CN115564490A - A Supplier Evaluation Method Based on Attention Mechanism - Google Patents

Abstract

The invention belongs to the technical field of supplier management, and provides a supplier evaluation method based on an attention mechanism, which comprises the following steps: step 1: determining evaluation items and weights to obtain maximum characteristic root lambda _max Constructing an evaluation result iteration operator A; step 2: introducing a regular term and an attention mechanism, improving an iteration target L, constructing a neural network, iterating until an iteration threshold constraint is met, and outputting a weight output matrix A of an evaluation result iteration operator A _output (ii) a And step 3: based on maximum characteristic root lambda _max Constructing a consistency check operator and judging the current weight output matrix A _output Whether it is valid. The method constructs an evaluation result iterative operator as a measure selected by a supplierAnd an attention mechanism and a consistency checking operator are introduced into the function, so that the evaluation accuracy is improved, and the evaluation quality of a supplier is guaranteed.

Description

A Supplier Evaluation Method Based on Attention Mechanism

technical field

The invention relates to the technical field of supplier management, in particular to a supplier evaluation method based on an attention mechanism.

Background technique

With the rapid development of economy and electronic information technology, supplier management is becoming more and more important to the development and competition of enterprises, and has become a research hotspot. Supplier evaluation is a basic part of supplier management. It refers to the process of assessing and evaluating potential suppliers in various ways before establishing a cooperative relationship with suppliers, and continuously tracking and giving feedback during the process of establishing cooperation. process. The evaluation and selection of suppliers is the prerequisite for the smooth development of procurement management and supply management, and it is also the cornerstone for enterprises to maintain long-term competitive advantages under the globalization situation. Accurate and efficient supplier evaluation algorithms can not only reduce enterprise procurement costs and improve product quality, but also integrate industrial resources and promote the healthy and orderly growth of the entire industrial chain. Therefore, the research on supplier evaluation methods has great application value.

The existing supplier evaluation algorithm mainly uses the calculation of procurement cost for comparative analysis, and selects suppliers with lower procurement costs by calculating the sum of various expenses including selling price, procurement costs, and transportation costs. This method simply chooses from the perspective of procurement cost, which has great limitations and often runs counter to the strategic goals of the enterprise. Therefore, some scholars have introduced the data envelopment analysis method DEA, and increased the dimension of supplier evaluation by constructing multi-indicator input and output. Data envelopment analysis cannot effectively reflect decision makers' preferences for judgment criteria. At the same time, because the correlation between criteria is not analyzed, it leads to repeated evaluation. Based on this, scholars put forward the Analytic Hierarchy Algorithm (AHP), which determines the evaluation criteria through experts and adopts comprehensive consulting scores to determine the weight. However, AHP needs to meet the iterative goals and constraints, and it is difficult to meet the global consistency requirements when there are many element goals and the problem scale is large.

Contents of the invention

In view of this, an embodiment of the present invention provides a supplier evaluation method based on an attention mechanism to solve or partially solve the above problems.

An attention mechanism-based supplier evaluation method provided by an embodiment of the present invention includes the following steps:

Step 1: Determine the evaluation items and weights, obtain the largest characteristic root λ _max , and construct the evaluation result iteration operator A;

Step 2: Introduce the regular term and attention mechanism, improve the iterative target L, construct the neural network, iterate until the iterative threshold constraint is satisfied, and output the weight output matrix A _output of the iterative operator A of the evaluation result;

Step 3: Construct a consistency check operator based on the largest characteristic root λ _max , and judge whether the current weight output matrix A _output is valid.

According to a specific implementation of the embodiment of the present invention, the step 1 includes the following steps:

Step 1.1: Build a hierarchical evaluation model, determine the target layer, criterion layer and sub-criteria level of supplier evaluation decision-making events, and construct a judgment matrix A ₀ for the evaluation items of the sub-criteria level:

Step 1.2: Calculate the weight W of each factor a _ij to the target layer based on the judgment matrix, specifically:

次幂：For the judgment matrix A0, calculate the product by row elements, and then ball

power:

归一化：Will

Normalized:

Then the indicator vector is expressed as:

W = (W ₁ , W ₂ . . . W _n ) ^T (4);

Step 1.3: Based on the weight W, calculate the largest characteristic root λ _max :

AW = λ _max W (5);

And then get:

Step 1.4: Construct the evaluation result iteration operator, specifically:

其中m为历史决策事件的数目，d为数据库存储的决策用户数目，即决策并行数；Define the current supplier evaluation decision event as U _i , then U _i belongs to the decision event embedding matrix U,

Where m is the number of historical decision-making events, d is the number of decision-making users stored in the database, that is, the number of decision-making parallelism;

其中n为评价项数，Define the evaluation item matrix

where n is the number of evaluation items,

Construct the evaluation result iteration operator A:

A=UV ^T (7);

Then the iterative optimization problem can be defined as:

where ||A-UV ^T || _F represents the Frobenius norm of A and its approximation UV ^T.

According to a specific implementation of the embodiment of the present invention, in the step 1.1, the element assignment in A ₀ uses the 1-9 scaling method of Santy, and satisfies:

According to a specific implementation manner of the embodiment of the present invention, the step 2 is specifically:

Step 2.1: Introduce the l2 regular term and the regular term of the gravity implicit matrix to the objective function L:

Further get:

In the formula, L represents the objective function, i is the index of the decision-making event matrix, j is the index of the evaluation item matrix, Ω represents the subset of the judgment matrix A ₀ , λ _r and λ _g are two regularization coefficients, and M is the The total number of supplier evaluation decision-making events, N is the total number of evaluation items in the set Ω;

Step 2.2: Introducing the attention mechanism, U can be further expressed as:

In the formula, U _i is the evaluation score vector of the i-th decision, and U _p is the historical decision-making vector before the i-th time;

Step 2.3: Construct a neural network according to Step 2.1 and Step 2.2;

Step 2.4: Connect the decision event vector in the database to construct the decision event embedding matrix U;

Step 2.5: According to the serial number of the current iteration, construct the historical decision event vector U _p ;

Step 2.6: Initialize the supplier evaluation item matrix V as W in step 1.2, set the neural network learning rate γ, regular hyperparameters λ _r and λ _g , iteration threshold L ₀ , maximum iteration number Maxgen, initial iteration number k=1;

Step 2.7: According to step 2.2, U _i and the convolution vector of U _i and U _p are used as input data, input to the fully connected layer, and the decision event input vector is obtained through the nonlinear ReLU activation function:

Step 2.8: Obtain the supplier evaluation matrix V according to step 2.6. For the iteration number k, obtain the input vector of supplier evaluation items:

V _input = V _k (15);

Step 2.9: Input the decision event input vector and the supplier evaluation item input vector into the neural network to obtain the weight output matrix of the iterative operator A of the evaluation result:

A _output = Softmax(FCs2(FCs1(U _input , V _input )) (16);

Step 2.10: Calculate the objective function value L according to the formula (10) in step 2.1;

Step 2.11: The number of iterations k=k+1, if L≥L ₀ and k<Maxgen, return to step 2.7, otherwise output the current optimal solution, and enter step 3.1.

According to a specific implementation manner of the embodiment of the present invention, the step 3 is specifically:

Step 3.1: Obtain the maximum characteristic root λ _max of the initial evaluation matrix according to step 1.3, and construct the consistency check operator as follows:

n表示评价矩阵阶数，C₀为检验常数；in

n represents the evaluation matrix order, and C ₀ is a test constant;

Step 3.2: Compare the consistency of the one-time check operator in step 3.1 with the threshold θ ₀ , if consistency ≤ θ ₀ , the current optimal solution output in step 2.11 is valid, otherwise the current optimal solution output in step 2.11 is invalid, return to step 1.

Embodiments of the present invention have at least the following technical effects:

First, the attention mechanism introduced by the present invention comprehensively considers the relationship between historical decision-making and supplier evaluation items, and can refine core evaluation items according to the current focus of the enterprise to improve evaluation accuracy; the introduction of a consistency check operator, for The combination of evaluation items obtained by the neural network is tested to avoid logic problems in the evaluation matrix and ensure the quality of supplier evaluation.

Second, the present invention introduces a neural network to replace the traditional explicit solution. As the amount of data such as the number of decisions increases, the training data such as the number of historical decisions stored in the database increases accordingly, and the attention factor will become more accurate, realizing " The more you use it, the better it works".

Third, the present invention constructs an evaluation result iteration operator as a measurement function for supplier selection, comprehensively considers the impact of evaluation item weights and decision-making events, avoids waste of historical data, and improves evaluation accuracy.

Description of drawings

In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings that need to be used in the description of the specific embodiments or the prior art. Throughout the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, elements or parts are not necessarily drawn in actual scale.

Fig. 1 shows a flow chart of the steps of an attention mechanism-based supplier evaluation method provided by an embodiment of the present invention;

FIG. 2 shows a hierarchical diagram of supplier evaluation decision-making events in an embodiment of the present invention;

FIG. 3 shows a structural diagram of a neural network based on an attention mechanism in an embodiment of the present invention.

detailed description

Embodiments of the technical solutions of the present invention will be described in detail below in conjunction with the accompanying drawings. The following examples are only used to illustrate the technical solution of the present invention more clearly, so they are only examples, and should not be used to limit the protection scope of the present invention.

It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application shall have the usual meanings understood by those skilled in the art to which the present invention belongs.

Fig. 1 is a flow chart of the steps of an attention mechanism-based supplier evaluation method provided by an embodiment of the present invention. Referring to Fig. 1, the method includes the following steps:

Step 1: Determine the evaluation items and weights, obtain the largest characteristic root λ _max , and construct the evaluation result iteration operator A. details as follows:

Step 1.1: Combining the AHP supplier evaluation algorithm and the decision maker's preference for criteria, construct a hierarchical evaluation model, as shown in Figure 2, determine the target layer, criterion layer and sub-criteria level of supplier evaluation decision-making events, and evaluate the sub-criteria level Item construction judgment matrix A ₀ :

The element assignment in A ₀ uses Santy's 1-9 scale method, and satisfies:

Step 1.2: Calculate the weight W of each factor a _ij to the target layer based on the judgment matrix, specifically:

次幂：For the judgment matrix A ₀ , calculate the product by row elements, and then calculate

power:

归一化：Will

Normalized:

Then the indicator vector can be expressed as:

W = (W ₁ , W ₂ . . . W _n ) ^T (4).

Step 1.3: Based on the weight W, calculate the largest characteristic root λ _max :

AW = λ _max W (5);

Then you can get:

Step 1.4: Construct the evaluation result iteration operator, specifically:

其中m为历史决策事件的数目，d为数据库存储的决策用户数目，即决策并行数；Define the current supplier evaluation decision event as U _i , then U _i belongs to the decision event embedding matrix

Where m is the number of historical decision-making events, d is the number of decision-making users stored in the database, that is, the number of decision-making parallelism;

其中n为评价项数，构造评价结果迭代算子A：Define the evaluation item matrix

Where n is the number of evaluation items, construct the evaluation result iteration operator A:

A=UV ^T (7);

Then the iterative optimization problem can be defined as:

where ||A-UV ^T || _F represents the Frobenius norm of A and its approximation UV ^T.

Step 2: Introduce the regular term and attention mechanism, improve the iterative target L, construct the neural network, iterate until the iterative threshold constraint is satisfied, and output the weight output matrix A _output of the iterative operator A of the evaluation result. Specifically:

Step 2.1: Introduce the l2 regular term and the regular term of the gravity implicit matrix to the objective function L:

Further available:

In the formula, L represents the objective function, that is, the iterative loss function, i is the index of the decision-making event matrix, j is the index of the evaluation item matrix, Ω represents the subset of the judgment matrix A ₀ , λ _r and λ _g are two regularization coefficients, M is the total number of supplier evaluation decision-making events in the set Ω, and N is the total number of evaluation items in the set Ω.

The traditional neural network is prone to overfitting. The reason is generally that the network model takes into account every point in the sample, and the final function fluctuates greatly. A slight change in the independent variable will cause a drastic change in the function value. Here, r( U, V) and g(U, V) are two regular terms, which reduce the weight, thereby reducing the fluctuation of the function value, thereby avoiding the occurrence of overfitting.

Step 2.2: When evaluating suppliers, different companies have their own evaluation criteria. Therefore, it is necessary to fully consider the evaluation criteria of the current company and the preference of the current decision-makers for the criteria. As the name implies, the attention mechanism means that when the model is predicting, it pays different attention to different behaviors. The "related" behavior history is more important, and the "irrelevant" history can even be ignored. For the decision-making process of supplier evaluation, the attention mechanism is introduced, and the relationship between the current round of decision-making events and historical decision-making events is used as the weight. If the correlation is high, it proves that the current enterprise pays more attention to this evaluation item.

In the traditional supplier evaluation algorithm, the decision event embedding matrix U of formula (10) decision event is generally a simple summation of a single evaluation decision U _i or a weighted sum with fixed weights, namely

Introducing the attention mechanism, U can be further expressed as:

In the formula, U _i is the evaluation score vector of the i-th decision, and U _p is the historical decision-making vector before the i-th time. Due to the addition of the attention mechanism, the weight w _i of U _i is determined by the relationship between U _p and U _i , that is, g(U _p , U _i ) in formula (12);

Step 2.3: Construct a neural network according to Step 2.1 and Step 2.2, as shown in Figure 3;

Step 2.4: Connect the decision event vector in the database to construct the decision event matrix U;

Step 2.5: According to the serial number of the current iteration, construct the historical decision event vector U _p ;

Step 2.6: Initialize the supplier evaluation item matrix V as W in step 1.2, set the neural network learning rate γ, regular hyperparameters λ _r and λ _g , iteration threshold L ₀ , maximum iteration number Maxgen, initial iteration number k=1;

Step 2.7: According to step 2.2, U _i and the convolution vector of U _i and U _p are used as input data, input into the fully connected layer, and the decision event input vector is obtained through the nonlinear ReLU activation function:

Step 2.8: Obtain the supplier evaluation matrix V according to step 2.6. For the iteration number k, obtain the input vector of supplier evaluation items:

V _input = V _k (15);

Step 2.9: Input the decision-making event input vector and the supplier evaluation item input vector into the neural network to obtain the weight output matrix of the iterative operator A for constructing evaluation results:

A _ouptut = Softmax(FCs2(FCs1(U _input , V _input )) (16);

Step 2.10: Calculate the objective function value L according to the formula (10) in step 2.1;

Step 2.11: increase the number of iterations k=k+1, if L≥L ₀ and k<Maxgen, return to step 2.7, otherwise output the current optimal solution, and enter step 3.1;

Step 3: Construct a consistency check operator based on the largest characteristic root λ _max , and judge whether the current weight output matrix A _output is valid.

Step 3.1: Construct a consistency check operator;

According to step 1.3, the largest characteristic root λ _max of the initial evaluation matrix is obtained, and the consistency check operator is constructed as follows:

n表示评价矩阵阶数，C₀为检验常数，可通过随机一致性指标取值表查得。in

n represents the order of the evaluation matrix, and C ₀ is a test constant, which can be found through the value table of the random consistency index.

Step 3.2: According to step 3.1, calculate the result of the consistency test, and set the consistency test threshold as θ ₀ . If consistency≤θ ₀ , the current optimal solution output in step 2.11 is valid; otherwise, the current optimal solution output in step 2.11 is invalid, and return to step 1.1.

If the current optimal solution output in step 2.11 is valid, then input the scores of each index item according to the evaluation matrix, and obtain the objective score of each supplier after the corresponding weight and level cross calculation. But the current optimal solution output in step 2.11 is invalid, return to step 1.1, that is, re-input a reasonable initial evaluation matrix A ₀ , and execute the method again.

It should be noted that the arrangement of the various modules according to the flow layout is only an embodiment of the present invention, and other arrangements may also be adopted, which is not limited in the present invention.

Embodiments of the invention have the following technical effects:

First, the present invention obtains the initial index weight based on the AHP supplier evaluation algorithm, retains the advantages of the AHP algorithm, pays attention to the correlation between the analysis criteria, and refines the decision maker's preference for the criteria;

Second, the present invention constructs an evaluation result iteration operator as a measurement function for supplier selection, comprehensively considers the impact of evaluation item weights and decision-making events, avoids waste of historical data, and improves evaluation accuracy;

Third, the attention mechanism introduced by the present invention comprehensively considers the relationship between historical decision-making and supplier evaluation items, and can refine core evaluation items according to the current focus of the enterprise to improve evaluation accuracy;

Fourth, the present invention introduces a neural network to replace the traditional explicit solution. As the amount of data such as the number of decisions increases, the training data such as the number of historical decisions stored in the database increases accordingly, and the attention factor will become more accurate, realizing " The more you use it, the better it works”;

Fifth, the present invention introduces a consistency check operator to check the combination of evaluation items obtained by the neural network to avoid logic problems in the evaluation matrix and ensure the quality of supplier evaluation.

The above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Anyone skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present invention. All should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (5)

1. A supplier evaluation method based on attention mechanism, is characterized in that, comprises the following steps: Step 1: Determine the evaluation items and weights, obtain the largest characteristic root λ _max , and construct the evaluation result iteration operator A; Step 2: Introduce a regular term and attention mechanism, improve the iterative target L, construct a neural network, iterate until the iterative threshold constraint is met, and output the weight output matrix A _output of the iterative operator A of the evaluation result; Step 3: Construct a consistency check operator based on the maximum characteristic root λ _max , and judge whether the current weight output matrix A _output is valid. 2. The supplier evaluation method according to claim 1, wherein said step 1 comprises the following steps: Step 1.1: Build a hierarchical evaluation model, determine the target layer, criterion layer and sub-criteria level of supplier evaluation decision-making events, and construct a judgment matrix A ₀ for the evaluation items of the sub-criteria level:

Step 1.2: Calculate the weight W of each factor a _ij to the target layer based on the judgment matrix, specifically: For the judgment matrix A ₀ , calculate the product by row elements, and then calculate

power:

Will

Normalized:

Then the indicator vector is expressed as: W=(W ₁ ,W ₂ ...W _n ) ^T (4); Step 1.3: Based on the weight W, calculate the largest characteristic root λ _max : AW = λ _max W (5); And then get:

Step 1.4: Construct the evaluation result iteration operator, specifically: Define the current supplier evaluation decision event as U _i , then U _i belongs to the decision event embedding matrix U,

Where m is the number of historical decision-making events, d is the number of decision-making users stored in the database, that is, the number of decision-making parallelism; Define the evaluation item matrix

where n is the number of evaluation items, Construct the evaluation result iteration operator A: A=UV ^T (7); Then the iterative optimization problem is defined as:

where ||A-UV ^T || _F represents the Frobenius norm of A and its approximation UV ^T. 3. The supplier evaluation method according to claim 2, characterized in that: in the step 1.1, the element assignment in A ₀ uses the 1-9 scaling method of Santy, and satisfies:

4. The supplier evaluation method according to claim 2, characterized in that the step 2 is specifically: Step 2.1: Introduce to the objective function L

Regular term and regular term of gravity implicit matrix:

Further get:

In the formula, L represents the objective function, i is the index of the decision-making event matrix, j is the index of the evaluation item matrix, Ω represents the subset of the judgment matrix A ₀ , λ _r and λ _g are two regularization coefficients, and M is the The total number of supplier evaluation decision-making events, N is the total number of evaluation items in the set Ω; Step 2.2: Introducing the attention mechanism, U is further expressed as:

In the formula, U _i is the evaluation score vector of the i-th decision, and U _p is the historical decision-making vector before the i-th time; Step 2.3: Construct a neural network according to Step 2.1 and Step 2.2; Step 2.4: Connect the decision event vector in the database to construct the decision event matrix U; Step 2.5: According to the serial number of the current iteration, construct the historical decision event vector U _p ; Step 2.6: Initialize the supplier evaluation item matrix V as W in step 1.2, set the neural network learning rate γ, regular hyperparameters λ _r and λ _g , iteration threshold L ₀ , maximum iteration number Maxgen, initial iteration number k=1; Step 2.7: According to step 2.2, U _i and the convolution vector of U _i and U _p are used as input data, input to the fully connected layer, and the decision event input vector is obtained through the nonlinear ReLU activation function:

Step 2.8: Obtain the supplier evaluation matrix V according to step 2.6. For the iteration number k, obtain the input vector of supplier evaluation items: V _input = V _k (15); Step 2.9: Input the decision event input vector and the supplier evaluation item input vector into the neural network to obtain the weight output matrix of the iterative operator A of the evaluation result: A _output = Softmax(FCs2(FCs1(U _input ,V _input )) (16); Step 2.10: Calculate the objective function value L according to the formula (10) in step 2.1; Step 2.11: The number of iterations k=k+1, if L≥L ₀ and k<Maxgen, return to step 2.7, otherwise output the current optimal solution. 5. The supplier evaluation method according to claim 4, characterized in that the step 3 is specifically: Step 3.1: Obtain the maximum characteristic root λ _max of the initial evaluation matrix according to step 1.3, and construct the consistency check operator as follows:

in

n represents the evaluation matrix order, and C ₀ is a test constant; Step 3.2: Compare the one-time check operator consistenxy in step 3.1 with the threshold θ ₀ , if consistency≤θ ₀ , then the current optimal solution output in step 2.11 is valid, otherwise the current optimal solution output in step 2.11 is invalid, return to step 1.

Images