection device for time series prediction and causal factor analysis according to an embodiment as an example. .

3 is a block diagram illustrating a dynamic high-dimensional feature selection device for time series prediction and causal factor analysis according to an embodiment.

Referring to FIG. 3 , an apparatus 300 for dynamic high-dimensional feature selection for time series prediction and causal factor analysis according to an embodiment may include a feature selection unit 310 and a time series prediction modeling unit 320 .

In step S110, the feature selector 310 may select and learn a causal factor that is helpful in predicting a predicted value in time series prediction. At this time, it can be learned through a deep learning method. Here, the causal factor may include at least one or more of metadata directly connected to the predicted value, time-series input data, and input covariates.

를 의미할 수 있다.The feature selector 310 learns a combination of a plurality of input spatial features and a metric about how effective the combination of input spatial features is in predicting a prediction value through an algorithm, and then performs data-driven It is possible to select valid input space features for prediction of predicted values. Here, the algorithm is the above-mentioned algorithm

can mean

을 제안하는 에이전트를 구성하고, 에이전트가 입력공간 특징들의 조합

및 입력공간 특징들의 조합이 예측 값을 예측하는데 얼마나 유효한지에 대한 메트릭(metric)

을 학습할 수 있다. 즉, 에이전트가

) 쌍들을 지속적으로 학습할 수 있다. 특징 선택부(310)는 학습이 진행됨에 따라, 메타데이터

를 기반으로 주어진 인스턴스(instance) 예측 값

을 예측하는데 도움이 되는 에이전트가 입력공간 특징을 선택할 수 있다. More specifically, the feature selection unit 310 is a combination of input space features

Construct an agent that proposes, and the agent is a combination of input space features

and a metric of how effective the combination of input space features is in predicting the predicted value.

can learn That is, the agent

) Pairs can be continuously learned. As learning progresses, the feature selection unit 310 provides metadata

predicted value for a given instance based on

An agent that helps predict the can select input space features.

내의 요소(element) 개수에 제한을 두지 않는 경우, 디코더(decoder)를 구성할 수 있다. 또한, 특징 선택부(310)는 입력공간 특징

내의 요소(element) 개수에 제한을 둘 경우, 임의의 특징 인스턴스 제안(feature instance proposal)이 가능한 모듈을 사용하여 해당 개수만큼 특징 인스턴스(feature instance)를 출력하도록 할 수 있다. Here, the feature selection unit 310 is a feature of the input space.

If there is no limit on the number of elements within, a decoder can be configured. In addition, the feature selection unit 310 is a feature of the input space

When the number of elements within the block is limited, a module capable of making arbitrary feature instance proposals can be used to output as many feature instances as the corresponding number.

On the other hand, if the feature selector 310 is based on reinforcement learning, it is possible to fix an instance to be predicted and optimize a loss.

를 구성하는 방법은 학습 기반이 아닌 접근 방식으로 제공될 수도 있다. greedy한 특징 선택(feature selection) 점수가 높은 특징들을 우선 중심으로 하여 탐색의 범위를 확장해 나간다. 즉, 의미 없는 특징들을 한번 제거해놓고 시작할 수 있다.On the other hand, the algorithm

The method of constructing may be provided as a non-learning-based approach. In greedy feature selection, the range of search is expanded by focusing on features with high scores first. In other words, you can start by removing meaningless features once.

In step S120, the time series prediction modeling unit 320 selects and learns a causal factor, and then outputs a final predicted value through a time series prediction model.

를 구성할 수 있으며, 딥러닝 기반 학습을 통한 접근 방식으로 구성할 수 있다. The time series prediction modeling unit 320 selects a causal factor, applies a loss of a time series prediction model based on a value derived during learning, and can learn end-to-end. in other words,

can be configured, and it can be configured with an approach through deep learning-based learning.

Here, the time series prediction modeling unit 320 may additionally configure an encoder when the number of elements in the input spatial feature is not limited.

Since the time series prediction modeling unit 320 is not consistent in the shape of the input manifold when the feature is changed for every prediction instance, the manifold between instances after feature selection can be transformed so that the consistency of

의 구성은 학습 기반이 아닌 방식으로도 접근 가능하다. 학습 기반이 아닌 경우,

와 값이 유사한 그룹과

와 값이 유사하지 않은 그룹 간의 차이가 커지도록

를 구성한다. Meanwhile,

The composition of is also accessible in a non-learning-based way. If not learning-based,

groups with values similar to

so that the difference between groups with dissimilar values of

make up

According to the embodiments, a combination of a plurality of input spatial features and a metric for how effective a combination of input spatial features is in predicting a predicted value are learned to predict each predicted value in a data-driven manner. By selecting valid input space features, the input dimensionality can be reduced while minimizing the loss of information about the problem to be solved.

The devices described above may be implemented as hardware components, software components, and/or a combination of hardware components and software components. For example, devices and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It may be implemented using one or more general purpose or special purpose computers, such as a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. A processing device may run an operating system (OS) and one or more software applications running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of software. For convenience of understanding, there are cases in which one processing device is used, but those skilled in the art will understand that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it can include. For example, a processing device may include a plurality of processors or a processor and a controller. Other processing configurations are also possible, such as parallel processors.

Software may include a computer program, code, instructions, or a combination of one or more of the foregoing, which configures a processing device to operate as desired or processes independently or collectively. You can command the device. Software and/or data may be any tangible machine, component, physical device, virtual equipment, computer storage medium or device, intended to be interpreted by or provide instructions or data to a processing device. can be embodied in Software may be distributed on networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer readable media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program commands recorded on the medium may be specially designed and configured for the embodiment or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. - includes hardware devices specially configured to store and execute program instructions, such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler.

As described above, although the embodiments have been described with limited examples and drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques may be performed in an order different from the method described, and/or components of the described system, structure, device, circuit, etc. may be combined or combined in a different form than the method described, or other components may be used. Or even if it is replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents of the claims are within the scope of the following claims.

Claims (15)

A dynamic ultra-high-dimensional feature selection method performed by a computer device, comprising:
The step of selecting and learning causal factors that are helpful in predicting the predicted value in time series prediction
including,
The causative factor is
Including at least one or more of metadata directly linked to the predicted value, time-series input data, and input covariates
Characterized by, feature selection method. According to claim 1,
After learning by selecting the causal factor, outputting a final predicted value through a time series prediction model
Further comprising a feature selection method. According to claim 1,
The step of selecting and learning the causative factor,
Through the algorithm, a combination of a plurality of input spatial features and a metric of how effective the combination of the input spatial features are in predicting a predicted value are learned to predict each predicted value in a data-driven manner. selecting a valid input space feature;
Including, feature selection method. According to claim 1,
The step of selecting and learning the causative factor,
constructing an agent that proposes a combination of input space features;
learning, by the agent, the combination of input spatial features and a metric about how effective the combination of input spatial features is in predicting a prediction value; and
As learning progresses, the agent selecting the input space feature that is helpful in predicting a given instance prediction value based on the metadata.
Including, feature selection method. According to claim 4,
The step of selecting and learning the causative factor,
Configuring a decoder when the number of elements in the input space feature is not limited
Including, feature selection method. According to claim 4,
The step of selecting and learning the causative factor,
When limiting the number of elements in the input space feature, outputting as many feature instances as the corresponding number using a module capable of making arbitrary feature instance proposals.
Including, feature selection method. According to claim 1,
The step of selecting and learning the causative factor,
When based on reinforcement learning, fixing the instance to be predicted and optimizing the loss
Characterized by, feature selection method. According to claim 2,
The step of outputting the final predicted value through the time series prediction model,
Applying the loss of the time series prediction model based on the value derived during learning by selecting the causal factor, and end-to-end learning
Characterized by, feature selection method. According to claim 2,
The step of outputting the final predicted value through the time series prediction model,
If the number of elements in the input space feature is not limited, additionally configuring an encoder.
Including, feature selection method. According to claim 2,
The step of outputting the final predicted value through the time series prediction model,
Since the shape of the input manifold is inconsistent when the feature is changed for every prediction instance, conversion is performed to ensure consistency of the manifold between instances after feature selection. step
Including, feature selection method. In the dynamic ultra-high-dimensional feature selection device,
A feature selection unit that selects and learns causal factors that are helpful in predicting predicted values in time series prediction.
including,
The causative factor is
Including at least one or more of metadata directly linked to the predicted value, time-series input data, and input covariates
Characterized in that, feature selection device. According to claim 11,
A time series prediction modeling unit that selects and learns the causal factor and outputs a final predicted value through a time series prediction model.
Further comprising a feature selection device. According to claim 11,
The feature selection unit,
Through the algorithm, a combination of a plurality of input spatial features and a metric of how effective the combination of the input spatial features are in predicting a predicted value are learned to predict each predicted value in a data-driven manner. selecting a valid said input space feature;
Characterized in that, feature selection device. According to claim 11,
The feature selection unit,
An agent that proposes a combination of input spatial features is configured, the agent learns the combination of the input spatial features and a metric about how effective the combination of the input spatial features is in predicting a predicted value, and the learning proceeds Selecting the input space feature by the agent to help predict a given instance prediction value based on the metadata, according to
Characterized in that, feature selection device. According to claim 14,
The feature selection unit,
Constructing a decoder when the number of elements in the input space feature is not limited
Characterized in that, feature selection device.

Images