JP7841897B2 - Matching device and program - Google Patents

Description

This invention relates to a matching device and program.

Companies that conduct project-based work often temporarily bring in personnel from other companies as partners to address labor shortages or acquire technologies they do not possess internally. In particular, highly skilled partners often play a central role in projects, and maintaining long-term collaboration with such partners, rather than just temporary partnerships, is a crucial management challenge.

To maintain long-term collaboration with partners, it's necessary to go beyond use within specific departments and collaborate company-wide across departments. However, cross-departmental collaboration with partners is currently not progressing. This is because the required skills and customer characteristics differ from department to department, making it difficult to determine whether partners belonging to different departments are suitable for one's own department's projects.

As a technique for predicting the degree of compatibility between partners and projects, Patent Document 1 discloses a technique for learning a method for calculating the degree of compatibility from past partner-project matching results.

Specifically, a deep neural network (DNN) is used to extract features useful for calculating the degree of fit from the attribute values of the partner and project, and the degree of fit between the partner and project is calculated based on these features. By training the parameters so that the degree of fit predicted by the DNN matches the actual degree of fit for partner and project matching results from past matching data, accurate prediction of the degree of fit becomes possible.

Patent No. 6369053

However, conventional technologies faced the challenge of accurately predicting the compatibility between partners belonging to one department and projects belonging to different departments.

To achieve this using the technology described in Patent Document 1, it is necessary to have matching data between partners belonging to different departments and projects, which would serve as training data with similar characteristics to the target of prediction. However, given the current state of partner collaboration, such matching data is scarce, and a sufficient amount of training data cannot be obtained. Therefore, it is difficult to accurately predict the degree of fit between partners belonging to different departments and projects using the prior art.

This invention was made to solve these problems and aims to provide a matching device and program that can more accurately predict the degree of compatibility between partners belonging to different departments and projects.

An example of a matching device according to the present invention is:
A matching device that predicts the degree of compatibility between a partner and a project,
The matching device comprises a learning means and a prediction means,
The learning means generates a trained model by performing machine learning processing based on partner attribute values, project attribute values, partner correct answer category, project correct answer category, and correct answer fit.
The aforementioned machine learning process is
- A first learning process that takes the partner attribute value and the project attribute value as input and learns to output a predicted fit that is closer to the correct fit,
- A second learning process that takes the aforementioned partner attribute values and project attribute values as input and learns a common prediction method that can be used to predict the degree of fit of any department,
Includes,
The prediction means uses the trained model to predict the degree of fit based on information about the partner and information about the project.

In one example, the first learning process is:
A process to calculate partner features based on the aforementioned partner attribute values,
A process for calculating project features based on the aforementioned project attribute values,
A process for calculating the prediction fit based on the partner features and project features,
A process to learn so that the accuracy error, which is the difference between the correct accuracy and the predicted accuracy, becomes smaller,
Includes.

In one example, the second learning process is:
Based on the aforementioned partner features, a process is performed to predict the partner prediction department to which the partner belongs.
A process to calculate the partner division error as the error between the partner prediction division and the partner correct answer division,
A process to learn so that the aforementioned partner division error becomes larger,
Includes.

In one example, the second learning process is:
Based on the aforementioned project features, a process is performed to predict the project forecasting department to which the project belongs.
The process of calculating the project department error as the difference between the project prediction department and the project correct answer department,
The process involves learning to increase the error in the aforementioned project division,
Includes.

In one example, the partner feature and the project feature are represented by vectors.
The aforementioned machine learning process is performed using a deep neural network.

In one example, the prediction fit is calculated using the squared error or cosine similarity.

In one example, the partner prediction sector is represented by a vector that indicates the likelihood of the partner belonging to each sector.
The correct answer category for the partner is represented by a vector where the likelihood of the partner belonging to that category is 1, and the likelihood of the partner belonging to another category is 0.
The aforementioned partner division error is calculated using the cross-entropy error function.

In one example, the machine learning process is:
- A first deep neural network that takes the aforementioned partner attribute values as input and outputs the aforementioned partner features,
- A second deep neural network that takes the aforementioned project attribute values as input and outputs the aforementioned project features,
- A third deep neural network that takes the aforementioned partner features as input and outputs the partner prediction department to which the partner belongs,
- A fourth deep neural network that takes the aforementioned project features as input and outputs the project prediction department to which the project belongs,
It is done using
In the first learning process, the first deep neural network and the second deep neural network are trained.
In the second learning process, the first deep neural network, the second deep neural network, the third deep neural network, and the fourth deep neural network are trained.

In one example, in the first learning process,
- Update the parameters of the first deep neural network to reduce the aforementioned goodness-of-fit error,
- The parameters of the second deep neural network are updated to reduce the aforementioned goodness-of-fit error,
In the second learning process described above,
- Update the parameters of the first deep neural network to increase the partner division error,
- Update the parameters of the second deep neural network to increase the error in the aforementioned project division,
- Update the parameters of the third deep neural network to reduce the aforementioned partner division error,
- Update the parameters of the fourth deep neural network to reduce the project division error.

An example of a program according to the present invention causes a computer to function as the matching device described above.

According to this invention, the degree of compatibility between partners belonging to different departments and a project can be predicted more accurately.

A functional block diagram of a matching device 100 according to Embodiment 1 of the present invention. Hardware configuration of the matching device 100. A schematic flowchart illustrating the operation of the matching device 100. Example configuration of partner table T1. Example configuration for project table T2. Example configuration of matching results table T3. Overview of the learning process by the learning method 101. An overview of the effects of common feature learning processing. A mathematical formula that explains the details of the calculation process performed by the learning method 101. A diagram illustrating the details of the calculation process performed by the learning method 101. Flowchart of the operation of learning method 101. Example of the structure of the learning result table T4. Example configuration of the prediction target table T5. Flowchart of the operation of the prediction means 102. Example of the structure of prediction result table T6.

First, the outline of the present invention will be described. This specification discloses a matching device that can accurately predict the degree of fit between partners and projects belonging to different departments. This is achieved by excluding department-specific characteristics from the matching results of partners and projects belonging to the same department, for which a sufficient amount of training data is available, and by learning a common prediction method that can be used to predict the degree of fit for any department.

Specifically, the model is trained so that it can no longer predict the department to which a partner or project belongs based on its features, and then the goodness-of-fit is predicted based on those features. The inability to predict the department means that partners and projects from different departments share the same features, indicating that these are common features across all departments.

The embodiments of the present invention will be described in detail below with reference to the drawings. The following example illustrates how a partner procurement officer for a project selects partners to be used in that project from a list of partners used within the company, using one embodiment of the present invention.

[Embodiment 1]
Figure 1 shows a functional block diagram of a matching device 100 according to Embodiment 1 of the present invention. The matching device 100 is a device that predicts the degree of compatibility between a partner and a project. A partner refers to, for example, an individual, and a project refers to, for example, a group of individuals organized within a company for a certain purpose. However, a partner does not necessarily have to be an individual within a company, and a project does not necessarily have to be a group within a company.

The matching device 100 includes a learning means 101 and a prediction means 102.

Figure 2 shows the hardware configuration of the matching device 100. The matching device 100 has a hardware configuration similar to that of a known computer, and includes an input device 103, an output device 104, an arithmetic unit 105, and a storage device 106.

The input device 103 includes a keyboard, mouse, network interface, etc. The output device 104 includes a display device (such as a display), printing device (such as a printer), network interface, etc. The arithmetic unit 105 includes, for example, a processor. The storage device 106 includes, for example, storage media such as a semiconductor memory device and a magnetic disk device. The storage device 106 may include main memory and auxiliary storage.

The storage device 106 stores the Partner Table T1, Project Table T2, Matching Performance Table T3, Learning Results Table T4, Prediction Target Table T5, and Prediction Results Table T6. The contents of these tables will be described later.

The storage device 106 may store a program (not shown). The arithmetic unit 105 may execute this program, causing the computer to function as a matching device 100. In other words, this program may cause the computer to function as a matching device 100.

Figure 3 shows a schematic flowchart illustrating the operation of the matching device 100. In step S1, the matching device 100 starts processing. In step S2, the matching device 100 decides whether to operate in learning mode or prediction mode. This decision is made, for example, in response to instructions from the user, such as a procurement officer.

When the matching device 100 operates in learning mode, in step S3, the learning means 101 of the matching device 100 learns how to calculate the degree of fit between the partner and the project based on past matching results.

When the matching device 100 operates in prediction mode, in step S4, the prediction means 102 of the matching device 100 predicts the degree of fit between the user-specified partner and the project based on the degree of fit calculation method learned during the learning process. Note that the processing in step S4 is performed using the learning result table T4, but since the learning result table T4 does not contain data in its initial state, the matching device 100 must operate in learning mode (step S3) at least once.

After steps S3 and S4, the process ends in step S5.

<Learning process in Embodiment 1>
Step S3 will be explained in more detail below. In the learning process of Step S3, the method for calculating the degree of fit between partners and projects is learned based on past matching results.

Before executing the learning process, the user stores the necessary data in the Partner Table T1, Project Table T2, and Matching Performance Table T3 using the input device 103.

Figure 4 shows an example of the structure of Partner Table T1. Partner Table T1 stores teacher-partner information related to partners. Teacher-partner information is information related to partners. In the example in Figure 4, the key is information identifying the partner (Partner ID), and it associates information such as the department to which the partner belongs, whether the partner has skills (C language skills and Java language skills), the partner's personality (whether they are cheerful or not, and whether they are calm or not), the partner's employment rate, and the partner's years of experience.

In this embodiment, the teacher partner information includes, as mandatory items, a partner ID and information representing the department. Other items can be modified as appropriate by those skilled in the art, and it is preferable to include various items other than those shown in the illustration.

Teacher partner information is preferably in a format suitable for learning. Since a DNN is used in this embodiment, teacher partner information is preferably numerical data that can be input into the DNN. In the example in Figure 4, for explanatory purposes, "department" is represented as text data, but in reality, departments are also represented as numerical data.

Of the information shown in Figure 4, the presence or absence of C language skills, the presence or absence of Java language skills, whether or not the person has a cheerful personality, whether or not they have a calm personality, the employment rate, and the number of years of experience represent attribute values related to the partner.

Figure 5 shows an example of the structure of Project Table T2. Project Table T2 stores teacher project information related to the project. Teacher project information is information related to the project. In the example in Figure 5, the key is information identifying the project (Project ID), and it associates information such as the department to which the project belongs, the necessity of partner technology in the project (C language technology and Java language technology), the characteristics of the customer related to the project (new customer or returning customer), and the personality of the employee related to the project (cheerful or not, and calm or not). Note that "employee" may include partners, or it may be defined regardless of whether they are partners or not.

In this embodiment, the teacher project information includes, as mandatory items, a project ID and information representing the department. Other items can be modified as appropriate by those skilled in the art, and it is preferable to include various items other than those shown in the illustration.

Of the information shown in Figure 5, the following represent attribute values related to the project: whether C language skills are required, whether Java language skills are required, whether the customer is new or not, whether the customer is a returning customer or not, whether the employee has a cheerful personality or not, and whether the employee has a calm personality or not.

Teacher project information is preferably in a format suitable for learning. Since a DNN is used in this embodiment, teacher project information is preferably numerical data that can be input into the DNN. In the example in Figure 5, for explanatory purposes, "department" is represented as text data, but in reality, departments are also represented as numerical data.

Figure 6 shows an example of the structure of the matching performance table T3. The matching performance table T3 uses pairs of current or past projects and the partners belonging to those projects as keys, and stores the degree of fit (correct fit) between the project and the partner for each pair. The correct fit is represented, for example, as a real number between 0 and 1, with a higher value indicating a better fit between the project and the partner.

For example, the first row of Figure 6 indicates that the partner with partner ID "PTN0001" has belonged to the project with project ID "PRJ0001," and that the compatibility score entered by the person in charge (e.g., the user of matching device 100) for these pairs was 0.9.

Figure 7 shows an overview of the learning process by the learning means 101. In this embodiment, the learning means 101 uses a DNN to learn how to calculate the goodness of fit between partners and projects based on past matching results. The DNN-related processing includes feature calculation, goodness of fit calculation, goodness of fit error calculation, departmental calculation, and departmental error calculation.

In Figure 7, the white squares represent scalar values, and the figure formed by three connected white squares represents a vector value.

The input to the DNN is information (attribute values) stored in partner table T1 and project table T2. Since it is difficult to predict the degree of fit by directly comparing attribute values, the feature extraction calculation process converts the information (features) into a format suitable for predicting the degree of fit, based on multiple attribute values.

The goodness-of-fit calculation process calculates the goodness-of-fit for a pair by comparing the features of the partner and the project. The goodness-of-fit error calculation process calculates the error (goodness-of-fit error) between the goodness-of-fit predicted in the goodness-of-fit calculation process and the correct goodness-of-fit stored in the matching performance table T3. By updating the parameters of the feature extraction calculation process to minimize this error, the learning means 101 learns how to calculate the goodness-of-fit.

The departmental calculation process and departmental error calculation process are designed to learn a common prediction method that can be used to predict the degree of fit for any department, by excluding department-specific characteristics from the matching history of partners and projects belonging to the same department.

The departmental calculation process and the departmental error calculation process will be collectively referred to as the "common feature learning process." The departmental calculation process predicts the department to which each partner belongs by extracting departmental-specific features from the partner and project features. The departmental error calculation process calculates the error (departmental error) between the department predicted in the departmental calculation process and the correct department stored in the partner table T1 or project table T2. By updating the parameters of the feature calculation process to increase this error (so that departments cannot be predicted from the features), features that exclude departmental-specific features are learned. As a result, common features that can be used to calculate the goodness-of-fit for any department are learned, enabling accurate prediction of the goodness-of-fit for partners and projects in different departments. Furthermore, the parameters of the departmental calculation process are also sequentially updated to extract departmental-specific features from the features at each learning stage. Specifically, the parameters are updated to reduce the departmental error.

Figure 8 illustrates the general effect of the common feature learning process. In Figure 8, the white figures and the hatched figures represent partners belonging to different departments.

Figure 8(a) is a conceptual diagram showing the case where common feature learning is not applied. Generally, most of the paired data in the matching performance table T3 consists of partner and project pairs belonging to the same department. Therefore, features with different distributions for each department are learned, allowing for accurate prediction of the goodness of fit only for pairs within each department.

Figure 8(b) is a conceptual diagram showing the case when common feature learning is applied. By learning the features in such a way that departments cannot be predicted from the features alone, department-specific features are removed, and the distribution of departmental features overlaps (they become similar in value). In this case, in the feature space, features of partners in different departments are placed near the features of a partner in a given department, meaning that features common to multiple departments are extracted.

The learning process by the learning means 101 is described below. The learning means 101 uses four DNNs and generates a trained model by performing machine learning processing based on partner attribute values, project attribute values, partner ground truth departments, project ground truth departments, and ground truth fit. While four DNNs are used in this embodiment, the number of DNNs can be appropriately changed by those skilled in the art. For example, if the partner or project attribute values obtained differ for each department, a DNN for feature calculation processing may be created for each department.

The details of the computation process by the learning means 101 will be explained using the mathematical formulas shown in Figure 9 and the diagram shown in Figure 10. In the following, a partner is represented by i ∈ S _ptn and a project is represented by j ∈ S _prj . S _ptn is the set of partners, and S _prj is the set of projects. t ∈ {ptn, prj} is an index representing a partner (ptn) or a project (prj).

The attribute values of partner i are represented by x _i (a vector with dimensions equal to the number of elements in partner's attribute values), and the attribute values of project j are represented by x _j (a vector with dimensions equal to the number of elements in project's attribute values).

The learning means 101 comprises a partner feature calculation DNN 201 (first deep neural network), a project feature calculation DNN 202 (second deep neural network), a partner division calculation DNN 203 (third deep neural network), and a project division calculation DNN 204 (fourth deep neural network). The partner feature calculation DNN 201 and project feature calculation DNN 202 are fully connected DNNs, where n ∈ {1, ..., N} represents the layer number and N is the number of layers. The partner division calculation DNN 203 and project division calculation DNN 204 are fully connected DNNs, where m ∈ {1, ..., M} represents the layer number and M is the number of layers.

In the feature calculation process, the partner feature calculation DNN201 calculates the partner feature y _i ^N based on the attribute value x _i of partner i using equations (1) and (2). y _i ^N is a vector with dimensions equal to the number of elements in the partner feature. _{A ptn} is the weighting coefficient for the inter-neuron connections of the partner feature calculation DNN201, where A _ptn ⁰ is a matrix of [number of elements in partner feature] x [number of elements in partner attribute value] dimensions, and A _ptn ⁿ is a matrix of [number of elements in partner feature] x [number of elements in partner feature] dimensions. _{a ptn} (i.e., a _ptn ⁰ and a _ptn ⁿ ) are the biases for each layer and are vectors with dimensions equal to the number of elements in the partner feature. LeakyReLU is a leaky normalized linear unit.

By calculating equation (2) for each layer, the partner feature y _i ^N , which is used to predict the degree of fit, is calculated as the output. Thus, the partner feature calculation DNN201 takes the attribute value x _i of partner i as input and outputs the partner feature y _i ^N.

Similarly, using the formulas obtained by changing the subscript i to j in formulas (1) and (2), the project feature calculation DNN202 calculates the project feature y _j ^N based on the attribute value x _j of project j. y _j ^N is a vector with dimensions equal to the number of elements in the project feature. A _prj is the weighting coefficient for the interneuron connections of the project feature calculation DNN202, A _prj ⁰ is a matrix of the dimension [number of elements in the project feature] x [number of elements in the project attribute value], and A _prj ⁿ is a matrix of the dimension [number of elements in the project feature] x [number of elements in the project feature]. a _prj (i.e., a _prj ⁰ and a _prj ⁿ ) are the biases of each layer and are vectors with dimensions equal to the number of elements in the project feature.

By calculating equation (2) for each layer, the project feature vector y _j ^N , which is used to predict the degree of fit, is calculated as the output. Thus, the project feature vector calculation DNN202 takes the attribute value x _j of project j as input and outputs the project feature vector y _j ^N.

It is preferable to configure the number of elements in each feature vector and the number of layers in each DNN so that they can be arbitrarily changed from the outside.

In the goodness-of-fit calculation process, the learning means 101 uses equation (3) to calculate the predicted goodness-of-fit p _ij from the feature pairs (y _i ^N , y _j ^N ) of partner i and project j, as the predicted goodness-of-fit for that pair. The predicted goodness-of-fit p _ij is a real number (scalar value). MSE is the squared error function, and tanh is the hyperbolic function. The predicted goodness-of-fit p _ij is calculated as a value between 0 and 1, where 1 means high goodness-of-fit and 0 means low goodness-of-fit.

In this embodiment, the prediction fit p _ij is calculated using the squared error, but it may also be calculated using the cosine similarity between (y _i ^N , y _j ^N ), or other calculation methods may be used.

In the goodness-of-fit error calculation process, the learning means 101 calculates the goodness-of-fit error e ij ^evl as the error between the predicted goodness-of-fit p _ij and the correct goodness-of-fit ^p _ij stored in the matching performance table _T3 , using equation (4). For the sake of notation, the hat symbol ^ is placed to the left of p, but the hat symbol ^ should correctly be placed above p as shown in equation (4). The same applies hereafter.

The goodness-of-fit error e _ij ^evl is a real number (scalar value). By updating the parameters of partner feature computation DNN201 and project feature computation DNN202 so that the goodness-of-fit error e _ij ^evl is small, the calculation method for calculating the predictive goodness-of-fit p _ij is learned as the parameters of each DNN.

Thus, the machine learning process by the learning means 101 includes a process (first learning process) that takes partner attribute values and project attribute values as input and learns to output a predicted degree of fit that is closer to the correct degree of fit. More specifically, the first learning process includes the following:
- A process to calculate the partner feature quantity y _i ^N based on the attribute value x _i of partner i.
- A process to calculate project features y _j ^N based on the attribute values x _j of project j.
- A process to calculate the goodness of fit _pij based on partner features y _i ^N and project features y _j ^N.
- A process to train the partner feature calculation DNN201 and project feature calculation DNN202 so that the fitness error e _ij ^evl , which is the difference between the correct fitness score ^p _ij stored in the matching performance table T3 and the predicted fitness score p _ij , becomes smaller.

In this embodiment, the first learning process involves learning the partner feature calculation DNN 201 and the project feature calculation DNN 202, but not the partner division calculation DNN 203 or the project division calculation DNN 204. (Note that in a modified example, the partner division calculation DNN 203 and the project division calculation DNN 204 may also be learned.)

In the departmental calculation process, the partner departmental calculation DNN203 calculates the predicted partner department _qi based on the partner features y _i ^N using equations (5) to (8). _qi is a vector with dimensions equal to the number of departments, where each element represents the probability or likelihood that partner i belongs to each department. _{B ptn} is the weighting coefficient for the inter-neuron connections of the partner departmental calculation DNN203, B _ptn ^m is a matrix of [number of partner feature elements] x [number of partner feature elements] dimensions, and B _ptn ^M+1 is a matrix of [number of departments] x [number of partner feature elements] dimensions. _{b ptn} is the bias for each layer, in particular b _ptn ^m is a vector with dimensions equal to the number of partner feature elements, and b _ptn ^M+1 is a vector with dimensions equal to the number of departments. Softmax represents the softmax function.

Thus, the partner division calculation DNN203 takes the partner feature quantity y _i ^N as input and outputs the partner division prediction q _i .

Similarly, using formulas obtained by changing the subscript i to j in equations (5) to (8), the project department calculation DNN204 calculates project predicted departments q _j based on the project features y _j ^N. q _j is a vector with dimensions equal to the number of departments, where each element represents the probability or likelihood that project j belongs to each department. B _prj is the weighting coefficient for the inter-neuron connections of the project department calculation DNN204, B _prj ^m is a matrix of [number of project feature elements] dimensions × [number of project feature elements] dimensions, and B _prj ^M+1 is a matrix of [number of departments] dimensions × [number of project feature elements] dimensions. _{b prj} is the bias for each layer, in particular b _prj ^m is a vector with dimensions equal to the number of project feature elements, and b _prj ^M+1 is a vector with dimensions equal to the number of departments.

Thus, the project division calculation DNN204 takes project features y _j ^N as input and outputs project division prediction q _j .

In the division error calculation process, the learning means 101 calculates the partner division error e i dpt as the error between the predicted division for partner i (partner predicted division q _i ) and the correct division (partner correct division ^q _i ) using ^equation (9 ₎ . Partner correct division ^q _i is a one-hot vector representing the division to which partner i belongs (a vector where the probability or likelihood of partner i belonging to that division is 1, and the probability or likelihood of partner i belonging to another division is 0). CrossEntropy represents the cross-entropy error function.

Similarly, the learning means 101 uses a formula obtained by changing the subscript i in formula (9) to j to calculate the project department error e _j ^dpt as the error between the predicted department for project j (project predicted department q _j ) and the correct department (project correct department q _j ). The project correct department q _j is a one-hot vector representing the department to which project j belongs (a vector where the probability or likelihood of project j belonging to that department is 1, and the probability or likelihood of project j belonging to another department is 0).

The learning means 101 updates the parameters of the partner feature calculation DNN 201 and the project feature calculation DNN 202 to increase the partner department error e _i ^dpt and the project department error e _j ^dpt . Through this learning process, only the features common to the departments can be efficiently learned.
Furthermore, the learning means 101 updates the parameters of the partner department calculation DNN 203 and the project department calculation DNN 204 to reduce the partner department error e _i ^dpt and the project department error e _j ^dpt . Through this learning process, the partner department calculation DNN 203 and the project department calculation DNN 204 become able to extract features unique to each department from the features at each stage of learning.

As described above, the partner feature calculation DNN201 is updated in two types of learning processes: one that reduces the goodness-of-fit error e _i j ^{e vl} , and another that increases the partner department error e _i ^{d pt} . Similarly, the project feature calculation DNN202 is updated in two types of learning processes: one that reduces the goodness-of-fit error e _{i j} ^{e vl} , and another that increases the project department error e _j ^{d pt} . Therefore, in order to adjust the influence of learning on goodness-of-fit and learning on departments, the calculation of the department error is multiplied by a weight w (real number), as shown in equation (9).

Thus, the machine learning process by the learning means 101 takes partner attribute values and project attribute values as input and includes a process (second learning process) that makes predictions between the department to which the partner belongs and the department to which the project belongs more difficult. More specifically, the second learning process includes the following:
- A process to predict the partner prediction department _qi to which partner i belongs, based on partner feature y _i ^N.
- A process to calculate the partner division error e _i ^dpt as the error between the partner prediction division q _i and the partner correct answer division q _i .
- A process to train the partner feature computation DNN201 so that the partner division error e _i ^dpt becomes larger.
- A process to train the partner division calculation DNN203 so that the partner division error e _i ^dpt becomes smaller.
- A process to predict the project forecasting department _qj to which project j belongs, based on project features _yj ^N.
- A process to calculate the project department error _ej ^dpt as the error between the project prediction department _qj and the project correct answer department _qj .
- A process to train the project feature calculation DNN202 so that the project division error _ej ^dpt becomes larger.
- A process to train the project division calculation DNN204 so that the project division error e _j ^dpt becomes smaller.

In this embodiment, in this second learning process, the calculation method for partner features y _i ^N and project features y _j ^N can be modified to exclude features specific to each department, thereby learning a common prediction method that can be used to predict the degree of fit for any department. This allows for learning that can more accurately predict the degree of fit for partners and projects belonging to different departments.

Figure 11 shows a flowchart of the operation of the learning means 101. This flowchart illustrates the details of step S3 in Figure 3.

In step S11, the learning means 101 starts processing. Next, in step S12, the learning means 101 obtains partner and project pairs and the degree of fit (correct fit^p _ij ) of those pairs from the matching performance table T3.

Next, in step S13, the learning means 101 obtains the attribute values x _i for the partner and the attribute values x _j for the project from the partner table T1 and the project table T2. It also obtains the correct departments to which each belongs (correct departments), namely the partner correct department^q _i and the project correct department^q _j . These correct departments are then processed into one-hot vectors.

Next, in step S14, the learning means 101 initializes the parameters of each DNN. For example, the initial values of all parameters may be determined using uniformly distributed random numbers between -1 and 1, but this method is not limited to this.

Next, the learning means 101 repeats the process in steps S15 to S23 a predetermined number of times. Preferably, the number of repetitions is configured to allow the user of the matching device 100 to input or change it. In step S15, the learning means 101 starts the repetition unit.

In step S16, the learning means 101 calculates partner feature quantity y _i ^N and project feature quantity y _j ^N based on the partner attribute value x _i and project attribute value x _j obtained in step S13.

Next, in step S17, the learning means 101 calculates the prediction fit p _ij for the pair (i, j) based on the partner feature y _i ^N and project feature y _j ^N calculated in step S16.

Next, in step S18, the learning means 101 uses a goodness-of-fit error calculation process to calculate the goodness-of-fit error e _ij ^evl from the predicted goodness-of-fit p _ij calculated in step S17 and the correct goodness-of-fit p _ij obtained in step S12.

Next, in step S19, the learning means 101 uses the departmental prediction process to calculate the predicted partner department q _i and the predicted project department q _j based on the partner feature quantity y _i ^N and the project feature quantity y _j ^N calculated in step S16.

Next, in step S20, the learning means 101 uses a departmental error calculation process to calculate the partner department error e _i dpt and the project department error e _j ^dpt based on the partner predicted department q _i and project predicted department _{q j calculated} in step ^S19 and the partner correct department q _i and project correct department q j obtained in step S13.

Next, in step S21, the learning means 101 calculates the gradient of each DNN parameter based on the goodness-of-fit error e _{i j} ^evl calculated in step S18 and the partner division error e _i ^dpt and project division error e _j ^dpt calculated in step S20. For calculating the gradient, for example, the backpropagation method can be used, but is not limited to this.

Next, in step S22, the learning means 101 updates the parameters of each DNN as follows.
- Partner feature calculation DNN201: Minimize goodness-of-fit error e _{i j} ^{e vl} , maximize partner division error e _i ^{d pt} - Project feature calculation DNN202: Minimize goodness-of-fit error e _{i j} ^{e vl} , maximize project division error e _j ^{d pt} - Partner division calculation DNN203: Minimize partner division error e _i ^{d pt} - Project division calculation DNN204: Minimize project division error e _j ^{d pt}

Furthermore, Adam's rules or similar can be used to calculate the specific parameter update difference in S22.

Next, in step S23, the learning means 101 terminates the iteration unit. The iteration unit may be configured to update the parameters each time a partner-project pair is processed, or to update the parameters after a predetermined number of partner-project pairs have been processed, or the timing of parameter updates may be determined based on other criteria.

Upon completion of the repeated processing in steps S15 to S23, the machine learning process is finished, and each DNN is generated as a trained model. Next, in step S24, the learning means 101 stores the parameters resulting from the training performed in steps S15 to S23 in the learning result table T4. Then, in step S25, the learning means 101 terminates the process.

Figure 12 shows an example of the structure of the learning result table T4. This example shows some of the parameters that define the partner feature computation DNN 201. For the matrix specified by the parameter name "A _ptn ⁰ ", pairs of row and column numbers that identify each element of the matrix are defined, and numerical values are stored for each pair. Other parameters of the partner feature computation DNN 201, as well as other DNNs, can be expressed in a similar format.

<Prediction processing in Embodiment 1>
The following describes step S4 in Figure 3 in more detail. In the prediction process of step S4, the degree of fit between the user-specified partner and the project is predicted based on the degree of fit calculation method learned in the learning process of step S3.

The user of the matching device 100 generates the prediction target table T5 using the input device 103 before step S4 is executed.

Figure 13 shows an example of the structure of the prediction target table T5. The prediction target table T5 stores pairs of partners and projects for which the degree of fit to be predicted is to be determined. Specifically, the partner ID and project ID are used as keys.

Here, if you are looking for a partner to use in a specific project, you should enter all partner pairs and the corresponding project. Similarly, if you want to find recommended projects for a specific partner, you should enter all projects and the corresponding partner pairs. Figure 13 shows a scenario where you are looking for a partner to use in project PRJ1001, and all partner pairs for project PRJ1001 have been entered.

Figure 14 shows a flowchart of the operation of the prediction means 102. This flowchart illustrates the details of step S4 in Figure 3.

In step S31, the prediction means 102 starts processing. Next, in step S32, the prediction means 102 retrieves the partner and project pairs to be predicted from the prediction target table T5.

Next, in step S33, the prediction means 102 obtains attribute values for the partner and project acquired in step S32. These attribute values can be obtained, for example, from the partner table T1 and the project table T2, but may also be obtained from other sources (i.e., they may be different from the teacher partner information and teacher project information).

Next, in step S34, the prediction means 102 retrieves the parameters of each DNN from the learning result table T4 and sets them in each DNN, thereby reproducing each DNN as a trained model.

Next, in step S35, the prediction means 102 uses a feature calculation process to calculate partner features and project features from the partner attribute values and project attribute values obtained in step S33. This calculation is performed by the partner feature calculation DNN 201 and the project feature calculation DNN 202.

Next, in step S36, the prediction means 102 uses a goodness-of-fit calculation process to calculate the goodness-of-fit of the pair from the partner features and project features calculated in step S35.

Next, in step S37, the prediction means 102 writes the degree of fit calculated in step S36 to the prediction result table T6. Then, in step S38, the prediction means 102 terminates the process.

Figure 15 shows an example of the structure of the prediction results table T6. The prediction results table T6 uses all the pairs obtained in step S32 as keys, and stores the goodness-of-fit calculated in step S36 for each pair. For example, the first row of Figure 15 shows that the goodness-of-fit between partner PTN0001 and project PRJ1001 is 0.234.

The user of the matching device 100 (for example, a procurement officer) can select partners to use for a project based on the degree of compatibility between partners and projects stored in the prediction results table T6. For example, for each project, the partner with the highest degree of compatibility can be selected.

As mentioned above, since the parameters of each DNN learn a common, department-independent prediction method, it can more accurately predict the degree of fit even when partners and projects belonging to different departments are input.

100... Matching device 101... Learning means 102... Prediction means 103... Input device 104... Output device 105... Calculation unit 106... Memory device T1... Partner table T2... Project table T3... Matching performance table T4... Learning result table T5... Prediction target table T6... Prediction result table 201... Partner feature calculation DNN (First Deep Neural Network)
202...Project Feature Calculation DNN (Second Deep Neural Network)
203... Partner Division Computational DNN (Third Deep Neural Network)
204...Project Division Computational DNN (Fourth Deep Neural Network)

Claims (7)

A matching device that predicts the degree of compatibility between a partner and a project,
The matching device comprises a learning means and a prediction means,
The learning means generates a trained model by performing machine learning processing based on partner attribute values, project attribute values, partner correct answer category, project correct answer category, and correct answer fit.
The aforementioned machine learning process is
- A first learning process that takes the partner attribute value and the project attribute value as input and learns to output a predicted fit that is closer to the correct fit,
- A second learning process that takes the aforementioned partner attribute values and project attribute values as input and learns a common prediction method that can be used to predict the degree of fit of any department,
Includes,
The prediction means uses the trained model to predict the degree of fit based on information about the partner and information about the project.
The first learning process described above is:
A process to calculate partner features based on the aforementioned partner attribute values,
A process for calculating project features based on the aforementioned project attribute values,
A process for calculating the prediction fit based on the partner features and project features,
A process to learn so that the accuracy error, which is the difference between the correct accuracy and the predicted accuracy, becomes smaller,
Includes,
The second learning process described above is:
Based on the aforementioned partner features, a process is performed to predict the partner prediction department to which the partner belongs.
A process to calculate the partner division error as the difference between the partner prediction division and the partner correct answer division,
A process to learn so that the aforementioned partner division error becomes larger,
Includes,
The second learning process described above is:
Based on the aforementioned project features, a process is performed to predict the project forecasting department to which the project belongs.
The process of calculating the project department error as the difference between the project prediction department and the project correct answer department,
The process involves learning to increase the error in the aforementioned project division,
including,
Matching device. A matching device according to claim 1 ,
The aforementioned partner features and project features are represented by vectors.
The aforementioned machine learning process is performed using a deep neural network.
Matching device. A matching device according to claim 1,
The aforementioned goodness of prediction is calculated using the squared error or cosine similarity.
Matching device. A matching device according to claim 1 ,
The aforementioned partner prediction sector is represented by a vector that indicates the likelihood of the partner belonging to each sector.
The correct answer category for the partner is represented by a vector where the likelihood of the partner belonging to that category is 1, and the likelihood of the partner belonging to another category is 0.
The aforementioned partner division error is calculated using the cross-entropy error function.
Matching device. A matching device according to claim 2 ,
The aforementioned machine learning process is
- A first deep neural network that takes the aforementioned partner attribute values as input and outputs the aforementioned partner features,
- A second deep neural network that takes the aforementioned project attribute values as input and outputs the aforementioned project features,
- A third deep neural network that takes the aforementioned partner features as input and outputs the partner prediction department to which the partner belongs,
- A fourth deep neural network that takes the aforementioned project features as input and outputs the project prediction department to which the project belongs,
It is done using
In the first learning process, the first deep neural network and the second deep neural network are trained.
In the second learning process, the first deep neural network, the second deep neural network, the third deep neural network, and the fourth deep neural network are trained.
Matching device. A matching device according to claim 5 ,
In the first learning process described above,
- Update the parameters of the first deep neural network to reduce the aforementioned goodness-of-fit error,
- The parameters of the second deep neural network are updated to reduce the aforementioned goodness-of-fit error,
In the second learning process described above,
- Update the parameters of the first deep neural network to increase the partner division error,
- Update the parameters of the second deep neural network to increase the error in the aforementioned project division,
- Update the parameters of the third deep neural network to reduce the aforementioned partner division error,
- Update the parameters of the fourth deep neural network to reduce the project division error.
Matching device. A program that causes a computer to function as a matching device according to any one of claims 1 to 6 .