KR102093789B1 - Explainable deep learning method and system - Google Patents

Abstract

A computer implemented method, comprising: (a) performing deep learning on a data set to obtain results regarding a given problem; (b) creating a decision tree comprising a plurality of classification rules for the given problem; (c) matching the classification result of the decision tree having the result most similar to the result through the deep learning to the result through the deep learning; And (d) acquiring a basis from which the corresponding result through the deep learning is reflected from the decision tree by reflecting the matching result.

Description

EXPLAINABLE DEEP LEARNING METHOD AND SYSTEM}

The present invention relates to a deep learning technology, and more particularly, to a deep learning method and system for securing explanatory power in a deep learning result by using a decision tree having a result value most similar to the deep learning result .

Deep learning is a neural network technology that is widely used in image classification, voice recognition, stock price prediction, and natural language chat bots. However, neural networks have weak explanation because the grounds for explaining the characteristics (ie, the decision rules for the corresponding result, which is the basis for providing the explanation) are scattered among feature-based hidden layers. There is a disadvantage (see FIG. 1).

In other words, deep learning provides accurate classification results for a given classification problem through neural network learning for a given data set (eg, convolutional neural network). At this time, deep learning has excellent learning results such as accuracy of classification, but it is often the case that the mathematical classifiers set for the classification divide the given problem into multiple boundaries and do not know the boundary when classifying the given data. Even if it is found, it is difficult to express in words because it is not a non-orthogonal boundary.

Therefore, there is a need for a new method and system capable of securing explanatory power in deep learning that actually shows excellent results in various problems.

The present invention is to provide a deep learning method and system that secures explanatory power in a deep learning result by using a decision tree having a result value most similar to the deep learning result.

According to an aspect of the present invention, a computer implemented method, comprising: (a) performing deep learning on a data set to obtain a result of a given problem; (b) creating a decision tree comprising a plurality of classification rules for the given problem; (c) matching the classification result of the decision tree having the result most similar to the result through the deep learning to the result through the deep learning; And (d) acquiring a basis from which the corresponding result through the deep learning is reflected from the decision tree by reflecting the matching result.

Here, step (b) may be a step of generating using a random forrest (Random Forrest) or a gradient boosting decision tree (Gradient boosting decision tree).

Here, in step (b), a plurality of decision trees are generated based on the data set, and step (c) has a result most similar to the result through the deep learning among the generated plurality of decision trees. Matching the result of the classification of the decision tree to the result through the deep learning,

In the step (d), the classification result related to the most similar classification result extracted from the decision tree may be used to obtain a basis for the corresponding result through the deep learning.

According to another aspect of the present invention, a deep learning system comprising: a deep learning performing unit to obtain a result of a given problem by performing deep learning on a data set; A decision tree generator for generating a decision tree including a plurality of classification rules for the given problem; The classification result of the decision tree having the result most similar to the result through the deep learning is matched to the result through the deep learning, and the matching result is reflected to indicate the basis on which the corresponding result through the deep learning appears. A deep learning system including a matching unit obtained from all trees is provided.

Here, the decision tree generation unit may be generated using a random forrest (Random Forrest) or a gradient boosting decision tree (Gradient boosting decision tree).

Here, the decision tree generation unit generates a plurality of decision trees based on the data set,

The matching unit matches the classification result of the decision tree having the most similar result to the result through the deep learning among the generated plurality of decision trees to the result through the deep learning, and is extracted from the decision tree The classification rule related to the most similar classification result may be used to obtain a basis for the corresponding result through the deep learning.

According to the deep learning method and system according to an embodiment of the present invention, by using a decision tree having a result value most similar to the deep learning result, it is possible to secure an explanatory power in the deep learning result.

1 is a reference diagram for explaining a problem of the existing deep learning technology.
2 is a view for explaining an explanatory deep learning method and system according to an embodiment of the present invention.
FIG. 3 is a schematic diagram for explaining the decision tree, FIG. 4 is a reference diagram for explaining the position and meaning of a node in the decision tree, FIG. 5 is an exemplary view of the decision tree, and FIG. 6 is a decision tree A diagram for describing a classification space divided by a tree.

The present invention can be applied to various transformations and can have various embodiments, and specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all conversions, equivalents, and substitutes included in the spirit and scope of the present invention.

In describing the present invention, when it is determined that a detailed description of related known technologies may unnecessarily obscure the subject matter of the present invention, the detailed description will be omitted. In addition, the numbers (for example, first, second, etc.) used in the description process of the present specification are only identification symbols for distinguishing one component from other components.

Further, in the specification, when one component is referred to as "connected" or "connected" with another component, the one component may be directly connected to the other component, or may be directly connected, but in particular It should be understood that, as long as there is no objection to the contrary, it may or may be connected via another component in the middle.

In addition, throughout the specification, when a part "includes" a certain component, this means that other components may be further included instead of excluding other components unless specifically stated otherwise. In addition, terms such as “part” and “module” described in the specification mean a unit that processes at least one function or operation, which means that it can be implemented by one or more hardware or software or a combination of hardware and software. .

The present invention utilizes the results of a random forrest or a gradient boosting decision tree that generates a highly descriptive decision tree as an existing data mining technique in order to find an answer to a given specific problem. Thus, by matching the result of the decision tree that is most similar to the deep learning result, the best result presented in deep learning is visualized to secure the explanatory power of the result from deep learning from the above-described decision tree technology.

According to this, it was difficult to provide a rationale for the result because of the hidden layer, which is a characteristic of the neural network, but a problem can be solved by creating a decision tree that generates explanatory rules. The method of generating a representative decision tree, such as random forrest (Random Forrest) or gradient boosting decision tree (Gradient boosting decision tree), and the results of deep learning are presented in connection with various results that want to explain the results after applying deep learning. It is expected that the user's reliability of the results from the problems can be significantly increased.

In this regard, embodiments of the present invention will be described in detail with reference to the accompanying drawings. Here, FIG. 2 is a diagram for generally explaining an explanatory deep learning method and system according to an embodiment of the present invention. Also, FIG. 3 is a schematic diagram for explaining the decision tree, FIG. 4 is a reference diagram for explaining the position and meaning of a node in the decision tree, FIG. 5 is an exemplary view of the decision tree, and FIG. 6 is a decision tree It is a diagram for explaining the classification space divided by the entire tree.

Referring to FIG. 2, an explanatory deep learning method and system according to an embodiment of the present invention obtains a result of a given problem by performing deep learning on a data set, and Decision trees including a plurality of classification rules related to a given problem are generated, and among the generated plurality of decision trees, the classification result of the decision tree having the most similar result to the result through the deep learning May be implemented in a manner of matching the result through the deep learning and obtaining the basis for the result through the deep learning by using the classification rule related to the most similar classification result extracted from the decision tree. .

That is, in the explanatory deep learning method in the present invention, first, deep learning (eg, CNN, etc.) with high accuracy is first performed to determine a given problem. Subsequently, in order to secure explanatory power for the deep learning result, a decision tree that secures the boundary of classification is generated. At this time, the decision tree may be generated using a random forrest or a gradient boosting decision tree.

3, the decision tree classifies a given problem as a set of questions and answers, as shown through FIG. 3, and a tree structure of questions and answers (that is, a multi-layer IF-THEN structure from upper to lower nodes). (See 4), it has the advantage of creating a lot of orthogonal borders through which explanatory power can be secured. Accordingly, the classification rules of the decision tree for generating the orthogonal boundary line provide a clear explanatory power to classify the given data as illustrated in FIG. 3, and thus the disadvantage of deep learning that is difficult to secure the explanatory power by the hidden layer. Can complement.

As a brief example of this, FIG. 5 shows a decision tree for classifying diabetes risk groups, and FIG. 5 is an example of a classification tree for classifying diabetes risk groups, in which three classification rules are composed of IF-THEN statements. Showing. The applicable classification rules applied in this example are as follows.

Rule A: If (IF) age> 40 (THEN) is dangerous for diabetes.

Rule B: (IF) If the value of glycated hemoglobin is less than or equal to 120 (THEN), it is dangerous for diabetes.

Rule C: If (IF) urine foams (THEN), it is dangerous for diabetes.

As in the above-described example, by using classification rules of a tree structure constituting each IF-THEN statement, it is possible to construct initial root nodes with high explanatory power.

At this time, each node in the decision tree is divided into spaces (that is, classification of data based on a classification boundary) by each node as shown in FIG. 6. Referring to FIG. 6, the root node in a state where no branching occurs is A (the variable is three, so a cube in 3D space is assumed to be A as in the top figure), and A is determined by B by some criteria. Assuming that it is divided into C, you can see that A is divided into two subsets as shown in the second picture. Finally, if B is divided into D and C, as shown in the figure at the bottom, the terminal nodes are 3 of C, D, and E. If you think about this in relation to the data space, you can see that the entire data A is divided into three subsets. .

Described using the rules of the upper node of the decision tree according to the above, explanation power for most situations is provided, and even when very detailed rules are required, it can be explained by using the rules of the lower nodes of the decision tree. do.

Therefore, according to an embodiment of the present invention, it is possible to solve two problems that have been an issue in the existing deep learning. The first of the two issues is a legal issue, although it is true that deep learning is an algorithm with high accuracy, but it lacks the explanatory power of the algorithm. Due to the legal problems could arise. The second issue is service convenience, so it is necessary to be able to explain why the problem was classified as A when providing the service as a solution to the given classification problem. . However, according to an embodiment of the present invention, by utilizing a decision tree that generates explanatory rules in connection with the results of deep learning, the user reliability of the results is significantly improved in various problems that want to explain the results after applying deep learning. It has the effect of increasing.

The deep learning method according to the above-described embodiment of the present invention may be implemented as computer readable codes on a computer readable recording medium. The computer-readable recording medium includes any kind of recording medium storing data that can be read by a computer system. For example, there may be a read only memory (ROM), a random access memory (RAM), a magnetic tape, a magnetic disk, a flash memory, and an optical data storage device. In addition, the computer-readable recording medium may be distributed over computer systems connected by a computer communication network, and stored and executed as code readable in a distributed manner.

Although described above with reference to embodiments of the present invention, those skilled in the art variously modify the present invention without departing from the spirit and scope of the present invention as set forth in the claims below. And it can be easily understood.

Claims (6)

As a computer implemented method,
(a) performing deep learning on a data set to obtain a result of a given problem;
(b) creating a decision tree comprising a plurality of classification rules for the given problem;
(c) matching the classification result of the decision tree having the result most similar to the result through the deep learning to the result through the deep learning; And
(d) reflecting the matching result, obtaining a basis for the corresponding result through the deep learning from the decision tree
Computer implemented method comprising a.
According to claim 1,
The step (b), characterized in that generated using a random forrest (Random Forrest) or a gradient boosting decision tree (Gradient boosting decision tree), computer-implemented method.
According to claim 1,
In step (b), a plurality of decision trees are generated based on the data set,
The step (c) is a step of matching the classification result of the decision tree having the result most similar to the result through the deep learning among the generated plurality of decision trees to the result through the deep learning,
In step (d), the computer-implemented method is characterized in that a basis for obtaining the corresponding result through the deep learning is obtained by using a classification rule related to the most similar classification result extracted from the decision tree.
As a deep learning system,
A deep learning unit performing deep learning on a data set to obtain a result of a given problem; A decision tree generator for generating a decision tree including a plurality of classification rules for the given problem; The classification result of the decision tree having the result most similar to the result through the deep learning is matched to the result through the deep learning, and the matching result is reflected to indicate the basis on which the corresponding result through the deep learning appears. Deep learning system including a matching unit obtained from all trees.
According to claim 4,
The decision tree generation unit, characterized in that generated by using a random forest (Random Forrest) or a gradient boosting decision tree (Gradient boosting decision tree), deep learning system.
According to claim 4,
The decision tree generation unit generates a plurality of decision trees based on the data set,
The matching unit matches the classification result of the decision tree having the most similar result to the result through the deep learning among the generated plurality of decision trees to the result through the deep learning, and is extracted from the decision tree Deep learning system, characterized in that to obtain the basis for the corresponding result through the deep learning by using the classification rule associated with the most similar classification results.

Images