WO2021075742A1 - Procédé et appareil d'évaluation de valeur basée sur un apprentissage profond - Google Patents

Procédé et appareil d'évaluation de valeur basée sur un apprentissage profond Download PDF

Info

Publication number
WO2021075742A1
WO2021075742A1 PCT/KR2020/012694 KR2020012694W WO2021075742A1 WO 2021075742 A1 WO2021075742 A1 WO 2021075742A1 KR 2020012694 W KR2020012694 W KR 2020012694W WO 2021075742 A1 WO2021075742 A1 WO 2021075742A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
evaluation
deep learning
discriminators
score
Prior art date
Application number
PCT/KR2020/012694
Other languages
English (en)
Korean (ko)
Inventor
박훈
박현우
이종택
Original Assignee
한국과학기술정보연구원
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 한국과학기술정보연구원 filed Critical 한국과학기술정보연구원
Publication of WO2021075742A1 publication Critical patent/WO2021075742A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0637—Strategic management or analysis, e.g. setting a goal or target of an organisation; Planning actions based on goals; Analysis or evaluation of effectiveness of goals
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/10—Office automation; Time management

Definitions

  • the present disclosure relates to a deep learning-based value evaluation method and an apparatus therefor.
  • a method for generating virtual evaluation basic data using a deep learning model based on a Generative Adversarial Network (GAN), and evaluating the value of an evaluation target using the generated evaluation basic data, and an apparatus for performing the method It is about.
  • GAN Generative Adversarial Network
  • the cash flow discount model is a model in which the subject of commercialization uses a product or service to which technology is grafted to discount and evaluate the economic added value that can be created in the future with the present value. It is evaluated as.
  • the cash flow discount model requires valuation basic data such as the economic life of the technology, free cash flow, discount rate, and technology contribution.
  • actual data of the company e.g., financial data
  • Need it is very difficult to secure the actual data of a company in a sufficient volume enough to calculate the evaluation value distribution of the evaluation target. For this reason, it is practically not easy to properly use the cash flow discount model in the field of valuation with the existing method.
  • the technical problem to be solved by the present disclosure is to provide a value evaluation method capable of securing a reliable and sufficient volume of evaluation basic data using artificial intelligence based on deep learning and an apparatus for performing the method.
  • Another technical problem to be solved by the present disclosure is to provide a value evaluation method with improved user convenience and an apparatus for performing the method, since a value evaluation result can be obtained only by inputting a simple evaluation item.
  • a value evaluation method includes: generating a virtual data sample using a generator; Determining the generated data sample using a plurality of discriminators; Determining a policy gradient according to the determination result by the plurality of discriminators; And learning the generator or the plurality of discriminators by using the determined policy gradient.
  • the generator may include an artificial neural network including an input layer, an output layer, and a hidden layer, and may receive random noise through the input layer and output the data sample through the output layer.
  • the data sample is a data sample corresponding to classification information
  • the generator may generate the data sample by referring to the classification information
  • each of the plurality of discriminators includes an artificial neural network including an input layer, an output layer, and a hidden layer, and may be configured to have an artificial neural network structure different from that of another discriminator among the plurality of discriminators.
  • the different artificial neural network structures may be configured by differentiating the number of hidden layers, the number of units, or the types of activation functions.
  • the determining of a policy gradient according to the determination result may include determining whether or not a predetermined number or more of the plurality of discriminators has made the same or similar determination; And determining the policy gradient to reduce the weight of the data sample according to the determination result.
  • the determining of a policy gradient according to the determination result may include determining whether or not a predetermined number or more of the plurality of discriminators has made the same or similar determination; And determining the policy gradient to increase the weight of the data sample according to the determination result.
  • a value evaluation method includes: obtaining classification information of an evaluation object; Generating a data set including a virtual data sample corresponding to the classification information using a deep learning model; Calculating basic evaluation data using the generated data set; And calculating an evaluation result by applying the calculated evaluation basic data to a value evaluation model.
  • the evaluation basic data may include discount rate data, technology life data, technology contribution data, and free cash flow data for the evaluation target.
  • the input information is a core factor score including at least one of a technical score, a right score, a market score, and a business score for the evaluation target, the classification information And information on the size of the company to be evaluated, and the discount rate data may be defined as a function of the core factor score, the classification information, and the company size information.
  • the input information is a core factor score including at least one of a technical score, a right score, a market score, and a business score for the evaluation target, and the evaluation It includes IPC information of the subject, and the technical life data may be defined as a function of the core factor score and the IPC information.
  • the input information is a core factor score including at least one of a technical score, a right score, a market score, and a business score for the evaluation target, and the evaluation It includes classification information of an object, and the technology contribution data may be defined as a function of the core factor score and the classification information.
  • the free cash flow data may be defined as a function of financial item data included in the virtual data sample.
  • the calculating of the evaluation result may include processing a distribution of result values calculated by applying the evaluation basic data to the value evaluation model in the form of data or a graph.
  • a value evaluation apparatus for solving the above-described technical problem includes: a processor; A memory for loading a computer program executed by the processor; And a storage storing the computer program, wherein the computer program includes an operation of generating a virtual data sample using a generator, an operation of determining the generated data sample using a plurality of discriminators, and the plurality of determinations And instructions for determining a policy gradient according to a result of the determination by a group and an operation for learning the generator or the plurality of discriminators using the determined policy gradient.
  • a computer program for executing a value evaluation method for solving the above-described technical problem includes the steps of: storing in a computer-readable recording medium and generating a virtual data sample using a generator; Determining the generated data sample using a plurality of discriminators; Determining a policy gradient according to the determination result by the plurality of discriminators; And learning the generator or the plurality of discriminators by using the determined policy gradient.
  • FIG. 1 is a schematic diagram schematically showing a value evaluation method according to an embodiment of the present disclosure.
  • FIG. 2 is a block diagram illustrating a value evaluation system according to an embodiment of the present disclosure.
  • FIG. 3 is a block diagram illustrating a value evaluation apparatus according to an embodiment of the present disclosure.
  • FIG. 4 is an exemplary diagram illustrating an example of data provided to calculate basic evaluation data that may be referred to in some embodiments of the present disclosure.
  • FIG. 5 is a block diagram illustrating an exemplary configuration of the virtual data generating unit 140 of FIG. 1.
  • FIG. 6 is a diagram illustrating a structure of a deep learning model of the virtual data generator 140.
  • FIG. 7 and 8 are flowcharts illustrating a method of generating and learning virtual data by the value evaluation apparatus 100 according to some embodiments of the present disclosure.
  • FIG. 9 is a diagram illustrating a value evaluation method according to some embodiments of the present disclosure and a performance of the apparatus 100 in comparison with the prior art.
  • FIG. 10 is a flowchart illustrating a method of calculating a value evaluation result of an evaluation target using virtual data according to some embodiments of the present disclosure.
  • 11 is a diagram for describing a method of calculating evaluation basic data using virtual data and input information in some embodiments of the present disclosure.
  • FIG. 12 is a diagram for explaining a method of calculating an evaluation value distribution by applying the calculated evaluation basic data to a cash flow discount model.
  • FIG. 13 is a hardware configuration diagram illustrating an exemplary computing device capable of implementing a device according to various embodiments of the present disclosure.
  • first, second, A, B, (a) and (b) may be used. These terms are for distinguishing the constituent element from other constituent elements, and the nature, order, or order of the constituent element is not limited by the term.
  • an evaluation object literally means an object of value evaluation.
  • the evaluation target may include anything that can give a quantitative value, such as technology, company, and assets.
  • some embodiments of the present invention will be described on the assumption that the evaluation target is “technology”.
  • evaluation basic data refers to data that is the basis for value evaluation.
  • the evaluation basic data is data processed to apply to a valuation model in relation to the evaluation target (e.g., in the case of a cash flow discount model, technology life data, free cash flow data, discount rate data, technology contribution data) Etc.) and the like.
  • classification information of an evaluation object means class information defined according to a predetermined classification criterion.
  • the classification information may be preferably defined based on standard criteria.
  • the classification information may be a KSIC code defined according to the Korea Standard Industrial Classification (KSIC) classification criteria, and other IPC codes defined according to the International Patent Classification (IPC), and SIC (Standard Industrial Classification) SIC codes defined according to the classification criteria may also be the classification information.
  • KSIC Korean Standard Industrial Classification
  • IPC International Patent Classification
  • SIC Standard Industrial Classification
  • an instruction refers to a set of instructions grouped on the basis of a function, which is a component of a computer program and is executed by a processor.
  • FIG. 1 is a schematic diagram schematically showing a value evaluation method according to an embodiment of the present disclosure.
  • the valuation method according to the present disclosure obtains the evaluation basic data (1) for the evaluation target and applies it to the cash flow discount model (2), and evaluates according to the formula of the model (2). Calculate the estimated value distribution (5) of the object.
  • the evaluation basic data is basic data provided as an input value to the cash flow discount model 2, and is obtained according to a predetermined calculation formula based on a data set using deep learning and an input variable input by a user.
  • the cash flow discount model (2) is a present value measurement model obtained by discounting the expected cash flow in the future at an appropriate discount rate.For example, discounted cash flow is evaluated by the most objective method among valuation calculation methods. , DCF) method may be the cash flow discount model (2).
  • the evaluation value distribution (3) is a result value output by applying the evaluation basic data (1) to the cash flow discount model (2), and since the evaluation basic data is composed of a series of data sets, the result value is also It comes out as a series of output values, and the evaluation value distribution (3) is the display of the distribution of the series of output values.
  • the evaluation value distribution 3 is provided to the user as a result value, and through this, the user determines the estimated value of the evaluation object.
  • the evaluation basic data (1), the cash flow discount model (2), and the evaluation value distribution (3) will be described in more detail later in FIG. 11 and later.
  • FIG. 2 is a block diagram illustrating a value evaluation system according to an embodiment of the present disclosure.
  • the value evaluation system 1000 is illustrated as an example of providing a value evaluation service to a plurality of user terminals 200.
  • the value evaluation system 1000 includes a value evaluation apparatus 100, a user terminal 200, and a network 10 connecting the value evaluation apparatus 100 and the user terminal 200.
  • the valuation system may further include an external DB system that stores and manages various data used for valuation.
  • each of the components of the value evaluation system illustrated in FIG. 2 represents functional elements that are functionally divided, and at least one component may be implemented in a form in which they are integrated with each other in an actual physical environment.
  • the value evaluation device 100 is a computing device that performs a value evaluation on an evaluation object.
  • the computing device may be a notebook, a desktop, a laptop, etc., but is not limited thereto, and may include all types of devices equipped with a computing function and a communication function.
  • the computing device may preferably be implemented as a high-performance server computing device.
  • the value evaluation apparatus 100 provides a value evaluation service in response to an evaluation request from the user terminal 200.
  • the value evaluation apparatus 100 may provide the value evaluation service through a web-based user interface or an APP-based user interface.
  • the value evaluation device 100 is an input variable (eg, KSIC, IPC, company size, technology score, rights score, marketability) from the user terminal 200 through the user interface. Score, business score, etc.) are input, and evaluation basic data is obtained based on the input variable. Then, the obtained evaluation basic data is applied to a value evaluation model (for example, a cash flow discount model) to calculate the evaluation value distribution (3) of the evaluation object as a value evaluation result.
  • a value evaluation model for example, a cash flow discount model
  • the calculated value evaluation result is provided to the user terminal 200 through the user interface.
  • the user can easily obtain a value evaluation result by inputting only the input variable to be evaluated, even if the user does not have expert knowledge about the value evaluation model. Accordingly, not only the user satisfaction with the value evaluation service can be improved, but the overall service use environment of the value evaluation service can be improved.
  • the value evaluation apparatus 100 automatically generates a data set including a virtual data sample related to classification information of the evaluation target by using a deep learning model. And, based on this, the evaluation basic data may be calculated. According to this, it is possible to solve the conventional problem in which it was difficult to secure enough data samples to calculate the evaluation value distribution.
  • the value evaluation apparatus 100 proposes a new deep learning method with significantly improved reliability and efficiency than a general deep learning method in a method of generating a virtual data sample. According to the above embodiment, since more reliable (ie, similar to real data) virtual data samples are generated by the improved deep learning method, the reliability and accuracy of evaluation basic data calculated therefrom can also be guaranteed. A detailed description of the new deep learning method will be described later with reference to the accompanying drawings in FIG. 5.
  • the user terminal 200 communicates with the valuation device 100 and provides an appropriate interface to the user so that the valuation service provided by the valuation device 100 can be used.
  • the user terminal 200 may include a mobile terminal such as a mobile phone or a smart phone, a laptop, a desktop, and a computing device such as a workstation, but is not limited thereto.
  • the network 10 relays communication between the value evaluation apparatus 100 and the user terminal 200.
  • the network 10 is a local area network (LAN), a wide area network (WAN), a mobile radio communication network (mobile radio communication network), various types of wired / wireless such as Wibro (Wireless Broadband Internet). It may be configured as a network and a combination thereof, but is not limited thereto.
  • a user can obtain a value evaluation result for an evaluation target only by inputting a simple evaluation item corresponding to an input variable. Accordingly, the user's convenience of using the valuation service is improved, and the valuation result for the evaluation target can be easily obtained even without specialized knowledge of a complex valuation model.
  • the value evaluation device 100 since the value evaluation device 100 generates a reliable and sufficient volume of virtual data sample using artificial intelligence based on deep learning and secures evaluation basic data based on this, it is a conventional problem that it is difficult to secure evaluation basic data. Can be solved easily. Furthermore, since an improved deep learning method is proposed in generating a virtual data sample, the reliability and accuracy of the generated virtual data sample can be greatly improved.
  • the value evaluation apparatus 100 includes an interface unit 110, an evaluation data calculation unit 120, a virtual data generation unit 130, a value evaluation unit 130, and an actual data DB 150. Includes. However, only the components related to the embodiment of the present invention are shown in FIG. 3. Accordingly, those of ordinary skill in the art to which the present invention pertains can recognize that other general-purpose components may be further included in addition to the components illustrated in FIG. 3. In addition, since the embodiment shown in FIG. 3 shows an exemplary configuration of the value evaluation apparatus 100, some of the components shown in FIG. 3 may be omitted in other embodiments.
  • each of the components of the value evaluation apparatus 100 shown in FIG. 3 represents functional elements that are functionally divided, and at least one component may be implemented in a form in which they are integrated with each other in an actual physical environment. do.
  • the interface unit 110 provides an interface for communicating with an external device or providing an input/output means to a user.
  • the interface unit 110 receives input information from the user or the user terminal 200 and provides a result of processing the input information.
  • the received input information may include an input variable used to calculate basic evaluation data.
  • the input variable may include KSIC, IPC, company size (eg, large, medium, small, start-up company, etc.), technology score, rights score, marketability score, or business score.
  • the technology score, rights score, marketability score, or business score is obtained by evaluating the technology, rights, marketability, and business value of the evaluation target, respectively.
  • the technical score, right score, market score, or business score may be grouped into key factor scores.
  • the processing result provided by the interface unit 110 may include a value evaluation result for an evaluation object.
  • the value evaluation result may be provided in the form of an evaluation value distribution for an evaluation object.
  • the evaluation data calculation unit 120 calculates evaluation basic data necessary for value evaluation by using an input variable, a virtual data sample, or an actual data sample.
  • the evaluation basic data calculated by the evaluation data calculation unit 120 may include a discount rate (r), a technical life (n), a free cash flow (FCF), and a technology contribution (T.F.).
  • each item calculated by the evaluation data calculation unit 120 may be defined as a function of the input variable, the virtual data sample, or individual items of the actual data sample.
  • the function may be variously defined according to the field and situation to which this valuation method is applied. That is, since factors and weights to be considered when evaluating value may be different for each individual field and situation, the function may be variously adjusted according to the situation.
  • FCF free cash flow
  • the free cash flow may be defined as a function of the financial data item of the company in the field to which the evaluation target belongs.
  • the free cash flow (FCF) may be defined as a function such as Equation 1 below.
  • z1 is operating income after tax
  • z2 is depreciation
  • z3 is capital expenditure
  • z4 is net working capital increase or decrease
  • z5 is investment recovery.
  • free cash flow is defined as a function that takes an entity's financial data item as an input.
  • the evaluation data calculation unit 120 uses the virtual financial data sample provided by the virtual data generating unit 140 or the actual financial data sample provided by the real data DB 150 to determine the free cash flow (FCF). Will yield.
  • a specific example of a virtual financial data sample or an actual financial data sample used by the evaluation data calculation unit 120 is illustrated in FIG. 4.
  • values 22 of input items for calculating the free cash flow are presented.
  • the'code' area is an area in which a classification code according to the Korean Standard Industrial Classification (KSIC) is described, and is a field included in actual financial data, but may not be included in virtual financial data. .
  • KSIC Korean Standard Industrial Classification
  • a code area may not be necessary, and thus code information may not be included in the data sample.
  • the value evaluation unit 130 evaluates the value of the evaluation target by applying the evaluation basic data calculated by the evaluation data calculation unit 120 to the value evaluation model.
  • the valuation model may be a cash flow discount model.
  • the value evaluation unit 130 may provide a distribution of a series of evaluation values calculated using the value evaluation model in the form of data or graphs.
  • the value evaluation model may be a machine learning model built through machine learning such as deep learning.
  • machine learning such as deep learning.
  • the valuation device 100 provides a valuation service according to a cash flow discount model for a certain period of time and accumulates cumulative data consisting of classification information and valuation results of an evaluation target, machine learning based on the accumulated data is performed. Through this, you can build your own valuation model.
  • a valuation model based on a machine learning model with higher reliability than a typical cash flow discount model may be constructed.
  • a valuation service may be provided by mixing a machine learning-based first valuation model and a second valuation model according to a cash flow discount model.
  • the valuation unit 153 provides a valuation service according to the second valuation model when a certain condition (for example, when the learning maturity/accuracy of the second valuation model is greater than or equal to the threshold value) is satisfied, and the opposite
  • a hybrid value evaluation model may be constructed in a form of providing a value evaluation service according to the first value evaluation model.
  • the value evaluation unit 150 provides a final value evaluation result for the evaluation object according to the sum of the weights of the first value evaluation value of the first value evaluation model and the second value evaluation value of the second value evaluation model. You can build a hybrid valuation model in the form of.
  • the result calculated or evaluated by the value evaluation unit 130 may be provided to the user terminal 200 as a value evaluation result.
  • the virtual data generator 120 receives classification information and generates a virtual data set corresponding to the classification information using a deep learning model.
  • the deep learning model may be a model based on a Generative Adversarial Networks (GAN) algorithm.
  • GAN Generative Adversarial Networks
  • the virtual data set generated by the virtual data generator 120 is composed of virtual data samples generated by simulating the distribution of real data. That is, the virtual data is data generated by a deep learning model based on actually collected data.
  • the classification information received by the virtual data generator 120 may be classification information according to KSIC.
  • the virtual data generation unit 120 receives classification information according to KSIC, and generates a virtual data set that simulates actual data of companies belonging to the industrial field indicated by the classification information.
  • the generated data set is provided to the evaluation data calculation unit 120 and is used to calculate basic evaluation data.
  • a specific method for the virtual data generation unit 120 to learn based on a deep learning model and to generate a virtual data set will be described in detail later in FIG. 5 or less.
  • the real data DB 150 is a storage storing real data for providing to the virtual data generating unit 140 or the evaluation data calculating unit 120.
  • the actual data is data composed of an actual data sample, and may be data collected in association with classification information such as KSIC.
  • the real data DB 150 may provide the stored real data for deep learning of the virtual data generator 140.
  • the actual data DB 150 may provide the stored actual data to the evaluation data calculation unit 120 to calculate basic evaluation data.
  • the evaluation data calculation unit 120 may calculate basic evaluation data using real data provided by the real data DB 150 together with the virtual data set generated by the virtual data generation unit 140.
  • each component of FIG. 3 may mean software or hardware such as a Field Programmable Gate Array (FPGA) or an Application-Specific Integrated Circuit (ASIC).
  • FPGA Field Programmable Gate Array
  • ASIC Application-Specific Integrated Circuit
  • the components are not limited to software or hardware, and may be configured to be in an addressable storage medium, or may be configured to execute one or more processors.
  • the functions provided in the components may be implemented by more subdivided components, or may be implemented as one component that performs a specific function by combining a plurality of components.
  • FIG. 5 is a block diagram illustrating an exemplary configuration of the virtual data generating unit 140 of FIG. 1.
  • the virtual data generation unit 140 performs deep learning learning to generate virtual data based on the GAN algorithm, and generates a virtual data set corresponding to the input classification information using the learned deep learning model.
  • the GAN consists of a generator that generates a virtual data sample and a discriminator that determines whether the input data sample is real data, and through adversarial training between the generator and the discriminator. It refers to the machine learning model being built.
  • a new GAN model that exhibits more improved performance is proposed as a deep learning model for generating virtual data without using a conventional general GAN model.
  • the deep learning model includes a generator 141, a plurality of discriminators 142a, 142b, and 142n, and a reinforcement learning performing unit 143.
  • the deep learning model is characterized by learning through hostile training between a plurality of discriminators 142a, 142b, 142n and one generator 141, unlike a general GAN model.
  • the generator 141 is a component that generates virtual data that simulates real data by a deep learning model.
  • the generator 141 receives classification information and random noise, and generates a virtual data sample corresponding to the classification information by its own artificial neural network.
  • the virtual data samples generated by the generator 141 are provided to the plurality of discriminators 142a, 142b, and 142n.
  • the plurality of discriminators 142a, 142b, 142n is a component that receives the virtual data sample generated by the generator 141 and determines in parallel whether the received data is real data or virtual data.
  • the virtual data sample and actual data 30 generated by the generator 141 are input to each of the plurality of discriminators 142a, 142b, 142n, and each of the plurality of discriminators 142a, 142b, 142n is input data. Is determined whether is virtual data or real data.
  • each of the plurality of discriminators 142a, 142b, and 142n outputs a confidence score for whether the input data sample corresponds to actual data, and the input data sample is It can be made to determine whether it is real data.
  • each of the plurality of discriminators 142a, 142b, 142n is learned through real data 30 or virtual data samples.
  • learning about the plurality of discriminators 142a, 142b, 142n is preceded by the initial learning. After sufficiently performed, learning about the generator 31 may be performed. If the error calculated during the inaccurate discrimination process at the beginning of learning is backpropagated, it may adversely affect the training of the generator 141.In order to avoid this, learning is performed centering on the plurality of discriminators 142a, 142b, 142n at the beginning of learning to avoid this. You can do it.
  • learning about the generator 141 and the plurality of discriminators (142a, 142b, 142n) is performed alternately, and the learning of the plurality of discriminators (142a, 142b, 142n) is repeatedly performed more than a specified number of times.
  • training for the generator 141 may be performed in a form that is performed less than a specified number of times.
  • the plurality of discriminators 142a, 142b, and 142n may be configured to have different structures.
  • each of the plurality of discriminators 142a, 142b, 142n may be configured with each neural network such that the number of hidden layers, the number of units, or the type of activation function is different. have.
  • the reinforcement learning performing unit 143 receives a result determined by the plurality of discriminators 142a, 142b, 142n, determines an error, and generates a policy gradient according to the result, a generator 141 and a plurality of discriminators 142a, 142b, 142n) so that the generator 141 and the plurality of discriminators 142a, 142b, 142n perform deep learning learning.
  • the reinforcement learning performing unit 143 reduces the value of the sample (ie, decreases the weight), and the plurality of discriminators (
  • the policy gradient is determined to increase the value of the sample (i.e., increase the weight), and the generator 141 and the plurality of discriminators 142a, 142b, 142n). This is to learn by giving more weight to the data samples that appear to have been incorrectly judged so that they can learn and classify them better by focusing more on the data samples that are judged incorrectly.
  • the reinforcement learning performing unit 143 outputs whether the plurality of discriminators 142a, 142b, and 142n make similar or different judgments, and the plurality of discriminators 142a, 142b, 142n It can be judged based on the standard deviation of the data. For example, if the plurality of discriminators 142a, 142b, 142n outputs the result of determining whether the sample is real or virtual data as a value between 0 and 1 (in this case, the determination is virtual data).
  • the reinforcement learning execution unit 143 considers that the standard deviation of the data output from the plurality of discriminators 142a, 142b, and 142n is 0.3 or more, is deemed to have made different judgments, and the corresponding sample A policy gradient may be determined to increase the weight of and, if the standard deviation is less than 0.3, the policy gradient may be determined to decrease the weight of the corresponding sample, as deemed similar to each other.
  • a plurality of discriminators determine for the generator 141, and determine whether the discrimination results are similar to each other, by combining a plurality of discriminators.
  • the effect of constructing a strong discriminator can be exerted.
  • the generator 141 After the generator 141 is sufficiently learned by the deep learning method described above, the generator 141 receives classification information and generates a virtual data set corresponding thereto. It is provided to the evaluation data calculation unit 120.
  • FIG. 6 is a diagram illustrating a structure of a deep learning model of the virtual data generator 140.
  • the deep learning structure of the generator 141 and the first discriminator 142a described in FIG. 5 will be exemplarily described.
  • the first discriminator 142a of the plurality of discriminators 142a, 142b, 142n is illustrated and described, but a plurality of discriminators with different descriptions of the first discriminator 142a It will be apparent to those skilled in the art that the same can be applied to (142b, 142n).
  • the generator 141 has a neural network structure for generating virtual data.
  • the generator 141 may have at least one input layer, at least one output layer, and at least one hidden layer.
  • the generator 141 receives random noise through an input layer. In this case, classification information (not shown) for the virtual data to be generated may be input together.
  • the generator 141 outputs a virtual data sample that simulates the real data 30 through an artificial neural network action of the generator 141.
  • the output virtual data sample is provided to the first discriminator 142a.
  • the first discriminator 142a has an artificial neural network structure for determining whether the input data sample is actual data. Similarly, the first discriminator 142a may have at least one input layer, at least one output layer, and at least one hidden layer. The first discriminator 142a receives the virtual data sample and the actual data sample 30 generated by the generator 141, and determines whether the data sample received through its own artificial neural network is actually data. Then, the determination result is provided to the reinforcement learning performing unit 143.
  • FIG. 7 and 8 are flowcharts illustrating a method of generating and learning virtual data by the value evaluation apparatus 100 according to some embodiments of the present disclosure.
  • step S110 the value evaluation apparatus 100 generates a virtual data sample that simulates an actual data sample using the generator 141.
  • the virtual data sample generated at this time may be a data sample including financial information of a company.
  • step S120 the value evaluation apparatus 100 determines whether or not the generated data sample is an actual data sample through the plurality of discriminators 142a, 142b, and 142n. At this time, each of the plurality of discriminators 142a, 142b, and 142n independently determines whether or not the data sample is the actual data sample.
  • step S130 the value evaluation apparatus 100 uses the reinforcement learning performing unit 143 to determine the determination results performed by the plurality of discriminators 142a, 142b, and 142n. And according to the decision result, the policy gradient is determined.
  • the value evaluation apparatus 100 may determine a policy gradient by determining whether the determination results of the plurality of discriminators 142a, 142b, and 142n are similar to each other.
  • FIG. 8 a specific method of determining a policy gradient through determination of the determination result of a plurality of discriminators by the value evaluation apparatus 100 is illustrated.
  • step S131 the value evaluation apparatus 100 determines whether the standard deviation of data output from the plurality of discriminators 142a, 142b, and 142n is less than a predetermined value.
  • the predetermined value may be 0.3. If the standard deviation is less than a predetermined value, the present embodiment proceeds to step S132. Otherwise, the present embodiment proceeds to step S133.
  • step S132 the value evaluation apparatus 100 determines that the plurality of discriminators 142a, 142b, and 142n have made similar determinations, and determines a policy gradient to reduce the weight of the corresponding data sample.
  • the value evaluation device 100 considers that the plurality of discriminators 142a, 142b, and 142n have made different discriminations, and the policy slope to increase the weight of the corresponding sample. Decide.
  • step S140 the value evaluation apparatus 100 backpropagates the determined policy gradient to the generator 141 and the plurality of discriminators 142a, 142b, and 142n, respectively, and the generator 141 and A plurality of discriminators 142a, 142b, 142n enables deep learning learning.
  • 9 is a diagram illustrating a value evaluation method according to some embodiments of the present disclosure and a performance of the apparatus 100 in comparison with the prior art.
  • 9, (a) shows the virtual data generation performance of the deep learning model according to the conventional GAN algorithm, and (b) shows the virtual data generation of the deep learning model according to the improved GAN algorithm proposed by the present disclosure.
  • real loss(D) is a graph of the loss rate for real data determined by the discriminant ratio, which means the probability that the discriminator determines that the real data is fake, and the fake loss(D) is discriminated.
  • a loss rate graph 41a for real data and a loss rate graph 41b for virtual data determined by the discriminator converge around about 40 generations (epoch).
  • the loss rate graph 42a for real data and the loss rate graph 42b for virtual data determined by the discriminator converge around about 20 generations. You can see that the running model is generating virtual data that resembles real data much faster.
  • the graph of FIG. 9 (b) shows a much lower value than the graph of FIG. 9 (a), so that the deep learning model proposed by the present disclosure has a lower error rate. Can be seen to show.
  • FIG. 10 is a flowchart illustrating a method of calculating a value evaluation result of an evaluation target using virtual data according to some embodiments of the present disclosure.
  • the method for evaluating a value according to the present embodiment includes a series of steps S210 to S240.
  • the value evaluation apparatus 100 acquires classification information of an evaluation target.
  • the classification information may be classification information according to the Korean Standard Industrial Classification (KSIC).
  • step S220 the value evaluation apparatus 100 generates a data set including a virtual data sample corresponding to the obtained classification information by using a deep learning model.
  • the deep learning model may be a deep learning model according to the algorithm and learning method described in FIGS. 5 to 8.
  • step S230 the value evaluation apparatus 100 calculates evaluation basic data for an evaluation target by using the generated data set.
  • the calculated evaluation basic data is data to be provided as an input value to the value evaluation model, and may further include various data in addition to data calculated using the generated data set. This will be described in more detail with reference to FIG. 11.
  • the evaluation basic data 1 includes four data of free cash flow (FCF), discount rate (r), technology life (n), and technology contribution (T.F.).
  • the free cash flow (FCF) is calculated from the data set 50 including a virtual data sample provided by the virtual data generating unit 140.
  • Free cash flow (FCF) may be defined as a function that receives each item of financial data of a company as an input value, as exemplified in Equations 1 and 2 above.
  • the virtual data generator 140 may generate a virtual data set corresponding to the classification information by referring to classification information (eg, KSIC) included in the input information 4.
  • classification information eg, KSIC
  • the user since the free cash flow (FCF) can be calculated according to the generated virtual data set, the user simply enters the classification information of the evaluation object to calculate the expected free cash flow (FCF) for the evaluation object. can do.
  • the discount rate (r), the technical life (n), and the technical contribution (T.F.) can also be defined as functions of input variables obtained from input information, similar to the free cash flow (FCF). However, the discount rate (r), technical life (n), and technical contribution (T.F.) are different in that they do not depend on the virtual data set and can be calculated based on the input information 4 input by the user.
  • the user inputs KSIC, IPC, company size, and key factor scores (ie, technology score, rights score, marketability score, and business score) as input information (4).
  • the discount rate (r) may be defined as a function in which a value is determined based on the score of the core factor that evaluates the industry group to which the evaluation object belongs, the size of the enterprise to be evaluated, and the technology of the evaluation object.
  • the technical life (n) can be defined as a function whose value is determined based on the IPC and the core factor score.
  • technology contribution can be defined as a function whose value is determined based on KSIC and the score of the key factors.
  • the evaluation basic data (1) includes items calculated based on a virtual data set, such as free cash flow (FCF), as well as items calculated from the input information (4) without relying on the virtual data set.
  • FCF free cash flow
  • the value evaluation apparatus 100 calculates an evaluation result by applying the calculated evaluation basic data to a value evaluation model.
  • the value evaluation apparatus 100 may calculate an evaluation value distribution for the evaluation target in the form of data or a graph using a plurality of evaluation results. This will be described in more detail with reference to FIG. 12.
  • FIG. 12 is a diagram for explaining a method of calculating an evaluation value distribution by applying the calculated evaluation basic data to a cash flow discount model. Referring to FIG. 12, an example of applying each item of the calculated evaluation basic data 1 to the cash flow discount model 2 is shown.
  • the cash flow discount model 2 is defined as a function of free cash flow (FCF), discount rate (r), technical life (n) and technical contribution (T.F.), as shown in FIG. 12. Therefore, if the evaluation basic data 1 calculated in step S230 is applied to the cash flow discount model 1, the evaluation value of the evaluation target can be easily calculated.
  • the FCF item among the evaluation basic data 1 may be composed of a data set including a plurality of data samples.
  • the result of applying the evaluation basic data 1 to the cash flow discount model 2 may include the same number of values as the number of a plurality of data samples included in the data set. That is, in order to calculate a more objective evaluation value of the evaluation target, the valuation apparatus 100 generates a plurality of virtual data simulating the corporate financial data of the industry to which the evaluation target belongs and applies it to the cash flow discount model 2. . This is because the estimation by multiple sample data can be statistically more meaningful and accurate than the estimation by single sample data, and accordingly, multiple result values applied to the cash flow discount model (2) can also be calculated. .
  • the value evaluation apparatus 100 may process the distribution of the calculated plurality of result values in the form of data or graphs. Referring to FIG. 12, an example (3) of an evaluation value distribution processed in the form of a graph is shown.
  • the calculated evaluation result value or evaluation value distribution 3 may be provided to the user terminal 200 as a value evaluation result.
  • the user simply enters the classification information of the evaluation target, IPC, company size, and core factor scores (i.e., technical score, rights score, market score, and business score). You can get the result of the valuation of the object to be evaluated. Therefore, even a user who does not have expertise in value evaluation can easily evaluate the value of the evaluation target, and user accessibility and convenience in the value evaluation field can be greatly improved.
  • a computing device 2000 includes one or more processors 2100, a bus 2500, a communication interface 2400, and a memory 2200 for loading a computer program executed by the processor 2100.
  • a storage 2300 for storing the value evaluation software 2310.
  • FIG. 13 only the components related to the embodiment of the present invention are shown in FIG. 13. Accordingly, those of ordinary skill in the art to which the present invention pertains can recognize that other general-purpose components may be further included in addition to the components shown in FIG. 13.
  • the computing device 2000 may be implemented in a form in which some of the components illustrated in FIG. 13 are omitted.
  • the processor 2100 controls the overall operation of each component of the computing device 2000.
  • the processor 2100 includes a CPU (Central Processing Unit), MPU (Micro Processor Unit), MCU (Micro Controller Unit), GPU (Graphic Processing Unit), or any type of processor well known in the technical field of the present invention. Can be. Also, the processor 2100 may perform an operation on at least one application or program for executing the method according to the embodiments of the present invention.
  • the computing device 2000 may include one or more processors.
  • the memory 2200 stores various types of data, commands, and/or information.
  • the memory 2200 may load one or more programs 2310 from the storage 2300 into the reception buffer 2210 of the memory 2200 in order to execute the value evaluation method according to embodiments of the present invention.
  • the bus 2500 provides a communication function between components of the computing device 2000.
  • the bus 2500 may be implemented as various types of buses such as an address bus, a data bus, and a control bus.
  • the communication interface 2400 supports wired/wireless Internet communication of the computing device 2000.
  • the communication interface 2400 may support various communication methods other than Internet communication.
  • the communication interface 2400 may be configured to include a communication module well known in the art.
  • the storage 2300 may non-temporarily store the one or more computer programs 2310.
  • the storage 2300 is a nonvolatile memory such as a ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), and flash memory, a hard disk, a removable disk, or well in the technical field to which the present invention pertains. It may be configured to include any known computer-readable recording medium.
  • the computer program 2310 includes instructions that when loaded into the memory 2200 cause the processor 2100 to perform a deep learning based valuation method according to some embodiments of the present invention.
  • the computer program 2310 is an operation of acquiring classification information of an evaluation target, an operation of generating a data set including a virtual data sample associated with the classification information using a deep learning model, and using the data set. Instructions for calculating evaluation basic data and calculating evaluation results by applying the calculated evaluation basic data to a value evaluation model may be included.
  • the value evaluation method according to the present disclosure described with reference to FIGS. 2 to 13 so far may be implemented as a computer-readable code on a computer-readable medium.
  • the computer-readable recording medium is, for example, a removable recording medium (CD, DVD, Blu-ray disk, USB storage device, removable hard disk) or a fixed recording medium (ROM, RAM, computer-equipped hard disk).
  • the computer program recorded on the computer-readable recording medium may be transmitted to another computing device through a network such as the Internet and installed in the other computing device, thereby being used in the other computing device.

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Human Resources & Organizations (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Strategic Management (AREA)
  • Theoretical Computer Science (AREA)
  • Economics (AREA)
  • Physics & Mathematics (AREA)
  • Educational Administration (AREA)
  • General Physics & Mathematics (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

L'invention concerne un procédé et un appareil d'évaluation de valeur basée sur un apprentissage profond. Le procédé d'évaluation de valeur réalisé par l'appareil d'évaluation de valeur selon la présente invention comprend les étapes consistant à : générer, par un générateur, un échantillon de données virtuelles ; déterminer, par chacun d'une pluralité de déterminateurs, l'échantillon de données en parallèle ; recevoir, par une unité de réalisation d'apprentissage par renforcement, des résultats de la détermination par chacun de la pluralité de déterminateurs et déterminer un gradient de politique en fonction des résultats reçus de la détermination ; et effectuer l'apprentissage, par le générateur ou la pluralité de déterminateurs, à l'aide du gradient de politique déterminé. Grâce aux configurations ci-dessus, le volume fiable et suffisant de données à base d'évaluation peut être sécurisé en utilisant une intelligence artificielle basée sur un apprentissage profond. En outre, la commodité pour l'utilisateur peut être améliorée par l'obtention d'un résultat d'évaluation de valeur uniquement avec une entrée d'un élément d'évaluation simple.
PCT/KR2020/012694 2019-10-16 2020-09-21 Procédé et appareil d'évaluation de valeur basée sur un apprentissage profond WO2021075742A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
KR10-2019-0128431 2019-10-16
KR1020190128431A KR102152081B1 (ko) 2019-10-16 2019-10-16 딥러닝 기반의 가치 평가 방법 및 그 장치

Publications (1)

Publication Number Publication Date
WO2021075742A1 true WO2021075742A1 (fr) 2021-04-22

Family

ID=72470905

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/KR2020/012694 WO2021075742A1 (fr) 2019-10-16 2020-09-21 Procédé et appareil d'évaluation de valeur basée sur un apprentissage profond

Country Status (2)

Country Link
KR (1) KR102152081B1 (fr)
WO (1) WO2021075742A1 (fr)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114581146A (zh) * 2022-03-10 2022-06-03 北京乐我无限科技有限责任公司 模型评估方法、装置、电子设备和存储介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR102152081B1 (ko) * 2019-10-16 2020-09-04 한국과학기술정보연구원 딥러닝 기반의 가치 평가 방법 및 그 장치
KR102236336B1 (ko) * 2020-10-14 2021-04-05 해양수산과학기술진흥원 인공 지능(ai) 모델에 기반한 매출 성장률 추정을 위한 전자 장치 및 그 동작 방법

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101536535B1 (ko) * 2014-09-26 2015-07-15 기술신용보증기금 동태적 현금흐름법에 따른 기술가치평가 방법
KR101843066B1 (ko) * 2017-08-23 2018-05-15 주식회사 뷰노 기계 학습에 있어서 데이터 확대를 이용하여 데이터의 분류를 수행하는 방법 및 이를 이용한 장치
KR101936029B1 (ko) * 2018-07-18 2019-01-11 한국과학기술정보연구원 딥러닝 기반의 가치 평가 방법 및 그 장치
KR20190049549A (ko) * 2017-10-30 2019-05-09 한국과학기술원 집단지성과 실시간 인공지능학습엔진을 이용한 확률기반 기술가치평가 및 기술거래 온라인플랫폼
KR101988504B1 (ko) * 2019-02-28 2019-10-01 아이덴티파이 주식회사 딥러닝에 의해 생성된 가상환경을 이용한 강화학습 방법
KR102152081B1 (ko) * 2019-10-16 2020-09-04 한국과학기술정보연구원 딥러닝 기반의 가치 평가 방법 및 그 장치

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20170049328A (ko) 2015-10-27 2017-05-10 중소기업은행 기술 또는 ip의 가치 평가 장치 및 방법

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101536535B1 (ko) * 2014-09-26 2015-07-15 기술신용보증기금 동태적 현금흐름법에 따른 기술가치평가 방법
KR101843066B1 (ko) * 2017-08-23 2018-05-15 주식회사 뷰노 기계 학습에 있어서 데이터 확대를 이용하여 데이터의 분류를 수행하는 방법 및 이를 이용한 장치
KR20190049549A (ko) * 2017-10-30 2019-05-09 한국과학기술원 집단지성과 실시간 인공지능학습엔진을 이용한 확률기반 기술가치평가 및 기술거래 온라인플랫폼
KR101936029B1 (ko) * 2018-07-18 2019-01-11 한국과학기술정보연구원 딥러닝 기반의 가치 평가 방법 및 그 장치
KR101988504B1 (ko) * 2019-02-28 2019-10-01 아이덴티파이 주식회사 딥러닝에 의해 생성된 가상환경을 이용한 강화학습 방법
KR102152081B1 (ko) * 2019-10-16 2020-09-04 한국과학기술정보연구원 딥러닝 기반의 가치 평가 방법 및 그 장치

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114581146A (zh) * 2022-03-10 2022-06-03 北京乐我无限科技有限责任公司 模型评估方法、装置、电子设备和存储介质

Also Published As

Publication number Publication date
KR102152081B1 (ko) 2020-09-04

Similar Documents

Publication Publication Date Title
WO2021075742A1 (fr) Procédé et appareil d'évaluation de valeur basée sur un apprentissage profond
WO2020022639A1 (fr) Procédé et appareil d'évaluation à base d'apprentissage profond
CN111310814A (zh) 利用不平衡正负样本对业务预测模型训练的方法及装置
WO2022163956A1 (fr) Procédé et système à base d'intelligence artificielle pour prédire la demande au niveau d'un magasin
CN110033026A (zh) 一种连续小样本图像的目标检测方法、装置及设备
WO2022146050A1 (fr) Procédé et système d'entraînement d'intelligence artificielle fédéré pour le diagnostic de la dépression
CN113610069B (zh) 基于知识蒸馏的目标检测模型训练方法
US20170372069A1 (en) Information processing method and server, and computer storage medium
CN110929785A (zh) 数据分类方法、装置、终端设备及可读存储介质
US9269055B2 (en) Data classifier using proximity graphs, edge weights, and propagation labels
CN111931809A (zh) 数据的处理方法、装置、存储介质及电子设备
CN107180190A (zh) 一种基于混合特征的Android恶意软件检测方法及系统
CN114419363A (zh) 基于无标注样本数据的目标分类模型训练方法及装置
CN115114329A (zh) 数据流异常检测的方法、装置、电子设备和存储介质
WO2022145800A1 (fr) Appareil informatique de fourniture d'informations d'analyse à propos de quantité d'énergie d'après reconnaissance d'images
Baradie et al. Managing the Fifth Generation (5G) Wireless Mobile Communication: A Machine Learning Approach for Network Traffic Prediction
CN113283388B (zh) 活体人脸检测模型的训练方法、装置、设备及存储介质
CN111062019A (zh) 用户攻击检测方法、装置、电子设备
CN113806501B (zh) 意图识别模型的训练方法、意图识别方法和设备
CN117876018A (zh) 潜在客户识别和预测的方法、装置、电子设备及存储介质
JP2020030698A (ja) 学習装置、学習方法および予測システム
CN115994839A (zh) 作答准确率的预测方法、装置、设备及介质
CN109436980B (zh) 电梯部件的状态检测方法和系统
CN114445656A (zh) 多标签模型处理方法、装置、电子设备及存储介质
CN113569879A (zh) 异常识别模型的训练方法、异常账号识别方法及相关装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20877208

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20877208

Country of ref document: EP

Kind code of ref document: A1