KR20210157302A - Method and Apparatus for Automatic Predictive Modeling Based on Workflow - Google Patents

Abstract

Disclosed are an automatic prediction modeling method based on a workflow and a device for thereof. The automatic prediction modeling method according to an embodiment of the present invention may comprise: an input step of obtaining the initial input data for the source data and the prediction work information; a data analysis step of generating at least one node based on the initial input data, and constructing a prediction workflow based on at least one generated node to convert thereof into the workflow output data; and a workflow output step of outputting the converted workflow output data so that an automatic prediction modeling is performed. Therefore, the present invention is capable of having an effect of being able to process automatic predictions.

Description

Method and Apparatus for Automatic Predictive Modeling Based on Workflow

The present invention relates to a workflow-based automatic predictive modeling method comprising at least one node and an apparatus therefor.

The content described in this section merely provides background information on the embodiments of the present invention and does not constitute the prior art.

For data analysis using machine learning, various steps such as pre-processing, feature extraction, classification, prediction, post-processing, and visualization must be performed. Nodes of operations and algorithms required in each process can be implemented based on different analysts or various programming languages.

Node for data analysis has different grammar, implementation method, and operating environment according to each analyst or programming language, and each analyst who performs data analysis selects and learns a programming language that an individual prefers and mainly uses it.

In the process where multiple analysts must collaborate in an environment that requires complex system analysis, collaboration difficulties arise when multiple data analysts use different programming languages and different node environments. In addition, in the process of performing collaboration of a plurality of analysts, a problem arises in that a number of algorithms configured in stages are connected to each other to increase the complexity of composing an integrated workflow.

The present invention generates a workflow for automatic predictive modeling by creating an algorithm node optimized for each step required for prediction after analyzing a data pattern based on the source data input by the user and the initial input data for the prediction work information. A main object of the present invention is to provide a workflow-based automatic predictive modeling method and an apparatus for the same.

According to one aspect of the present invention, an automatic predictive modeling method for achieving the above object includes an input step of acquiring initial input data for source data and predictive work information; a data analysis step of generating at least one node based on the initial input data, constructing a prediction workflow based on the generated at least one node, and converting it into workflow output data; and outputting the converted workflow output data to perform automatic predictive modeling.

In addition, according to another aspect of the present invention, an automatic predictive modeling apparatus for achieving the above object includes: an input unit for obtaining source data and initial input data for predictive work information; a data analysis unit that generates at least one node, configures a prediction workflow based on the generated at least one node, and converts it into workflow output data; and a workflow output unit for outputting the converted workflow output data to perform automatic predictive modeling.

As described above, the present invention is effective in providing users with an optimal workflow for performing automatic prediction on predetermined tasks and data.

In addition, the present invention has an effect that can process automatic prediction by integrating a plurality of analysts or nodes implemented in heterogeneous programming languages.

1 is a block diagram schematically illustrating an automatic predictive modeling apparatus according to an embodiment of the present invention.
2 is a block diagram schematically illustrating the configuration of a data analysis unit according to an embodiment of the present invention.
3 is a flowchart illustrating an automatic predictive modeling method according to an embodiment of the present invention.
4 is a block diagram schematically illustrating the configuration of a prediction node configuration unit according to an embodiment of the present invention.
5 is a flowchart illustrating an operation of configuring a prediction node according to an embodiment of the present invention.
6 is an exemplary diagram illustrating a workflow for automatic predictive modeling according to an embodiment of the present invention.
7 is a diagram showing the configuration of an automatic predictive modeling apparatus according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description thereof will be omitted. In addition, preferred embodiments of the present invention will be described below, but the technical spirit of the present invention is not limited thereto or may be variously implemented by those skilled in the art without being limited thereto. Hereinafter, a workflow-based automatic predictive modeling method and an apparatus therefor proposed by the present invention will be described in detail with reference to the drawings.

1 is a block diagram schematically illustrating an automatic predictive modeling apparatus according to an embodiment of the present invention.

The automatic predictive modeling apparatus 100 according to the present embodiment includes an input unit 110 , a data analysis unit 120 , and a workflow output unit 130 . The automatic predictive modeling apparatus 100 of FIG. 1 is according to an embodiment, and not all blocks shown in FIG. 1 are essential components, and in another embodiment, some blocks included in the automatic predictive modeling apparatus 100 are added. , may be changed or deleted. On the other hand, the automatic predictive modeling apparatus 100 may be implemented as a computing device, and each component included in the automatic predictive modeling apparatus 100 is implemented as a separate software program, or as a separate hardware device combined with software. can be implemented.

The automatic prediction modeling apparatus 100 analyzes a data pattern based on the prediction job information and source data selected by the user's manipulation or input, selects an algorithm optimized for each step required for prediction, generates a node, and the generated node Constructs a workflow for automatic predictive modeling based on Here, the nodes of the workflow are implemented with multiple algorithms to perform various functions required for prediction (data preprocessing, model learning, etc.) can Each node can be implemented in various programming languages (R, Python, Java, C, GO, Julia, etc.), and the automatic predictive modeling device 100 configures an optimal workflow considering the speed and function of each node. can The automatic predictive modeling apparatus 100 provides the configured final workflow to the user for automatic predictive analysis and visualizes it to be executed.

Hereinafter, each component included in the automatic predictive modeling apparatus 100 will be described.

The input unit 110 obtains initial input data including source data and prediction work information. Here, the prediction job information and the source data may be obtained through upload after being selected by a user's manipulation or input, but is not limited thereto, and may be information and data obtained by fetching from an external device.

The input unit 110 obtains initial input data including source data for at least one of continuous numbers, character variables, image data, etc. can

The input unit 110 may include a task selection unit 112 and an input data acquisition unit 114 .

The input data acquisition unit 114 acquires source data for at least one of a continuous number, a character variable, and image data. Here, the source data means data that a user wants to predict through an algorithm.

Source data may be data composed of continuous numeric or character variables of predetermined bits (eg 64 bits), image data such as jpeg, etc. have.

For example, the source data acquired by the input data acquisition unit 114 may include numeric and character data in the form of tables as shown in [Table 1].

In addition, the source data acquired by the input data acquisition unit 114 is an image in the form of a file such as File 1: image_1.jpeg (1080x720), File 2: image_2.jpeg (1080x720), File 3: image_3.jpeg (1080x720), etc. It may also contain data.

The job selection unit 112 acquires prediction job information about a prediction variable and a prediction type determined by a user's manipulation or input.

The job selection unit 112 may obtain a predictor variable and a prediction type for the source data.

The predictor variable of the job selection unit 112 used for the source data in the form of a table may be nominal data, ordinal data, interval data, ratio data, time series data, and the like. Here, the nominal data means values indicating categories such as gender, food menu, and zip code, and the order data means values with only large and small meanings such as ranking, academic background, and grades. In addition, interval data are quantitatively meaningful, such as temperature, time, and weight, but have no absolute zero point. Ratio data are quantitatively meaningful, such as age and number of defectives, and are values in which absolute zero points exist. In addition, time series data means data with numerical values according to time sequence, such as sales by date.

The prediction type of the task selection unit 112 may be a classification prediction, a regression prediction, a time series prediction, a clustering prediction, and the like.

Here, classification prediction means predicting a variable in the form of a nominal data, grasping the relationship between the explanatory variable and the response variable, and identifying the possible value of the response variable when a new explanatory variable is given. For example, classification prediction may be learning multiple images to predict whether the image is a lion or a tiger.

In addition, regression prediction predicts a variable in a form other than nominal data, and it means identifying the possible value of a response variable when a new explanatory variable is given by understanding the relationship between the explanatory variable and the response variable. For example, the regression prediction may be to predict the fuel efficiency of a vehicle through explanatory variables such as a car maker, maximum speed, weight, and size.

In addition, time series prediction refers to inferring patterns that have not yet occurred in future data by identifying patterns in past data, and time series data is utilized. For example, the time series prediction may be to predict tomorrow's rainfall based on past rainfall data.

In addition, cluster prediction is independent of the shape of a variable and means forming a cluster based on the similarity between observations. For example, the cluster prediction may be to form a personality type cluster of people through a questionnaire related to personality and predict the cluster of a specific person.

The data analysis unit 120 configures a prediction workflow for automatic prediction modeling based on initial input data. Specifically, the data analysis unit 120 generates at least one node by analyzing the initial input data, and configures a prediction workflow based on the generated at least one node. In addition, the data analysis unit 120 converts the prediction workflow into workflow output data.

The configuration and specific operation of the data analysis unit 120 will be described in detail with reference to FIGS. 2 and 4 .

The workflow output unit 130 outputs the workflow output data in which the predictive workflow is converted so that automatic predictive modeling is performed.

The workflow output unit 130 may output the workflow output data using a display equipped with it, but is not limited thereto, and transmits the workflow output data to perform automatic predictive modeling in a separate external device or user terminal. may be

The workflow output unit 130 outputs the workflow output data to show the node configuration of the workflow to the user, and to modify the workflow or execute the workflow-based automatic predictive modeling.

2 is a block diagram schematically illustrating the configuration of a data analysis unit according to an embodiment of the present invention.

The data analysis unit 120 according to the present embodiment may include a prediction node configuration unit 210 and a workflow visualization processing unit 220 . The data analysis unit 120 of FIG. 2 is according to an embodiment, and not all blocks shown in FIG. 2 are essential components, and in another embodiment, some blocks included in the data analysis unit 120 are added or changed. Or it can be deleted. Each component included in the data analysis unit 120 may be implemented as a separate software program, or as a separate hardware device combined with software.

The prediction node configuration unit 210 generates at least one node for each step for automatic prediction, and configures a prediction workflow for automatic prediction based on the at least one node.

The prediction node configuration unit 210 generates nodes necessary for automatic prediction based on initial input data including source data and prediction work information, selects all or some of the generated nodes, and connects them to configure a prediction workflow. .

The prediction node configuration unit 210 may configure a prediction workflow by selecting a node through an optimization algorithm because it is inefficient to test the connectivity of all possible nodes. That is, the prediction node configuration unit 210 may configure the prediction workflow through four steps of configuring a verification method, performing pre-processing, model learning processing, and verification processing. Each step included in the prediction node configuration unit 210 will be described in detail with reference to FIG. 4 .

The workflow visualization processing unit 220 converts the predictive workflow into workflow output data in order to provide it to the user in the form of a visualization.

The workflow visualization processing unit 220 is described as converting the prediction workflow including the finally selected final nodes into the workflow output data, but is not necessarily limited thereto, and is associated with the final node so that the user can change it at each step. A prediction workflow may be converted into workflow output data in a form including at least one candidate node.

Meanwhile, the data analysis unit 120 according to the present embodiment may further include an optimization language selection unit (not shown).

The optimization language selection unit (not shown) performs the same function as a predetermined node included in the prediction workflow output from the prediction node configuration unit 210, but when a node written in a different language (programming language) exists, the function is The operation of composing the final prediction workflow is performed by selecting the optimal node composed of the performing optimization language.

The optimization language selector (not shown) performs the same function when composing the prediction workflow, but when there are many nodes written in different languages, the processing speed of the workflow is optimized by considering the connection state of the previous node and the next node Select the optimal node that makes it possible. Here, the selection criterion of the optimal node composed of the optimization language may be determined through statistical verification.

The prediction node configuration unit 210 is configured based on prediction performance regardless of the language used in the node. However, the prediction performance has a characteristic that varies depending on the dataset and the initial model value. Therefore, the optimization language selection unit (not shown) is the optimization language when a plurality of prediction workflows are configured and recommended by the prediction node configuration unit 210 and there is a difference in prediction performance according to languages in each of the plurality of prediction workflows. can be selected to select the final prediction workflow.

On the other hand, the optimization language selection unit (not shown) is recommended when a plurality of prediction workflows are configured in the prediction node configuration unit 210 and there is no significant difference in prediction performance in each of the plurality of prediction workflows. You can also choose one final prediction workflow that reduces the system load by taking additional factors into consideration.

For example, a predetermined node (step) may be created in the R language or Python language using the same method of generating a derived variable. Random Forest, a classification algorithm, also has versions for the R language and Python language, and there is a difference in prediction performance and operation speed. Accordingly, the selection of an algorithm for model learning in the prediction node configuration unit 210 may include not only the type of algorithm but also the language in which the algorithm is written.

In general, since the operation of the same algorithm is similar even in different languages, if there is no difference in performance, the optimization language selection unit (not shown) may select a language with a low memory usage or a high speed.

The optimization language selection unit (not shown) utilizes statistical verification as a criterion for determining whether there is a difference in prediction performance in the prediction workflow configured by the prediction node configuration unit 210 .

For example, when k prediction performances are measured in one prediction workflow by the verification method configuration unit 410 of the prediction node configuration unit 210, ANOVA verification is performed because each workflow has a prediction performance index. can verify the difference in average prediction performance. The nodes used in each workflow may overlap each other, and according to the genetic algorithm principle, if there is a node that contributed to good performance in the previous stage, it will remain and overlap because the node in another stage is changed. There are many.

The optimization language selection unit (not shown) determines the memory usage and execution time of the high-level workflow with no statistically significant difference in prediction performance.

The memory usage and execution time of the prediction workflow may be stored in an optimization language selection unit (not shown), and each prediction workflow has one value.

In principle, priorities for memory and time are determined by the user, but some conditions may be introduced into the optimization language selection unit (not shown).

Since the stability of the automatic predictive modeling apparatus 100 is the best, if the memory load of the current system is below a certain level, the optimization language selection unit (not shown) selects a predictive workflow with a low memory usage as the top priority. On the other hand, if the memory load of the automatic predictive modeling apparatus 100 is less than a certain level, the optimization language selection unit (not shown) selects the execution time as the top priority.

In addition, if the condition is unclear, the optimization language selection unit (not shown) tests the memory and execution time of all prediction workflows for the optimization language, configures the prediction workflow with a probability distribution according to the test results, and selects Calculate the probability density value of the memory and execution time of the desired prediction workflow.

The optimization language selection unit (not shown) selects the best prediction workflow based on time, excluding the corresponding prediction workflow, if it exceeds the top 10% among the probability density values of memory and execution time. For example, memory usage may be located in the top 5% of the memory probability distribution (memory is very intensive), and execution time may be located in 50% of the time probability distribution (time consumption is average).

The optimization language selection unit (not shown) selects the prediction workflow based on the required execution time because the memory usage is extreme even if the automatic prediction modeling apparatus 100 has room to spare.

The optimization language selection unit (not shown) provides a method for interworking between the nodes connected to each other when the nodes are implemented in a heterogeneous language.

For example, nodes composed of different languages such as R and Python have different data structures, so data cannot be passed to the next node.

The optimization language selection unit (not shown) converts data using a common format that both R and Python languages can understand by utilizing meta information for interworking between nodes implemented in heterogeneous languages. The optimization language selection unit (not shown) unifies the input and output of each node in JSON format, and includes the process of converting JSON data in R script and Python script when creating a node.

Each workflow may exist in an independent environment Docker of the automatic predictive modeling apparatus 100 . The execution speed is determined by the time when the input value is passed to the first node of Docker and the output value is output to the last node, and in the distributed processing system, the resources required for each independent environment are automatically distributed, so you can understand the memory usage.

3 is a flowchart illustrating an automatic predictive modeling method according to an embodiment of the present invention.

The automatic predictive modeling apparatus 100 acquires initial input data including source data and predictive work information (S310, S320). Here, the automatic predictive modeling apparatus 100 includes source data for at least one of continuous numbers, character variables, image data, etc. Input data can be obtained.

The automatic predictive modeling apparatus 100 analyzes initial input data to generate at least one node (S330), and configures a predictive workflow based on the at least one generated node (S340).

The automatic prediction modeling apparatus 100 generates nodes necessary for automatic prediction based on initial input data including source data and prediction work information, selects all or some of the generated nodes, and connects them to configure a prediction workflow. . The automatic predictive modeling apparatus 100 converts the configured predictive workflow into workflow output data.

The automatic predictive modeling apparatus 100 outputs the workflow output data in which the predictive workflow is converted to perform automatic predictive modeling ( S350 ). The automatic predictive modeling apparatus 100 may output the workflow output data using a display equipped with it, but is not limited thereto, and transmits the workflow output data so that the automatic predictive modeling is performed in a separate external device or user terminal. may be

Although it is described that each step is sequentially executed in FIG. 3 , it is not necessarily limited thereto. In other words, since it may be applicable to changing and executing the steps described in FIG. 3 or executing one or more steps in parallel, FIG. 3 is not limited to a chronological order.

The automatic prediction modeling method according to the present embodiment described in FIG. 3 may be implemented as an application (or program) and recorded in a terminal device (or computer) readable recording medium. The recording medium in which the application (or program) for implementing the automatic prediction modeling method according to the present embodiment is recorded and the terminal device (or computer) can read is any type of recording device in which data that can be read by the computing system is stored. or media.

4 is a block diagram schematically illustrating the configuration of a prediction node configuration unit according to an embodiment of the present invention.

The prediction node configuration unit 210 according to the present embodiment includes a verification method configuration unit 410 , a preprocessing unit 420 , a model learning processing unit 430 , and a verification processing unit 440 . The prediction node configuration unit 210 of FIG. 4 is according to an embodiment, and not all blocks shown in FIG. 4 are essential components, and in another embodiment, some blocks included in the prediction node configuration unit 210 are added. , may be changed or deleted. Each of the components included in the prediction node configuration unit 210 may be implemented as a separate software program, or may be implemented as a separate hardware device combined with software.

The verification method configuration unit 410 classifies the initial input data into learning data and verification data.

In order to verify the prediction performance of the prediction model, the verification method configuration unit 410 separates the training data used for generating the prediction model and the verification data used for verification from each other.

There are various methods for dividing the initial input data obtained by the verification method configuration unit 410 into learning and verification data, and a classification method may be selected according to the data size. For example, the verification method configuration unit 410 may classify the initial input data by selecting one of Monte Carlo verification, k-fold cross-validation, leave-p-out cross-validation, and the like.

The Monte Carlo verification method uses data randomly sampled every time from the entire data as verification data, and is suitable for large data because of its high speed. In addition, the k-fold cross-validation method divides data into k datasets according to predetermined conditions, uses k-1 datasets as training data, and the remaining one as validation data, repeats k times to predict the average. It yields performance and is suitable for small data due to its slow speed.

In addition, the leave-p-out cross-validation method selects p pieces of data for validation and uses the remainder for training. It is suitable for small data below the established standard.

The preprocessing unit 420 performs an operation of transforming each of the training data and the verification data to be used in a predictive model.

Although the pre-processing unit 420 defines usable data types and types, it cannot be directly used in a predictive model. Since a predictive model is usually to learn the relationship between each variable, preprocessing must be performed to convert it into a format suitable for it.

The preprocessing unit 420 performs all or part of the process of checking and converting data types for each variable, processing of missing values, determining and removing outliers, and processing of generating derived variables for each of the training data and the verification data. pre-processing including

The process of checking and converting data types for each variable refers to a process of confirming whether data meets the prediction purpose selected by the user. For example, if the preprocessing unit 420 selects classification prediction and the response variable is a ratio data type, it performs a process of converting it into a nominal data type.

The missing value processing process refers to the process of removing missing values or replacing them with other values because missing values may not be learned or may affect prediction. For example, when predicting vehicle fuel efficiency, the preprocessing unit 420 performs a process of replacing an observed value including NA in a weight variable with an average value of the weight variable.

The process of determining and removing outliers refers to a process of removing or replacing each variable with another value when it greatly exceeds the range of each variable or has a large difference from other values. For example, the preprocessing unit 420 performs a process of removing 1,000 ton from the observation value when 0.1 to 2 ton and 1,000 ton are mixed in the values of the vehicle weight variable.

The process of generating a derived variable refers to the process of generating a new explanatory variable by transforming the explanatory variable in order to improve prediction performance. For example, when there is an exponentially increasing relationship between vehicle fuel efficiency and weight, the preprocessing unit 420 generates a weight derived variable obtained by squaring the weight variable to improve prediction performance.

The model learning processing unit 430 selects a specific algorithm among at least one algorithm by inputting the training data as an input, analyzes a pattern of the training data based on the at least one algorithm to select a candidate algorithm, and applies the verification data to the candidate algorithm. A predictive model is created based on the final algorithm selected by substituting it.

The model learning processing unit 430 may generate a predictive model using various statistical models, machine learning models, deep learning models, user-defined models, and the like, and each model may include various algorithms.

The prediction model generated by the model learning processing unit 430 receives the training data in step i, and applies the data for verification to the model generated in step ii and step ii of maximizing prediction performance by analyzing the pattern of the training data through an algorithm. It is processed by the same process as step iii to check the performance.

For example, when the model learning processing unit 430 processes the classification prediction model, the configuration and operation of the prediction model according to the function, and the representative algorithm are as follows.

Step i) The model learning processing unit 430 puts the explanatory variables and response variables specified by the user into the input values of the algorithm, and adjusts detailed parameters necessary for algorithm learning through the system. The prediction performance is different depending on the parameter value, and the setting is performed on how to explore and adjust it.

Step ii-a) The model learning processing unit 430 prepares a plurality of algorithms that can be used for classification prediction. Here, the classification algorithm for classification prediction may be a classification random forest, a logistic linear regression model, a classification neural network, a support vector machine, and the like.

Step ii-b) The model learning processing unit 430 searches for parameters to be used in the algorithm. Here, the parameters to be used in the algorithm are options that affect the complexity, speed, accuracy, etc. of the algorithm.

In the parameter search, the algorithm randomly tests some of all conditions of the parameters that can have in the data, removes the condition in which the performance of the initial test result is less than the preset standard, leaves the condition above the preset standard, and then adjusts other parameters. say action. Here, the algorithm for the parameter search is preferably a genetic algorithm, but is not necessarily limited thereto.

Step iii) The model learning processing unit 430 selects and stores the best algorithm and parameter as a result of the verification data test.

When the model learning processing unit 430 performs model learning on a regression prediction model, a time series prediction model, a cluster prediction model, etc., there is no significant difference from the above-described classification prediction model process, and there may be a difference in the algorithm. For example, the regression prediction model may include algorithms such as a regression random forest, a linear regression model, a principal component regression model, and a partial least squares regression model, and the time series prediction model may include an algorithm such as an ARIMA model and a GARCH model. Also, the cluster prediction model may include algorithms such as K-means and nearest neighbor selection model.

The verification processing unit 440 configures a prediction workflow by performing verification on the prediction model using a preset verification index.

The verification processing unit 440 performs an operation of selecting through which index to evaluate the learned predictive model. Here, various indicators for verification may exist.

Although the verification processing unit 440 is described as a separate operation from the model learning processing unit 430 , it is not necessarily limited thereto, and it may be included in the model learning processing unit 430 to verify the predictive model.

The verification processing unit 440 performs an operation of verifying whether the generated prediction model exhibits good prediction performance even in the verification data. The verification processing unit 440 performs relative evaluation between models rather than absolute performance evaluation. The verification processing unit 440 uses an index comparing the value predicted by the predictive model and the actual value.

For example, the verification index for the classification prediction model may be an F1 score, an area under curve, etc., and the verification index for the regression prediction model and the time series prediction model may be a root mean square error, r squared, and the like. Also, the verification index for the cluster prediction model may be a Dunn index, a Jaccard index, or the like.

Preferably, the verification processing unit 440 applies one index to each predictive model, but is not limited thereto. For example, the verification processing unit 440 performs primary verification as an evaluation criterion for the first among the indices listed for each predictive model, and uses the second indicator among the listed indicators as the evaluation criterion when the primary verification result performance is the same can perform secondary verification.

The prediction node configuration unit 210 according to the present embodiment may generate a prediction flow that proceeds in a plurality of steps. For example, when the prediction flow proceeds in four steps, there may be dozens to hundreds of options in each step. In this case, if there are n options in each step, the number of configurable types is up to n ⁴ , and there is a problem in that the time complexity and the space complexity increase enormously.

The prediction node configuration unit 210 according to the present embodiment may configure a prediction workflow using a genetic algorithm to solve this problem. Here, the genetic algorithm is one of the optimization techniques, and it is an algorithm that gradually evolves to have an optimal solution by selecting the best value at each step and remembering the characteristic to have the best value, rather than checking all near-infinite options.

The prediction node configuration unit 210 configures a workflow by initially selecting several nodes at random for automatic configuration, and leaves the nodes used in the workflow with excellent prediction performance in the selection. The prediction node configuration unit 210 continues the above process through a genetic algorithm, crosses nodes of a workflow with excellent predictive performance, constructs a new workflow, and randomly adds nodes to the process. At this time, the prediction node configuration unit 210 excludes nodes showing bad prediction performance, and continues iteratively until the prediction performance is no longer improved, ultimately selecting only the nodes that have a good influence on the prediction performance, and performing the final prediction work compose the flow.

Hereinafter, the operation of the prediction node configuration unit 210 for selecting an optimal node in the prediction flow composed of 4 steps will be described.

The initial step (step 1) and the final step (step 4) among steps 1 to 4 are performed by the operation of the verification method configuration unit 410 , the preprocessing unit 420 , the model learning processing unit 430 , and the verification processing unit 440 . In order to select an optimal prediction flow, the prediction node configuration unit 210 additionally performs an operation of verifying the connection combination of steps 2 and 3.

For example, if there are 10 methods in the step of generating a derived variable in step 2 and 10 algorithms that can be used in step 3, a total of 100 combinations are created, but it requires too much resources to test them all. Accordingly, the prediction node configuration unit 210 according to the present embodiment performs a process of finding an optimal combination based on a genetic algorithm.

The prediction node configuration unit 210 randomly selects 2 to 3 methods in steps 2 and 3, respectively, and performs a test using a combination of the selected methods. For example, the prediction node configuration unit 210 selects three of the various methods for generating the derived variable of step 2 and selects three algorithms that can be used in step 3 to test only 9 times in total.

Thereafter, the prediction node configuration unit 210 measures the performance of each model in steps 2 and 3 through 4 steps. At this time, the prediction node configuration unit 210 selects two to three models with excellent performance, and stores the derived variable generation method (the method selected in step 2) and the algorithm (the method selected in step 3) used at this time.

For example, the prediction node configuration unit 210 selects the top four performance among combinations close to infinity.

For example, the prediction node configuration unit 210 is configured for workflow 1: node (2.1) -> node (3.2) = performance 90%, workflow 2: node (2.2) -> node (3.6) = performance 88%, Workflow 3: Node(2.1) -> Node(3.1) = performance 85%, Workflow 4: Node(2.4) -> Node(3.7) = 80% performance. In method 4 and step 3, methods 2, 6, 1, and 7 can be selected and stored as high-performance results.

The prediction node configuration unit 210 fixes the derived variable method based on the selected result and selects a combination that has not been used in the three-step algorithm, or fixes the selected algorithm and selects a combination that has not been used in the two-step variable method .

For example, the prediction node configuration unit 210 may include workflow 1: node (2.1) -> node (3.11), workflow 2: node (2.2) -> node (3.3), and workflow 3: node (2.1). -> As with node (3.4), 1, 2, and 4 of the 2nd stage node are fixed and the 3rd stage node is configured to be a combination that has not been used before, or 2, 6, 1, 7 of the 3rd stage node is fixed The test can be performed by configuring the second stage node to be a combination that has not been used before.

The prediction node configuration unit 210 may repeat the above process until the performance of the prediction model is no longer improved in step 4 . That is, the prediction node configuration unit 210 may search for an optimal node by applying a method that has not been used to see if there is a better method while leaving the method and algorithm with good performance in the node of each stage.

5 is a flowchart illustrating an operation of configuring a prediction node according to an embodiment of the present invention.

FIG. 5 shows the concrete steps of steps S330 and S340 of FIG. 3 .

The automatic predictive modeling apparatus 100 classifies the initial input data into training data and verification data ( S510 ). The automatic predictive modeling apparatus 100 performs a process of separating the learning data used for generating the predictive model and the verification data used for the verification from each other in order to verify the predictive performance of the predictive model.

The automatic predictive modeling apparatus 100 performs an operation of transforming each of the training data and the verification data to be used in the predictive model ( S520 ). Automatic predictive modeling apparatus 100 for each of the data for training and verification data, check and transform the data type for each variable, missing value (missing value) processing, outlier determination and removal processing, all or part of the process, such as generation processing of derived variables Pre-processing including

The automatic predictive modeling apparatus 100 selects a specific algorithm among at least one algorithm by inputting the learning data as an input, analyzes a pattern of the learning data based on the at least one algorithm, selects a candidate algorithm, and selects the candidate algorithm for verification data A predictive model is generated based on the final algorithm selected by substituting (S530).

The automatic predictive modeling apparatus 100 configures a predictive workflow by verifying the predictive model using a preset verification index (S540).

Although it is described that each step is sequentially executed in FIG. 5 , the present invention is not limited thereto. In other words, since it may be applicable to changing and executing the steps described in FIG. 5 or executing one or more steps in parallel, FIG. 5 is not limited to a chronological order.

The prediction node configuration method according to the present embodiment described in FIG. 5 may be implemented as an application (or program) and recorded in a terminal device (or computer) readable recording medium. A recording medium in which an application (or program) for implementing the prediction node configuration method according to the present embodiment is recorded and a terminal device (or computer) readable recording medium is any type of recording device in which data that can be read by a computing system is stored or media.

6 is an exemplary diagram illustrating a workflow for automatic predictive modeling according to an embodiment of the present invention.

6 illustrates a lifespan prediction workflow 600 for predicting the remaining lifespan through a physical examination result of a patient.

The life prediction flow 600 may include five nodes 610 , 620 , 630 , 640 , and 650 . The first node 610 is a node that reads data to be used for analysis from the DB (Read CSV file), and the second node 620 is a node (Partitioner) that divides the data for learning and verification. In addition, the third node 630 is a node that learns data for training with a random forest algorithm (Random Forest Regression Learner), and the fourth node 640 is a node that predicts data for verification using the learned random forest algorithm. . In addition, the fifth node 650 is a node (Regression Score) for evaluating the performance of the prediction data for verification.

The automatic prediction modeling apparatus 100 may perform the following process to configure the lifetime prediction workflow 600 .

The automatic predictive modeling apparatus 100 performs a process of uploading analysis data and selecting a job. For example, the user wants to predict the remaining lifespan through the physical examination results of each patient, the automatic predictive modeling apparatus 100 obtains source data related to predicting the lifespan, and 'regression prediction' is selected as the prediction operation, Acquire predictive work information targeting variables related to lifespan.

Thereafter, the automatic prediction modeling apparatus 100 performs a prediction node configuration and recommendation process.

The automatic predictive modeling apparatus 100 automatically analyzes whether to add a node necessary for preprocessing by inserting a node that divides data for learning and for verification, and analyzing whether missing value processing and character variables are included.

The automatic predictive modeling apparatus 100 may analyze missing values and variable types in the DB system when data is uploaded. When constructing the second-stage node 620 using this information, selectable nodes are added to the options in a predefined manner. For example, if a missing value is found in a variable when data is uploaded to the automatic predictive modeling apparatus 100, a missing value processing node may be additionally selected in the detailed process of step 2. Thereafter, the automatic predictive modeling apparatus 100 automatically selects several algorithms most suitable for regression prediction.

Thereafter, the automatic predictive modeling apparatus 100 performs a process of selecting an optimization language.

The automatic predictive modeling apparatus 100 selects a case with the best prediction performance or the fastest speed among nodes implementing the same algorithm in different languages. For example, if the third node 630 has the R version of the random forest and the Python version of the random forest, the automatic prediction modeling device 100 configures the prediction workflow with the R version and the Python version, respectively, and the prediction performance is better. You can choose a good version.

Thereafter, the automatic predictive modeling apparatus 100 performs a workflow visualization process.

The automatic prediction modeling apparatus 100 configures one prediction workflow by connecting the finally selected nodes, and converts the configured prediction workflow into workflow output data and outputs it to be executed.

7 is a diagram showing the configuration of an automatic predictive modeling apparatus according to an embodiment of the present invention.

The automatic predictive modeling apparatus 700 illustrated in FIG. 7 may be implemented as a computing device, and includes at least one processor 710 , a computer-readable storage medium 720 , and a communication bus 760 .

The input unit 110 and the workflow output unit 130 of the automatic predictive modeling apparatus 700 may correspond to the input/output interface 740 or the communication interface 750 , and the data analysis unit 120 may correspond to the processor 710 . can

The processor 710 may control the automatic predictive modeling apparatus 700 to operate. For example, the processor 710 may execute one or more programs stored in the computer-readable storage medium 720 . The one or more programs may include one or more computer-executable instructions, which when executed by the processor 710 configure the automated predictive modeling apparatus 700 to perform operations in accordance with the exemplary embodiment. can be

Computer-readable storage medium 720 is configured to store computer-executable instructions or program code, program data, and/or other suitable form of information. The program 730 stored in the computer-readable storage medium 720 includes a set of instructions executable by the processor 710 . In one embodiment, computer-readable storage medium 720 includes memory (volatile memory, such as random access memory, non-volatile memory, or a suitable combination thereof), one or more magnetic disk storage devices, optical disk storage devices, It may be flash memory devices, other types of storage media that can be accessed by the automatic predictive modeling apparatus 700 and store desired information, or a suitable combination thereof.

The communication bus 760 interconnects various other components of the automatic predictive modeling apparatus 700 including the processor 710 and the computer-readable storage medium 720 .

The automatic predictive modeling apparatus 700 may also include one or more input/output interfaces 740 and one or more communication interfaces 750 that provide interfaces for one or more input/output devices. The input/output interface 740 and the communication interface 750 are coupled to the communication bus 760 . The input/output device may be connected to other components of the automatic predictive modeling device 700 through the input/output interface 740 .

The above description is merely illustrative of the technical idea of the embodiment of the present invention, and those of ordinary skill in the art to which the embodiment of the present invention pertains may modify various modifications and transformation will be possible. Accordingly, the embodiments of the present invention are not intended to limit the technical spirit of the embodiment of the present invention, but to explain, and the scope of the technical spirit of the embodiment of the present invention is not limited by these embodiments. The protection scope of the embodiment of the present invention should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be interpreted as being included in the scope of the present invention.

100: automatic predictive modeling device
110: input unit 120: data analysis unit
130: workflow output unit
210: prediction node configuration unit 220: workflow visualization processing unit

Claims (12)

A method for performing automatic predictive modeling based on a workflow in an automatic predictive modeling apparatus, the method comprising:
an input step of obtaining initial input data for the source data and the prediction work information;
a data analysis step of generating at least one node based on the initial input data, constructing a prediction workflow based on the generated at least one node, and converting it into workflow output data; and
A workflow output step of outputting the converted workflow output data so that automatic predictive modeling is performed
Automatic predictive modeling method comprising a. According to claim 1,
The input step is
Acquiring the initial input data including the source data for at least one of continuous numeric, character-type variables, and image data, and the predictive work information for predictive variables and prediction types determined by a user's manipulation or input, characterized in that An automatic predictive modeling method. According to claim 1,
The data analysis step is
a prediction node configuration step of generating the at least one node and configuring a prediction workflow for the automatic prediction based on the at least one node; and
A workflow visualization processing step of converting the prediction workflow into the workflow output data in order to provide it to the user in the form of a visualization
Automatic predictive modeling method comprising a. 4. The method of claim 3,
The prediction node configuration step includes:
a verification method configuration step of classifying the initial input data into training data and verification data;
performing a preprocessing step of converting each of the training data and the verification data to be used in a predictive model;
Selecting a specific algorithm among at least one algorithm by inputting the learning data as an input, analyzing a pattern of the learning data based on the at least one algorithm to select a candidate algorithm, and substituting the verification data into the candidate algorithm to select A model learning processing step of generating a predictive model based on the final algorithm; and
A verification processing step of configuring the prediction workflow by performing verification on the prediction model using a preset verification index
Automatic predictive modeling method comprising a. 5. The method of claim 4,
The pre-processing step is,
For each of the learning data and the verification data, a preprocessing including all or part of a process of checking and converting data types for each variable, processing of missing values, processing of outlier determination and removal, and generation of derived variables is performed. Automatic predictive modeling method, characterized in that. 5. The method of claim 4,
The model learning process step is,
i) receiving the training data as input, ii) maximizing the prediction performance by analyzing the pattern of the training data through an algorithm, and iii) substituting the verification data into the model generated in the maximizing step to check the performance Automatic predictive modeling method, characterized in that processed by a process comprising the step of. 4. The method of claim 3,
Further comprising an optimization language selection step of selecting an optimal node composed of an optimization language in the prediction workflow to select a final prediction workflow so that the final prediction workflow is converted into the workflow output data,
The optimization language selection step includes:
If there is a node that performs the same function as a predetermined node included in the prediction workflow but is written in a different language, an optimal node composed of an optimization language is selected in consideration of the connection state of the node for performing the function, and the final An automatic predictive modeling method, characterized in that the predictive workflow is selected. In an apparatus for performing automatic predictive modeling based on a workflow,
an input unit for obtaining initial input data for source data and prediction work information;
a data analysis unit that generates at least one node, configures a prediction workflow based on the generated at least one node, and converts it into workflow output data; and
A workflow output unit that outputs the converted workflow output data so that automatic predictive modeling is performed
Automatic predictive modeling device comprising a. 9. The method of claim 8,
The data analysis unit,
a prediction node configuration unit that generates the at least one node and configures a prediction workflow for the automatic prediction based on the at least one node; and
A workflow visualization processing unit that converts the prediction workflow into the workflow output data in order to provide it to the user in the form of a visualization
Automatic predictive modeling device comprising a. 10. The method of claim 9,
The prediction node configuration unit,
a verification method configuration unit for classifying the initial input data into learning data and verification data;
a pre-processing unit converting each of the training data and the verification data to be used in a predictive model;
Selecting a specific algorithm among at least one algorithm by inputting the learning data as an input, analyzing a pattern of the learning data based on the at least one algorithm to select a candidate algorithm, and substituting the verification data into the candidate algorithm to select a model learning processing unit for generating a predictive model based on the final algorithm; and
A verification processing unit configured to configure the prediction workflow by performing verification on the prediction model using a preset verification index
Automatic predictive modeling device comprising a. 10. The method of claim 9,
Automatic prediction, characterized in that it further comprises an optimization language selection unit that selects an optimal node composed of an optimization language in the prediction workflow to select a final prediction workflow so that the final prediction workflow is converted into the workflow output data modeling device. A computer program stored in a recording medium for executing the automatic predictive modeling method according to any one of claims 1 to 7 on the computer.

Images