KR102363373B1 - System of recommending activity using artificial neural network, method of recommending activity using artificial neural network and computer program therefor - Google Patents

Abstract

The present invention relates to an activity recommendation system using an artificial neural network, an activity recommendation method using an artificial neural network, and a computer program therefor. According to an embodiment of the present invention, in designing a business process model combining IT services, it is recommended to probabilistically select the next activity based on the order and/or context of input activities. Disclosed are the activity recommendation system using an artificial neural network, the activity recommendation method using an artificial neural network, and the computer program therefor.

Description

System of recommending activity using artificial neural network, method of recommending activity using artificial neural network and computer program therefor

The present invention relates to an activity recommendation system using an artificial neural network, an activity recommendation method using an artificial neural network, and a computer program therefor, and based on the order and/or context of input activities in designing a business process model combining IT services It relates to an activity recommendation system using an artificial neural network that recommends the next activity to be probabilistically selected as a method, an activity recommendation method using an artificial neural network, and a computer program for this.

In designing a business process model that combines various IT services including web services based on BPEL (Business Process Execution Language) or similar standard or non-standard process technology, various types of activities are connected sequentially or in parallel. will complete the design.

As a prior art related to business process model design, Republic of Korea Patent Registration 10-1596257 (February 16, 2016) uses services and service models as core assets in the Software Product Line, but uses new services In the course of development, we proposed a configuration in which business process models matching the goals were imported and reused or new business processes were modeled, and each activity of the business process was set.

As another prior art, Korean Patent Laid-Open Publication No. 10-2009-0131426 (December 29, 2009) discloses a step in which an electronic device receives a workflow pattern block additional command and the electronic device responds to the additional command Including the step of adding a workflow pattern block corresponding to the additional command to the position to be performed, a configuration was proposed to prevent unexpected errors without the user's separate action when performing workflow process modeling.

As another prior art, Korean Patent Registration No. 10-1127701 (Mar. 09, 2012) provides a business process family model that supports the expression of commonality and variability in a user's business process, This model is realized as a web service that can be executed by a computer system and a configuration is proposed to create a web service that meets the user's requirements.

By the way, although the prior arts suggested a configuration to prevent errors in the process modeling process or to increase efficiency, a configuration to increase the efficiency and accuracy of design work, such as recommending an activity required for process modeling to a process designer, was not proposed. .

For example, in the design of a process composed of activities, a designer or developer who designs and develops the process selects specific activities and proceeds by connecting them sequentially or in parallel.

However, while designing the process by selecting and connecting the next activity for the process currently being worked by, the designer frequently selects or incorrectly connects and corrects the activity due to a simple mistake or lack of information. I did not propose a configuration to solve the problem.

Republic of Korea Patent Registration 10-1596257 (February 16, 2016) Republic of Korea Patent Publication 10-2009-0131426 (December 29, 2009) Republic of Korea Patent Registration 10-1127701 (Mar. 09, 2012)

The present invention has been devised in consideration of the above problems, and in designing a business process model combining IT services, an artificial recommendation for probabilistically selecting the next activity based on the order and/or context of the activities inputted An object of the present invention is to provide an activity recommendation system using a neural network, an activity recommendation method using an artificial neural network, and a computer program for the same.

According to one aspect of the present invention in consideration of the above object, parameter data generated based on the activity sequence combination is received from the process design unit, the parameter data is input to the trained artificial neural network model unit, and the artificial neural network model an activity recommendation service unit returning the activity recommendation data obtained in the unit to the process design unit; and an artificial neural network model unit for obtaining activity recommendation data that recommends an activity in the next sequence of the sequence combination of activities included in the parameter data in the form of a probability value; an activity recommendation system using an artificial neural network configured to include.

Preferably, the present invention obtains a data set of training data and target data generated based on a combination of activity sequences included in the process data provided by the process design unit, and inputs the data set to train the artificial neural network model unit. Artificial neural network learning unit; includes.

Preferably, the present invention includes an artificial neural network configuration unit configured to set at least one of the number of neurons required for the construction of the artificial neural network model unit, the type of activation function, the size of the training epoch, and the type of the optimization method.

Preferably, the present invention, a data preprocessor for sequentially extracting text-based activity data included in the process data, and pre-processing it into number-based data to obtain the training data and target data; is configured to include.

Preferably, the training data is data related to a sequence combination of one or more activities included in the process data, and the target data is data related to an activity corresponding to a next sequence of sequence combinations of activities included in the training data.

Preferably, the training data includes a parallel connection activity included in the process data as an activity constituting the sequence combination.

Preferably, the training data includes an attribute of an activity included in the process data as an activity constituting the sequence combination.

Preferably, when one activity type included in the process data has two or more properties, the training data classifies the corresponding activity type as a separate activity type for each property to configure an order combination of activities.

Preferably, the artificial neural network model unit is based on a GRU-type Recurrent Neural Network (RNN) including an input layer, a Gated Recurrent Unit (GRU) cell, a hidden layer, and an output layer.

Preferably, the artificial neural network model unit is configured to further include a dropout layer to avoid overfitting after the hidden layer.

Preferably, the artificial neural network model unit uses a Softmax function as an activation function of the output layer, and each probability value for one or more activities constituting a next sequence of sequence combinations of activities included in the parameter data. is configured to obtain.

Preferably, the input layer of the artificial neural network model unit is based on the training data in which the largest number of activities among the data sets of the training data and the target data generated based on the sequence combination of activities included in the process data constitute the sequence combination. The number of input neurons is determined.

Preferably, the output layer of the artificial neural network model unit is output based on the number of types of different activities constituting the target data included in the data set of the training data and the target data generated based on the sequence of activities included in the process data. The number of neurons is determined.

According to another aspect of the present invention, there is provided an activity recommendation method executed in an activity recommendation system including a memory for storing one or more instructions and a processor for executing the one or more instructions stored in the memory, receiving the parameter data generated based on the process design unit; 2) inputting the parameter data into a trained artificial neural network model unit; 3) obtaining, by using the artificial neural network model unit, activity recommendation data for recommending an activity next to a sequence combination of activities included in the parameter data in the form of a probability value; and 4) returning the activity recommendation data obtained from the neural network model unit to the process design unit; an activity recommendation method using the configured artificial neural network is disclosed.

Preferably, the present invention, before step 1), P1) obtaining a data set of training data and target data generated based on a combination of activity sequences included in the process data provided by the process design unit; and P2) inputting the data set to train the artificial neural network model unit.

According to another aspect of the present invention, a computer-readable recording medium recording a program for executing each step of the activity recommendation method using the artificial neural network is disclosed.

As described above, the present invention recommends probabilistic selection of the next activity in designing a business process by connecting various types of activities sequentially or in parallel, but does not recommend the next activity based on only the currently connected last activity. Since it is recommended in consideration of the contextual element including the sequence combination of a plurality of activities configured before that, there is an advantage of improving the efficiency of the process design work and reducing errors.

1 is a block diagram of an activity recommendation system according to an embodiment of the present invention;
2 is a block diagram of an artificial neural network model unit according to an embodiment of the present invention;
3 is a schematic diagram of a hardware point of view of an activity recommendation system according to an embodiment of the present invention;
4 is an exemplary view of activity recommendation data according to an embodiment of the present invention;
5 and 6 are flowcharts of an activity recommendation method according to an embodiment of the present invention.

The present invention may be embodied in various other forms without departing from its technical spirit or main characteristics. Accordingly, the embodiments of the present invention are merely illustrative in all respects and should not be construed as limiting.

The terms 1st, 2nd, etc. are used only for the purpose of distinguishing one component from another component. For example, without departing from the scope of the present invention, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component.

When a component is referred to as being “connected” or “connected” to another component, it may be directly connected or connected to the other component, but another component may exist in between.

As used herein, the singular expression includes the plural expression unless the context clearly dictates otherwise. In the present application, terms such as "comprises" or "comprising", "have" and the like are intended to represent the presence of elements or combinations thereof described in the specification, and the possibility that other elements or features may be present or added. It is not precluded.

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a configuration diagram of an activity recommendation system according to an embodiment of the present invention, FIG. 2 is a configuration diagram of an artificial neural network model unit according to an embodiment of the present invention, and FIG. 3 is a hardware perspective view of an activity recommendation system according to an embodiment of the present invention 4 is an exemplary diagram of activity recommendation data according to an embodiment of the present invention.

The activity recommendation system (AS) of this embodiment combines various IT services including web services based on BPEL (Business Process Execution Language) or similar standard or non-standard process technology to design a business process model. It is recommended to probabilistically select the next activity based on their order and/or context.

A business process consists of one or more activities connected in sequence. The activity recommendation system (AS) of this embodiment recommends the next activity to the user (process designer) based on the information of the activities sequentially connected up to the present and the contextual information (eg, parallel connection state, activity properties) to the user (process designer) Probabilistic recommendations can improve process design efficiency. Context means not only sequential connection information of simple activities, but also various information such as parallel connection information, cancellation actions, and properties of activities. The activity recommendation system (AS) of the present embodiment considers the parallel connection information and the activity attribute together with the sequential connection information as a context for activity recommendation.

A number of techniques for designing a business process model by combining various IT services including web services based on BPEL (Business Process Execution Language) or similar standard or non-standard process technology are known, and detailed description thereof will be omitted. . In the description of the present embodiment, the process design unit 30 represents a system or function module for designing a business process model combining IT services.

The activity recommendation system (AS) of the present embodiment includes an activity recommendation service unit 10 and an artificial neural network model unit to probabilistically recommend an activity in the process of a user designing a business process model using the process design unit 30 . (20), and these elements provide the function of the model recommendation unit from a functional point of view.

The activity recommendation service unit 10 receives the parameter data generated based on the activity sequence combination from the process design unit 30, and inputs the parameter data to the learned artificial neural network model unit 20, The activity recommendation data obtained from the artificial neural network model unit 20 is returned to the process design unit 30 .

As an example, the activity recommendation service unit 10 is configured in a separate service form to facilitate the implementation and access to the request of the various process design unit 30, and is configured to be accessible in various methods and various languages such as Json and XML. it is preferable

Meanwhile, the activity recommendation service unit 10 may be located remotely from the process design unit 30 or may be included in one application. It differs depending on whether the entire system configuration is configured as a server structure or as a single application structure.

For example, if the process design unit 30 is implemented in the form of a web server, the activity recommendation service unit 10 is also configured in a web server method, and if the process design unit 30 is configured in an application method, the activity recommendation service unit 10 can be composed of a web server or a single application structure.

It is preferable that the activity recommendation service unit 10 of this embodiment can accommodate both of the above methods. For example, the remote server method is implemented in Json and XML based on HTTP so that it can be accessed from various client programming languages, and a single It is preferable that the application method be serviced by the API method so that the process design unit 30 can easily access it.

On the other hand, the process design unit 30 calls the activity recommendation service unit 10 based on the sequence data and context information of the currently designed activities to recommend the activity at the time the process is designed by the user (process designer). Thus, activity recommendation probability data is obtained, and activities are finally recommended to the user through the data.

The method of extracting the parameter data to be transmitted to the activity recommendation service unit 10 is, for example, based on the last and recently created activity, while traversing up to the starting activity in the reverse order, the type and order of the activities, their properties, whether they are parallel is extracted and vectorized.

The artificial neural network model unit 20 obtains activity recommendation data for recommending an activity of the next order of the order combination of the activities included in the parameter data in the form of a probability value.

4 is a diagram illustrating a scenario of activities recommended during process design, and illustrates a state in which four activities (D, E, F, G) recommended after activities A, B, and C are shown with probability. The user selects one of the one or more activities recommended by the activity recommendation system (AS) using the user interface means (eg, monitor, mouse, or keyboard) of the activity recommendation system (AS), Connect to the following sequence to complete the design process.

The artificial neural network model unit 20 will be described in more detail with reference to FIG. 2 .

The artificial neural network model unit 20 is a GRU-type recurrent neural network (RNN) including an input layer 21 , a Gated Recurrent Unit (GRU) cell 22 , a hidden layer 23 and an output layer 25 . ) is based on

The recurrent neural network is an artificial neural network useful for context-based prediction by analyzing sequential time series data. It is a neural network structure widely used for language transformation, etc. It is an artificial neural network with sequential memory as the N-1th output becomes the Nth input. . In this respect, the recurrent neural network is a technique suitable for the goal of the present invention having an ordinal context, and enables more efficient activity recommendation than other algorithms.

The GRU cell 22 of the recurrent neural network of this embodiment is a simplified version of an LSTM (Long Short-Term Memory) cell, and operates like an LSTM cell, and has a simple structure, so it is easy to implement and has the advantage of high speed.

In FIG. 2 , the N-1 th output of the GRU cell 22 is input to the next hidden layer 23 , and is also cycled to the input of the next GRU cell 22 .

The artificial neural network model unit 20 is configured to include a dropout layer 24 to avoid overfitting after the hidden layer 23 .

Dropout is a method that does not use any part of the artificial neural network in the learning process. For example, if the dropout ratio is set to 0.5, half neurons are not randomly used in each learning process, but only half neurons are used.

In general, dropout is used only for artificial neural network training and not for prediction. It prevents the artificial neural network from becoming too dependent on a specific neuron or a specific combination during training, and prevents overfitting by creating the same effect as using different neural networks as an ensemble because neurons are not used by random selection every time.

The artificial neural network model unit 20 of this embodiment increases the accuracy by placing the hidden layer 23 after the GRU cell 22, and then puts a dropout layer 24 to prevent overfitting of learning.

The artificial neural network model unit 20 uses a softmax function as an activation function of the output layer 25, and for one or more activities constituting the following sequence of sequence combinations of activities included in the parameter data. is configured to obtain each probability value. The softmax function is a function that receives a multidimensional vector corresponding to the number of classes to be classified and estimates the probability for each class. .

The input layer 21 of the artificial neural network model unit 20 is a training in which the largest number of activities among the data sets of the training data and the target data generated based on the sequence combination of activities included in the process data constitute the sequence combination The number of input neurons 21a is determined based on the data.

The output layer 25 of the artificial neural network model unit 20 includes the number of different types of activities constituting the target data included in the data set of the training data and the target data generated based on the sequence of activities included in the process data. The number of output neurons 25a is determined based on .

Each layer of the recurrent neural network consists of a plurality of neurons, and the number and main settings of the neurons may be configured as follows, for example. Such a configuration may be input or changed through an artificial neural network configuration unit 42 to be described later.

number hierarchy Number of neurons and their main settings One input layer Maximum number of connected activities in process data 2 GRU cell Default is 100 and can be changed when set separately 3 hidden layer Default is 200, and can be changed when set separately. 4 drop-out floor Rate is set to 0.5 by default and can be changed when set separately 5 output layer Number of different activity types

In the present embodiment, the number of neurons in the GRU cell 24 is configured by setting, but is set to 100 by default. The activation function of the GRU cell 10 can be variously designated, but is basically designated as a hyperbolic tangent (tanh) function. The input data is transmitted to the GRU cell 24, and is learned in time series.

The number of output neurons 25a of the output layer 24 is determined by the number of types of activities in the process, and the activation function uses a softmax function because the final output data is the probability of the recommended activity. The softmax function converts the values of the final output layer 24 into probabilities and outputs them.

The activity recommendation system (AS) of this embodiment includes an artificial neural network constructing unit 42, a data preprocessing unit 44, and an artificial neural network learning unit 40 to train the artificial neural network model unit 20, and these elements are It provides the function of the model learning unit from a functional point of view.

The above-described configuration in which the model recommendation unit and the model learning unit are divided and structured can develop, improve, and drive model learning and model recommendation separately, thereby increasing the efficiency of the activity recommendation system (AS). The operation timing of the model learning unit and the model recommendation unit is different. The model learning unit is involved in learning a new model or storing the learned model in a storage unit (eg, a non-volatile recording medium such as a database or disk) or loading it into the system's memory. It is used when there is training data and you want to retrain the model. The model recommendation unit is used when the process design unit 30 wants to receive an activity recommendation according to the user's needs without being involved in model training.

The artificial neural network configuration unit 42 sets at least one of the number of neurons required for the construction of the artificial neural network model unit 20 , the type of activation function, the size of the training epoch, and the type of the optimization method.

The number of neurons and the type of activation function can be set as in the example of Table 1 above.

There are various optimization methods for training the artificial neural network, and in this embodiment, Adam (Adopative Moment Estimating) is used, and the loss function uses least squares error, but is not limited thereto.

An epoch refers to a state in which forward propagation and back propagation have been completed for all data in an artificial neural network. For example, if the epoch is 50, it means that a total of 50 lessons are learned in terms of the entire data unit. If the epoch size is too small or too large, overfitting or underfitting can occur.

The data preprocessor 44 sequentially extracts text-based activity data included in the process data, and pre-processes it into number-based data to obtain the training data and target data.

Preferably, the training data is data related to a sequence combination of one or more activities included in the process data, and the target data is data related to an activity corresponding to a next sequence of sequence combinations of activities included in the training data.

Since the activity recommendation system (AS) of this embodiment basically recommends probabilistically selecting the next activity based on the order of the activities, the training data includes data on the order combination of one or more activities included in the process data. is composed

More preferably, the activity recommendation system (AS) of the present embodiment recommends probabilistically selecting the next activity based on the context of the activities including the parallel connection information and activity attribute information along with the order of the activities.

To this end, the training data includes a parallel connection activity included in the process data as an activity constituting the sequence combination.

In addition, the training data includes the properties of the activities included in the process data as activities constituting the sequence combination.

As another example, when one activity type included in the process data has two or more properties, the training data classifies the corresponding activity type as a separate activity type for each property to configure an order combination of activities.

The data preprocessor 44 receives process data, sequentially extracts activity data, and converts text into mapped numbers. In general, process data has sequential and parallel information of activities. In a process composed of N activities, data with N-1 sequence information is extracted and used based on the start activity.

For example, in data preprocessing, numeric mapping of activities starts from 0 as a positive number. However, since the number of activities to be trained is variable, in this embodiment, training data is set as data of a certain size, and the remaining space is padded with '-1' for learning. It is easier to configure and train than to remodel the structure of the recurrent neural network according to the data state, and convergence is also faster.

Activity extraction and pre-processing will be exemplarily described with reference to the bipel code in Table 2 below.

<process>
<sequence>
<receive/>
<assign/>
<invoke/>
<reply/>
</sequence>
</process>

The activity included in the bipel code of Table 2 may be extracted as follows.

- Preprocessing 1: Extraction: [receive, assign, invoke, reply]

- Preprocessing 2: mapping to numbers :　　[1, 2, 3, 4]

If the activity of the process data included in the bipel code of Table 2 is in the order of receive(1), assign(2), invoke(3), and reply(4), and the maximum number of activities included in the training process is 10, training Data and target data can be preprocessed as follows. (with -1 padding)

- Training data 1: 1,-1,-1,-1,-1,-1,-1,-1,-1,-1, target data: 2

- Training data 2: 1, 2,-1,-1,-1,-1,-1,-1,-1,-1, target data: 3

- Training data 3: 1, 2, 3,-1,-1,-1,-1,-1,-1,-1, target data: 4

For example, target data to be used for artificial neural network learning is input as one-hot-encoding as many as the number of activities. For example, if the total number of types of activities included in target data obtained from process data is 10, if the target data to be used for current learning is 3, [0,0,0,1,0,0,0,0,0 ,0] is converted into a vector of the form.

In general, process data includes not only order information of activities, but also context information on whether the activity is parallel to the flow and detailed attribute information of previous activities.

For example, information on parallelism and detailed attribute information of activities may be data preprocessed in a form similar to general activity information.

Activity extraction and pre-processing will be exemplarily described with reference to the bipel code of Table 3 below.

<process>
<parallel>
<sequence>
<receive/>
<reply/>
</sequence>
<sequence>
<invoke/>
<assign/>
</sequence>
</parallel>
</process>

- Pretreatment 1: Extraction:

[parallel, sequence, receive, reply]

[parallel, sequence, invoke, assign]

- Preprocessing 2: mapping to numbers :　 (when mapping numbers to receive(1), assign(2), invoke(3), reply(4), parallel(50), sequence(40))

[50, 40, 1, 4]

[50, 40, 3, 2]

The artificial neural network learning unit 40 obtains a data set of training data and target data generated based on a combination of activity sequences included in the process data provided by the process design unit 30, and inputs the data set. The artificial neural network model unit 20 is trained. Obtaining a data set of the training data and the target data may be obtained through the data preprocessing unit 44 as described above.

As training data and target data are input, batch learning is performed according to the batch processing set by the artificial neural network configuration unit 42 . The number of training is determined according to the setting of the number of epochs, and learning is performed.

In artificial neural network training, it is necessary to properly reconfigure the neural network and change other settings according to the accuracy and error rate. As for the final accuracy, since a recommended activity is not a fixed one, but a combination of various patterns of activities is possible, it is preferable to set a target of, for example, about 80%.

Learning of the artificial neural network is carried out with the input of the training data and the data set of the target data, and the training data is input and calculated with the weight values of the constructed artificial neural network to reduce the error between the final output value and the target data. According to the generated error value, the artificial neural network adjusts the weights of all neural networks using the backpropagation technique, and if training is repeated, these weight values are gradually learned to converge to the target value.

Referring to FIG. 3 , from a hardware point of view, the activity recommendation system AS of this embodiment includes a memory 2 that stores one or more instructions and a processor 4 that executes the one or more instructions stored in the memory 2 . It is a computing device in which the computer program stored in the medium is executed to execute the activity recommendation method of the present embodiment. The activity recommendation system AS of this embodiment may further include a data input/output interface 6 , a communication interface 8 , data display means 3 , and data storage means 5 .

5 and 6 are flowcharts of an activity recommendation method according to an embodiment of the present invention.

First, a process for learning the artificial neural network model unit 20 will be described.

In step PS1, the activity recommendation system AS obtains a data set of training data and target data generated based on a combination of activity sequences included in the process data provided by the process design unit 30 .

To this end, the activity recommendation system (AS), through the artificial neural network configuration unit 42, the number of neurons required to configure the artificial neural network model unit 20, the type of activation function, the size of the training epoch, the optimization method Set at least one of the types.

In addition, the activity recommendation system (AS) sequentially extracts the text-based activity data included in the process data through the data pre-processing unit 44, and pre-processes it into number-based data to obtain the training data and target data. to obtain Data preprocessing through the data preprocessing unit 44 may be performed in parallel with the artificial neural network configuration through the artificial neural network model unit 20 .

For example, the process data for obtaining a data set of training data and target data is obtained from the process data storage unit 50 in which information about a business process previously completed or in use by the process design unit 30 is stored. can do.

In step PS2, the activity recommendation system (AS) trains the artificial neural network model unit 20 by inputting the data set.

The trained artificial neural network model is stored in a storage unit (eg, a database, a non-volatile recording medium), and is loaded into the activity recommendation system (AS) for probabilistic recommendation of an activity using parameter data, as will be described later.

Next, a process of probabilistically recommending an activity in the process of a user designing a business process model using the process design unit 30 will be described.

In step S1 , the activity recommendation system AS receives the parameter data generated based on the activity sequence combination from the process design unit 30 .

The activity sequence combination refers to sequence combination information of activities sequentially connected up to now in a process being designed by a user (a process designer) using the process design unit 30 .

The parameter data includes various context information such as whether the activity is parallel with the order data and detailed properties of the activities, and may be understood as data in which the activity order combination is processed in the same data form as the training data by data preprocessing.

In step S2, the activity recommendation system (AS) inputs the parameter data to the trained artificial neural network model unit 20 .

To this end, the activity recommendation system (AS) loads the learned artificial neural network model stored in a storage unit (eg, a database, a non-volatile recording medium), and drives the artificial neural network model unit 20 .

In step S3, the activity recommendation system (AS) obtains activity recommendation data that recommends, in the form of a probability value, an activity of the next order of the sequence combination of activities included in the parameter data, using the artificial neural network model unit 20 . do.

To this end, the activity recommendation system AS inputs parameter data to the neural network model unit 20, and the neural network model unit 20 provides each recommendation probability of the activities based on the input parameter data.

In step S4 , the activity recommendation system AS returns the activity recommendation data obtained from the neural network model unit 20 to the process design unit 30 .

In this process, the activity recommendation service unit 10 formats and returns the recommendation probability of each provided activity according to the calling service, and the process design unit 30 shows the final delivered recommendation probability to the user through the user interface, let you choose

When the user selects an activity of the next order, the process design unit 30 combines the selected activity with a lower priority position of the activity connection currently being designed, and proceeds to the step of recommending and selecting the activity of the next order. When the working business process is completed and the user performs storage processing, information on the completed business process is stored in the process data storage unit 50 .

An activity recommendation method according to an embodiment of the present invention will be exemplarily described.

For example, in the loan application business process, in the case of the process of whether to proceed or cancel the loan application according to the activity of inquiring the credit limit of the applicant, the inquiry limit activity must be called first in order, so the sequential process corresponds to

As an example of a parallel process, the lowest loan interest inquiry process, for example, does not need to be in order when inquiring the loan interest of banks A, B, and C. It is a parallel process.

Regarding detailed properties, for example, the invoke activity is an activity that calls an external service. Depending on the type of the external call service, various IT services including web services, FTP, file, database, Hadoop, SAP, etc. There are detailed properties. Considering these detailed attributes, more accurate activity recommendation is possible. In the example of Table 4 below, the invoke activity has three detailed properties: rest, ftp, and webservice.

As above, the BPEL code of the loan application process including sequential and parallel elements and detailed attributes can be exemplified as shown in Table 4 below. As an example, this bipel code may be a code defining a business process previously completed in the process design unit 30 and stored in the process data storage unit 50, or may be a code input by a user for learning. .

<process name="Application for Loan">
<sequence name=""
<invoke name="Lending limit inquiry"binding=""
<if>
<then value=”Loan application>Loan limit”>
<terminate/>
</then>
</if>
<parallel>
<sequence>
<invoke target=”Bank A”binding=”rest” />
<assign source=”bank interest” target=”response document” />
</sequence>
<sequence>
<invoke target=”Bank B”binding=”ftp” />
<invoke target="database"/>
</sequence>
<sequence>
<invoketarget="BankC"binding="webservice"/>
<reply/>
</sequence>
</parallel>
</sequence>
</process>

When an order combination of activities for composing a data set of training data and target data is extracted with respect to the Bifel code of Table 4, it is configured as follows.

order Training data definition Target data definition One sequence invoke 2 sequence, invoke if 3 sequence, invoke, if then 4 sequence, invoke, if, then terminate 5 sequence, invoke, if parallel 6 sequence, invoke, if, parallel, sequence invoke 7 sequence, invoke, if, parallel, sequence, invoke-rest assign 8 sequence, invoke, if, parallel, sequence, invoke-ftp invoke 9 sequence, invoke, if, parallel, sequence, invoke-webservice reply

In the definition of the target data, the terminate activity is included only in the target data and is not included in the training data because the process is terminated in the corresponding activity and there is no recommended activity to be continued.

In the target data definition, the sequence activity following parallel is not included in the recommended activity. This is because, in the case of recommending an activity to be recommended after an activity, if the next activity is recommended and included without recommending the sequential activity for the sake of order, the sequence activity can be considered to be implicitly included.

For example, in the case of step 5, there is a meaning of invoking the loan limit system based on the request message of the service requester (invoke) and different actions according to the limit (if).

The invoke-rest, invoke-ftp, and invoke-webservice activities included in the training data of steps 7, 8, and 9 are the same activities as the target data invoke of step 6, but are separate activities according to the attributes (rest, ftp, webservice). As such, it is defined by dividing the training data.

In step 6, the target data is not divided into invoke-rest, invoke-ftp, and invoke-webservice activities, but only one invoke activity is set. This is because the training/target data policy was set in consideration of In other words, as the target data, the invoke activity is first recommended and the attribute is input (or recommended) separately. The data policy is set and reflected in the generation of training data. If the data policy is set such that three categories of invoke-rest, invoke-ftp, and invoke-webservice activities are recommended with the same probability, target data may be defined by dividing it into three categories.

As a variant, the data set of training data and target data may be configured as follows.

order Training data definition Target data definition One sequence invoke 2 sequence, invoke if 3 sequence, invoke, if then 4 sequence, invoke, if, then terminate 5 sequence, invoke, if parallel 6 sequence, invoke, if, parallel, sequence invoke 7 sequence, invoke, if, parallel, sequence, invoke, rest assign 8 sequence, invoke, if, parallel, sequence, invoke, ftp invoke 9 sequence, invoke, if, parallel, sequence, invoke, webservice reply

In Table 5, when one activity type included in the process data has two or more properties, the training data classifies the corresponding activity type as a separate activity type for each property to constitute an order combination of activities.

In comparison, in the example of Table 6, the training data is configured to directly include the attributes (rest, ftp, webservice) of the activity included in the process data as the activity constituting the sequence combination.

Since the extraction of training data varies depending on the recommended item or its environmental factors, one of the above two methods can be selected in consideration of the process design environment. In consideration of this, the artificial neural network configuration unit 42 sets the number of neurons required for the construction of the artificial neural network model unit 20 , the type of activation function, the size of the training epoch, the type of optimization method, and the like.

In the case of Table 5, the input layer 21 indicates that the number of activity sequences, invoke, if, parallel, sequence, and invoke-rest included in sequence 7 (or sequence 8, 9) with the largest number of training data is 6 Taking this into account, the number of input neurons 21a is set to six.

Since the output layer 25 is the number of types of activities constituting the target data, the output neuron 25a ) is set to 7.

As a result of the inventor's experiment, it was confirmed that the number of neurons in the GRU cell 22 converges best when there are about twice as many as the number of input neurons 21a in the input layer 21. In the case of the example, the number of neurons in the GRU cell 22 is defined as 12.

As a result of the present inventor's experiment, when the number of neurons in the GRU cell 22 is greater than twice the number of input neurons 21a in the input layer 21, convergence is similar and the execution speed is slow, and conversely, when the number of neurons is less than 2 times, the execution speed , but the convergence was confirmed to be slightly lower. As another example, when the number of input neurons 21a of the input layer 21 is 50, the number of neurons of the GRU cell 22 may be set to 100.

The number of neurons in the hidden layer 23 is defined as 24, which is twice the number of neurons in the GRU cell 22 .

As a result of the inventor's experiment, when the number of neurons in the hidden layer 23 is greater than twice the number of neurons in the GRU cell 22, convergence is similar and the execution speed is slow, and conversely, when it is less than 2 times, the execution speed increases, but the convergence is was found to be slightly lowered. As another example, when the number of neurons in the GRU cell 22 is 100, the number of neurons in the hidden layer 23 may be set to 200.

Batch processing and epoch settings vary depending on the error rate or accuracy during initial training of the artificial neural network and vary depending on the number of training data, but in general, batch processing can be set to 256 and epoch to 100.

The reason for batch processing is to converge well on correct answer data without being dependent on specific data. The value of batch processing is judged by looking at the number of training data and the accuracy or error rate during training. For example, if there are 100 training data and batch processing is 20, instead of training one by one, the artificial neural network is trained in blocks of 20 out of 100.

The typical setting value of RATE of dropout is 0.5. If you increase this value, the original data may be lost, and if it is set low, dropout may not be applied.

As an example, when a user who designs a business process using the process design unit 30 receives an activity recommendation using the artificial neural network model unit 20 trained with the data set of Table 5, the order of the currently designed activity If the combination is sequence-invoke-if, the “parallel” activity is recommended with the highest probability as shown in the target data of sequence 5 in Table 5.

As described above, the present invention provides a recommendation function as an auxiliary function used in process design, reduces errors in process design, and increases the speed of completion of process design. In particular, the present invention uses a recurrent neural network showing good performance in time series analysis and prediction to enable contextual recommendation for sequential and/or parallel activities of a process, rather than simply targeting the final activity information. In addition, since it is configured so that a user (process designer) can select a number of recommended activities by probabilistic recommendation, the design efficiency is improved, and the auxiliary utility in process design is high.

Embodiments of the present invention include a program for performing various computer-implemented operations and a computer-readable recording medium recording the same. The computer-readable recording medium may include program instructions, data files, data structures, etc. alone or in combination. The medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include hard disks, magnetic media such as floppy disks and magnetic tapes, optical recording media such as CD-ROMs, DVDs, and USB drives, magneto-optical media such as floppy disks, and ROM, RAM; Included are hardware devices specially configured to store and execute program instructions, such as flash memory and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

10: Activity Recommendation Service Department
20: artificial neural network model unit
21: input layer
22: Gated Recurrent Unit (GRU) cell
23: hidden layer
24: Dropout floor
25: output layer
30: Process Design Department
40: artificial neural network learning unit
42: artificial neural network component
44: data preprocessor
AS: Activity recommendation system

Claims (16)

Receive parameter data generated based on the combination of activity sequences of the business process being designed from the process design unit, input the parameter data to the trained artificial neural network model unit, and activity recommendation data obtained from the artificial neural network model unit an activity recommendation service unit returning to the process design unit; and
and an artificial neural network model unit for obtaining activity recommendation data that recommends an activity in the next order of the sequence combination of activities included in the parameter data in the form of a probability value based on an activation function;
The activity sequence combination of the business process being designed above is,
The activity recommendation system using an artificial neural network, characterized in that the process design unit receiving the activity recommendation data returned includes an activity sequence combination in which the activities of the next order selected through a user interface are combined with a lower priority position.
According to claim 1,
An artificial neural network learning unit that obtains a data set of training data and target data generated based on a combination of activity sequences included in the process data provided by the process design unit, and inputs the data set to learn the artificial neural network model unit; An activity recommendation system using an artificial neural network that includes
3. The method of claim 2,
An artificial neural network constructing unit configured to set at least one of the number of neurons, the type of activation function, the size of the training epoch, and the type of optimization required for the construction of the artificial neural network model unit; Activity recommendation system.
3. The method of claim 2,
A data pre-processing unit that sequentially extracts text-based activity data included in the process data and pre-processes it into number-based data to obtain the training data and target data; Activity recommendation system.
3. The method of claim 2,
The training data is data about an order combination of one or more activities included in the process data,
The target data is an activity recommendation system using an artificial neural network, characterized in that the data on the activity corresponding to the next order of the sequence combination of the activities included in the training data.
6. The method of claim 5,
The training data is an activity recommendation system using an artificial neural network, characterized in that it includes a parallel connection activity included in the process data as an activity constituting the sequence combination.
6. The method of claim 5,
The training data is an activity recommendation system using an artificial neural network, characterized in that it includes the properties of the activity included in the process data as activities constituting the sequence combination.
6. The method of claim 5,
When one activity type included in the process data has two or more attributes,
The training data is an activity recommendation system using an artificial neural network, characterized in that by classifying the corresponding activity type as a separate activity type for each attribute, an order combination of activities is configured.
According to claim 1,
The artificial neural network model unit,
An activity recommendation system using an artificial neural network, characterized in that it is based on a Recurrent Neural Network (RNN) of the GRU method including an input layer, a Gated Recurrent Unit (GRU) cell, a hidden layer, and an output layer.
10. The method of claim 9,
The artificial neural network model unit,
Activity recommendation system using an artificial neural network, characterized in that it further comprises a dropout layer to avoid overfitting after the hidden layer.
10. The method of claim 9,
The artificial neural network model unit,
As the activation function of the output layer, the Softmax function is used,
Activity recommendation system using an artificial neural network, characterized in that it is configured to obtain respective probability values for one or more activities constituting the next sequence of the sequence combination of activities included in the parameter data.
10. The method of claim 9,
The input layer of the artificial neural network model unit,
Artificial, characterized in that the number of input neurons is determined based on the training data in which the largest number of activities among the data sets of the training data and the target data generated based on the sequence combination of activities included in the process data constitute the sequence combination Activity recommendation system using neural network.
10. The method of claim 9,
The output layer of the artificial neural network model unit,
Artificial, characterized in that the number of output neurons is determined based on the number of types of different activities constituting the target data included in the data set of the training data and the target data generated based on the sequence of activities included in the process data Activity recommendation system using neural network.
An activity recommendation method executed in an activity recommendation system comprising a memory for storing one or more instructions and a processor for executing the one or more instructions stored in the memory, the method comprising:
1) receiving the parameter data generated based on the sequence of activities of the business process being designed from the process design unit;
2) inputting the parameter data into a trained artificial neural network model unit;
3) using the artificial neural network model unit, obtaining activity recommendation data for recommending an activity in the next order of the sequence combination of activities included in the parameter data in the form of a probability value based on an activation function; and
4) returning the activity recommendation data obtained from the artificial neural network model unit to the process design unit;
The activity sequence combination of the business process being designed above is,
The activity recommendation method using an artificial neural network, characterized in that the process design unit receiving the activity recommendation data returned includes an activity sequence combination in which the activities of the next order selected through a user interface are combined with a lower priority position.
15. The method of claim 14,
Before step 1) above,
P1) obtaining a data set of training data and target data generated based on a combination of activity sequences included in the process data provided by the process design unit; and
P2) learning the artificial neural network model unit by inputting the data set; activity recommendation method using an artificial neural network configured to further include a.
A computer-readable recording medium recording a program for executing each step of the activity recommendation method using an artificial neural network according to claim 14 or 15.

Images