A supplier recommendation system and method based on large model

Abstract

The present invention relates to the field of enterprise management technology, and specifically discloses a supplier recommendation system and method based on a large model, comprising: obtaining historical supplier data, constructing a supplier similarity matrix, constructing a classification model using a hierarchical clustering algorithm, determining suppliers of the same type, then analyzing the supplier data of the same type, obtaining a similar characterization value Ts and constructing a model training data group, constructing a supplier large model based on different algorithms, determining a target model algorithm through comparative analysis, adjusting the number of training times according to numerical analysis related to the target model algorithm, determining a recommended model algorithm, and outputting a recommended supplier; the system comprises a data acquisition module, a classification judgment module, a data analysis module, a model construction module, and a model determination module; the present invention can effectively process supplier data, improve the accuracy and reliability of recommended suppliers through a variety of algorithms and analysis methods, and provide strong support for enterprise procurement.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise management, and in particular to a supplier recommendation system and method based on a large model. Background Art

[0002] With the increasing complexity of the market and the continuous expansion of corporate business scale, how to accurately screen partners that meet corporate needs from a large number of suppliers has become a key challenge facing companies.

[0003] A Chinese patent application with announcement number CN112927037B discloses a supplier recommendation method and system, including: receiving at least one supplier sample; predicting the recommended operation corresponding to each supplier sample based on an adaptive multi-layer perception online transfer learning model; tracking and obtaining the actual adoption effect of each supplier sample by supplier users; adjusting the weight distribution and nonlinear parameter vector between the user features of the previous batch and the user features of the current batch of purchasing users served by the adaptive multi-layer perception online transfer learning model based on the comparison between the recommended operation and the actual adoption effect; re-acquiring new supplier samples and predicting the corresponding recommended operation using the adjusted online transfer learning model, and repeating the above steps until all supplier samples that meet the needs of the supplier users are screened out.

[0004] In model construction, the existing technology uses a relatively simple algorithm and only adopts an online transfer learning model with adaptive multi-layer perception. It has certain limitations in capturing the complex nonlinear relationships and dynamic changes in supplier data. It cannot adapt to the advantages brought by the collaboration of multiple algorithms and cannot meet the diverse needs of different types of supplier data. If a large supplier model is constructed through multiple algorithms and the optimal model algorithm is determined through comparison, it can better adapt to complex data characteristics and changes in enterprise needs.

[0005] When adjusting the model, the existing technology only considers the weight distribution of user characteristics and the nonlinear parameter vector, and does not comprehensively consider multiple factors such as weight ratio, misjudgment similarity mean, and similar representation values to dynamically adjust the number of model training times and determine the recommendation model algorithm; through the dynamic adjustment mechanism, the model performance is more comprehensively optimized and the reliability of supplier recommendations is improved. Summary of the Invention

[0006] The purpose of the present invention is to provide a supplier recommendation system and method based on a large model to solve the above background problems.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] A supplier recommendation method based on a large model includes the following steps:

[0009] Step 1: Obtain historical supplier data and store it in a supplier data group, perform similarity analysis on the supplier data group, and build a supplier similarity matrix ;

[0010] Step 2: Based on the supplier similarity matrix , build a supplier classification model to identify the same suppliers in different categories;

[0011] Step 3: Perform a numerical analysis on the supplier similarity of the same type of suppliers in the same category to obtain the central similarity difference Zx. Obtain the historical supply times of suppliers in the same category and perform numerical calculations to obtain the same type characterization value Ts. Rank the same type of suppliers based on the same type characterization value Ts and construct a model training data set.

[0012] Step 4: Based on the model training data set, use different algorithms to build a supplier model, output recommended suppliers based on the supplier model, and conduct comparative analysis to determine the target model algorithm;

[0013] Step 5: Based on the target model algorithm, perform numerical analysis on the output results of the target model algorithm to obtain the training enhancement ratio Xlq of the target model algorithm, and dynamically adjust the number of training times of the supplier large model of the target model algorithm. Based on the dynamically adjusted target algorithm model, determine the supplier large model corresponding to the recommended model algorithm and determine the recommended supplier.

[0014] As a further technical solution of the present invention: the supplier similarity matrix The way to obtain is:

[0015] Based on the enterprise's internal procurement management system, the historical transaction record data between the enterprise and different suppliers is obtained, and the historical transaction record data of different suppliers is used as the supplier data set a i ;

[0016] The historical transaction records include the supplier's product category and timeliness. i is the supplier's ID, i = 1, 2, ..., n, where n is the number of suppliers.

[0017] Put all supplier data sets into supplier data group, and construct supplier data group A, A={a1, a2, ...a n};

[0018] Extract supplier data set a i The goods category data in the i , calculate the supplier similarity G of different supplier data sets ij ;

[0019] By formula: Get the supplier similarity G of the supplier data set ij ;

[0020] where j=i+1,i+2,...,n, Represents the supplier data set c i with c j The number of intersections, Represents the supplier data set c i with c j The number of unions, X 12 The supplier similarity G between two suppliers numbered 1 and 2 ij ;

[0021] Based on the supplier similarity G in different supplier data ij , construct supplier similarity matrix S ij ;

[0022] Construct the supplier similarity matrix S ij = ;

[0023] The supplier similarity matrix S ij Perform preprocessing to remove the supplier similarity matrix S ij The meaningless data in the matrix, including the matrix diagonal and the upper triangle data, mark the pre-processed supplier similarity matrix as .

[0024] As a further technical solution of the present invention: the method of obtaining the same type of suppliers of different categories is:

[0025] Supplier Similarity Matrix Use hierarchical clustering algorithm to identify unique suppliers and homogeneous suppliers;

[0026] The supplier classification model is constructed as follows:

[0027] S1. Calculate the similarity distance d(i,j) using the formula: Get the similarity distance d(i,j), and build a similarity distance matrix based on the similarity distance d(i,j);

[0028] S2. Input the supplier matrix into the hierarchical clustering algorithm. The hierarchical clustering algorithm establishes different clusters based on the similarity distance d(i, j), calculates the distance between each cluster, and updates each cluster iteratively.

[0029] S3. Analyze the clusters output by the hierarchical clustering algorithm. If there is only one supplier in the cluster, mark the corresponding supplier as the only supplier. If there are multiple suppliers in the cluster, mark the suppliers in the cluster as the same type of suppliers.

[0030] As a further technical solution of the present invention: the model training data set is constructed as follows:

[0031] Sort the elements of the same level of detail sequence in descending order of the same level of characterization value Ts;

[0032] Get all unique suppliers and put them into the unique supplier sequence;

[0033] Obtain similarity sequences and unique supplier sequences of all categories that have been sorted, and build a model training data set.

[0034] As a further technical solution of the present invention: the same type characterization value Ts is obtained as follows:

[0035] Obtain the historical supply counts of the same type of supplier within the same category, and the historical supply counts of all suppliers of the same type;

[0036] The historical supply times of the same supplier are compared with the historical supply times of all suppliers of the same type to obtain the same type supply ratio, which is marked as T1;

[0037] Arrange the supplier similarities of the same type of suppliers in the same category from largest to smallest, and place the sorted suppliers as elements in the similarity sequence;

[0038] Get the total number of elements in the same similarity sequence, mark the total number of elements as Zs, and perform parity judgment on the total number of elements;

[0039] If the total number of elements Zs is an odd number, get the similarity sequence of the same type. The similarity of suppliers is used as the central similarity;

[0040] If the total number of elements Zs is an even number, obtain the similarity between the (Zs / 2)th and (Zs / 2+1)th suppliers in the similarity sequence, sum and average the similarities between the (Zs / 2)th and (Zs / 2+1)th suppliers in the similarity sequence, and obtain the central similarity.

[0041] Based on the similarity sequence of the same type, calculate the difference between each element and the center similarity of the similarity sequence to obtain the center similarity difference, and mark the center similarity difference as Zx;

[0042] Based on the center similarity difference Zx and the similar supply ratio TI, the similar characterization value Ts is calculated;

[0043] By formula: Get the same type characterization value Ts, where f and g are preset proportional coefficients.

[0044] As a further technical solution of the present invention: the target model algorithm is constructed as follows:

[0045] Based on the model training data set, different algorithms are used to build a large supplier model;

[0046] A supplier model built based on different algorithms takes into account the type of goods and timeliness, outputs recommended suppliers, and compares the recommended suppliers with the actual suppliers.

[0047] If the recommended supplier is inconsistent with the actual adopted supplier, the supplier will be marked as a misjudged supplier;

[0048] If the recommended supplier is consistent with the actual adopted supplier, the supplier will be marked as the adopted supplier;

[0049] Obtain the number of misjudged suppliers in the supplier macro models constructed using different algorithms, as well as the number of recommended suppliers output by the supplier macro models;

[0050] The number of misjudged suppliers is compared with the number of recommended suppliers output by the suppliers to obtain the misjudgment ratio.

[0051] Calculate the supplier model constructed by different algorithms, output recommended suppliers and misjudgment ratios multiple times, and sum and average the misjudgment ratios output multiple times to obtain the mean misjudgment ratio.

[0052] If the mean false positive ratio is lower than the false positive threshold, the algorithm corresponding to the supplier's large model is obtained, and the model constructed by the corresponding algorithm is marked as the target model algorithm.

[0053] As a further technical solution of the present invention: the method for dynamically adjusting the number of supplier large model training times of the target model algorithm is:

[0054] Obtain the number of supplier large model training times Mxc of the target model algorithm, calculate the additional training times of the target model algorithm, and mark the additional training times of the target model algorithm as Ew;

[0055] Calculate the additional training times Ew of the target model algorithm based on the additional training times Ew of the target model algorithm and the training enhancement ratio Xlq of the target model algorithm;

[0056] By formula: Get the additional training times of the target model algorithm;

[0057] Based on the additional training times of the target model algorithm, the training times of the supplier large model constructed by the target model algorithm are dynamically adjusted.

[0058] As a further technical solution of the present invention: the training enhancement ratio Xlq of the target model algorithm is obtained as follows:

[0059] Obtain the number of adopted suppliers output by the target model algorithm and the number of recommended suppliers output by the target model algorithm;

[0060] The number of adopted suppliers output by the target model algorithm is compared with the number of recommended suppliers output by the target model algorithm to obtain the adoption-supply ratio;

[0061] Obtain the adoption supply ratios of all target model algorithms, sum the adoption supply ratios of all target algorithms, and obtain the total adoption ratio value;

[0062] The adoption supply ratio is compared with the total adoption ratio to obtain the adoption weight ratio, which is marked as Cnq;

[0063] Obtain the target model algorithm output, the supplier similarity corresponding to the misjudged suppliers, and the similarity corresponding to the actually adopted suppliers;

[0064] The supplier similarity corresponding to the misjudged supplier is subtracted from the similarity corresponding to the actually adopted supplier to obtain the misjudged similarity difference;

[0065] Obtain the misjudgment similarity differences outputted multiple times by the target model algorithm, sum and average the misjudgment similarity differences outputted multiple times to obtain the misjudgment similarity mean, and mark the misjudgment similarity mean as Wp;

[0066] Obtain the misjudged suppliers output by the target model algorithm, and compare and analyze the misjudged suppliers with the supplier categories output by the hierarchical clustering algorithm;

[0067] If the supplier is misjudged as a similar supplier, obtain the similarity characterization value Ts of the misjudged supplier;

[0068] Based on the adopted weight ratio Cnq, the misjudgment similarity mean Wp, and the misjudgment supplier similarity representation value Ts, the training enhancement ratio Xlq of the target model algorithm is calculated;

[0069] By formula: Obtain the training enhancement ratio Xlq of the target model algorithm, where α=0.561, β=0.487, z is the total number of misjudged suppliers output by the target model algorithm, c=1, 2, ..., z; if the misjudged supplier is a similar supplier, then γ=1, otherwise, γ=0.

[0070] As a further technical solution of the present invention: the supplier macro model corresponding to the recommendation model algorithm is determined in the following manner:

[0071] Based on the dynamically adjusted target model algorithm and the mean error ratio of the target model algorithm, all the mean error ratios of the target model algorithms are arranged in ascending order, and the target model algorithm corresponding to the minimum mean error ratio is selected as the recommended model algorithm;

[0072] The recommended suppliers are obtained as follows:

[0073] Based on the supplier model constructed by the recommendation model algorithm, the recommended suppliers are output.

[0074] A supplier recommendation system based on a large model, including:

[0075] Data collection module: obtain historical supplier data and store it in the supplier data group, perform similarity analysis on the supplier data group, and build a supplier similarity matrix ;

[0076] Classification and discrimination module: based on supplier similarity matrix , build a supplier classification model to identify the same suppliers in different categories;

[0077] Data analysis module: This module performs numerical analysis on the supplier similarity of suppliers of the same type in the same category to obtain the central similarity difference Zx. It also obtains the historical supply count of suppliers of the same category and performs numerical calculations to obtain the similarity characterization value Ts. Suppliers of the same type are ranked based on the similarity characterization value Ts to construct a model training data set.

[0078] Model building module: Based on the model training data set, different algorithms are used to build a supplier model. Based on the supplier model output, recommended suppliers are recommended, and comparative analysis is performed to determine the target model algorithm.

[0079] Model determination module: Based on the target model algorithm, the output results of the target model algorithm are numerically analyzed to obtain the training enhancement ratio Xlq of the target model algorithm, and the training times of the supplier large model of the target model algorithm are dynamically adjusted. Based on the dynamically adjusted target algorithm model, the supplier large model corresponding to the recommended model algorithm is determined, and the recommended supplier is determined.

[0080] Beneficial effects of the present invention:

[0081] Through in-depth mining and analysis of historical supplier data, we build supplier similarity matrices, classification models, and supplier big models built using algorithms. We can identify and screen suppliers that meet the needs of the enterprise. By analyzing multiple dimensional factors such as product category, timeliness, and number of deliveries, we select multiple algorithms to build a supplier big model. Through complex calculations and model training, we continuously optimize the accuracy of the big model's supplier recommendations.

[0082] Based on the training reinforcement ratio of the target model algorithm, the training times of the supplier large model are dynamically adjusted. In actual application, as the market environment and corporate procurement needs continue to change, the model can continuously adapt to new data features and business needs. The dynamic optimization mechanism enables the model to maintain a high level of prediction accuracy and recommendation quality, and can provide enterprises with the most appropriate supplier recommendation solutions under different market conditions and corporate development stages. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] The present invention will be further described below with reference to the accompanying drawings.

[0084] Figure 1 is a flow chart of a supplier recommendation method based on a large model of the present invention;

[0085] Figure 2 This is a device module diagram of a supplier recommendation system based on a large model of the present invention;

[0086] Figure 3 The present invention provides a structural diagram of a device and storage medium for a supplier recommendation system based on a large model. DETAILED DESCRIPTION

[0087] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0088] Example 1: Please refer to Figure 1 As shown, the present invention is a supplier recommendation method based on a large model, comprising:

[0089] Step 1: Obtain historical supplier data and store it in the supplier data group, perform similarity analysis on the supplier data group, and build a supplier similarity matrix ;

[0090] Based on the enterprise's internal procurement management system, the historical transaction record data between the enterprise and different suppliers is obtained, and the historical transaction record data of different suppliers is used as the supplier data set a i ;

[0091] The supplier data group contains historical transaction records of multiple suppliers, including data such as the type of goods supplied by the supplier and the timeliness.

[0092] Where i is the supplier number, i=1,2,...,n, and n is the number of suppliers;

[0093] Put all supplier data sets into supplier data group, and build supplier data group A, A={a1, a2, ...a n};

[0094] Extract supplier data set a i The goods category data in the i , calculate the supplier similarity G of different supplier data sets ij ;

[0095] By formula: Get the supplier similarity G of the supplier data set ij ;

[0096] where j=i+1,i+2,...,n, Represents the supplier data set c i with c j The number of intersections, Represents the supplier data set c i with c j The number of unions, X 12 The supplier similarity G between two suppliers numbered 1 and 2 ij ;

[0097] Based on the supplier similarity G in different supplier data ij , construct supplier similarity matrix S ij ;

[0098] Among them, the supplier similarity matrix is constructed as S ij = ;

[0099] The supplier similarity matrix S ij Perform preprocessing to remove the supplier similarity matrix S ij The meaningless data in the matrix, including the matrix diagonal and the upper triangle data, mark the pre-processed supplier similarity matrix as ;

[0100] It should be noted that the matrix diagonal G ii Elements, such as G 11 Indicates the supplier similarity of the product category c1 in the supplier data set. The upper triangle elements and the lower triangle elements of the matrix have the same meaning. For example, G 12 , G 21 The supplier similarity between the goods category c1 in the supplier data set and the goods category c2 in the supplier data set has the same meaning, so it is removed;

[0101] For example, ;

[0102] Step 2: Based on the supplier similarity matrix , build a supplier classification model to identify the same suppliers in different categories;

[0103] Supplier Similarity Matrix Use hierarchical clustering algorithm to identify unique suppliers and homogeneous suppliers;

[0104] Specifically, the supplier classification model is constructed as follows:

[0105] S1. Calculate the similarity distance d(i,j) using the formula: Get the similarity distance d(i,j), and build a similarity distance matrix based on the similarity distance d(i,j);

[0106] It should be noted that the smaller the similarity distance d(i, j), the more similar the two suppliers are;

[0107] S2. Input the supplier matrix into the hierarchical clustering algorithm. The hierarchical clustering algorithm establishes different clusters based on the similarity distance d(i, j), calculates the distance between each cluster, and iterates each cluster until the distance between each cluster is less than the cluster threshold;

[0108] It should be noted that, in the initial state, the hierarchical clustering algorithm will be each supplier a i Treat it as a cluster, calculate the similarity distance d(i, j) between all clusters, find the two clusters with the closest similarity distance, merge the two clusters into one cluster, update the similarity distance between clusters, and stop the merging operation when the similarity distance between clusters is less than the cluster threshold. The cluster threshold is set by professionals in this field.

[0109] S3. Analyze the clusters output by the hierarchical clustering algorithm. If there is only one supplier in the cluster, mark the corresponding supplier as the only supplier. If there are multiple suppliers in the cluster, mark the suppliers in the cluster as the same type of suppliers.

[0110] It should be noted that there are multiple clusters in the hierarchical clustering algorithm, that is, there are different categories of the same suppliers;

[0111] The same type of suppliers belonging to the same cluster are marked as b ki , k = 1, 2, ..., m, m is the number of remaining clusters after removing the cluster containing the only supplier in the hierarchical clustering algorithm, that is, the number of categories of the same supplier, i is the supplier number;

[0112] For example, when k=1, i=1,2, b 11 、b 12 Indicates that two suppliers of the same type are both in category k 1, b 11 、b12 They are suppliers of the same type;

[0113] The technical solution of this embodiment is to obtain historical supplier data and store it in a supplier data group, perform similarity analysis on the supplier data group, and construct a supplier similarity matrix. , based on the supplier similarity matrix , build a supplier classification model, identify the same suppliers in different categories, and provide data support for subsequent supplier analysis.

[0114] Example 2: Please refer to Figure 1 As shown, the supplier recommendation method based on the large model of the present invention further includes:

[0115] Step 3: Perform a numerical analysis on the supplier similarity of the same type of suppliers in the same category to obtain the central similarity difference Zx. Combined with the historical supply times of the same category and numerically calculated, the same type characterization value Ts is obtained. Based on the same type characterization value Ts, the same type of suppliers are ranked and a model training data set is constructed.

[0116] Get the historical supply times of the same supplier in the same category;

[0117] The historical supply times of all suppliers of the same type in the same category;

[0118] The historical supply times of the same supplier are compared with the historical supply times of all suppliers of the same type to obtain the same type supply ratio, which is marked as T1;

[0119] Obtain the center similarity of the same type of suppliers in the same category;

[0120] Specifically, the center similarity is obtained as follows:

[0121] The similarities of suppliers of the same type within the same category are arranged in descending order, and the sorted suppliers of the same type are used as elements and placed in the similarity sequence of the same type;

[0122] Get the total number of elements in the same similarity sequence, mark the total number of elements as Zs, and perform parity judgment on the total number of elements;

[0123] If the total number of elements Zs is an odd number, get the similarity sequence of the same type. The similarity of suppliers is used as the central similarity;

[0124] It should be noted that Indicates that the total number of elements Zs is ratioed to 2 and rounded up. For example, if the total number of elements Zs is 5, the ratio of Zs to 2 is 2.5, which is rounded up to 3. The similarity of the third supplier in the similarity sequence is taken as the central similarity.

[0125] If the total number of elements Zs is an even number, obtain the similarity between the (Zs / 2)th and (Zs / 2+1)th suppliers in the similarity sequence, sum and average the similarities between the (Zs / 2)th and (Zs / 2+1)th suppliers in the similarity sequence, and obtain the central similarity.

[0126] Based on the similarity sequence of the same type, calculate the difference between each element and the center similarity of the similarity sequence to obtain the center similarity difference, and mark the center similarity difference as Zx;

[0127] Based on the center similarity difference Zx and the similar supply ratio TI, the similar characterization value Ts is calculated;

[0128] By formula: Obtain the same type of characterization value Ts, f = 0.683, g = 0.789;

[0129] Sort the elements of the same level of detail sequence in descending order of the same level of characterization value Ts;

[0130] Get all unique suppliers and put them into the unique supplier sequence;

[0131] Obtain similarity sequences and unique supplier sequences of all categories that have been sorted, and construct a model training data set;

[0132] It should be noted that the elements in the similarity sequence are sorted according to the similarity characterization value Ts. The priority of suppliers in the same category is determined by sorting. The similarity characterization value Ts comprehensively considers the central similarity difference Zx and the similar supply ratio Tl, which can more comprehensively reflect the importance of the supplier in the category. Through sorting, suppliers with larger similarity characterization values Ts can be regarded as more critical suppliers in the corresponding category, providing data support for the subsequent construction of the supplier model.

[0133] Step 4: Based on the model training data set, use different algorithms to build a supplier model, output recommended suppliers based on the supplier model, and conduct comparative analysis to determine the target model algorithm;

[0134] Based on the model training data set, different algorithms are used to build a large supplier model;

[0135] Among them, the supplier model is constructed by different algorithms such as graph convolutional network (GCN), graph attention network (GAT), multi-layer perceptron (MLP);

[0136] It should be noted that the graph convolutional network can use suppliers and buyers as nodes in the graph, and purchase behaviors as edges. Using the convolution operation of the graph, it learns the supplier similarity and timeliness of the supplier nodes. Since the model training data set sorts similar suppliers, the sorting order serves as the weight of each supplier in the model training. The weight is input into the graph convolutional network and the model is trained to obtain recommended suppliers.

[0137] The graph attention network uses the attention mechanism to learn the weights between nodes, integrating supplier similarity and timeliness into the attention mechanism to enhance the influence of similar suppliers. The graph attention network can capture the complex relationship between buyers and suppliers, and thus obtain recommended suppliers.

[0138] The multi-layer perceptron model is used to learn the nonlinear relationship between suppliers and goods. It uses supplier similarity and timeliness as input features. After encoding the supplier similarity and timeliness information of similar suppliers, it is combined with the feature information of other suppliers and input into the multi-layer perceptron model for training to obtain recommended suppliers.

[0139] A supplier model built based on different algorithms takes into account the type of goods and timeliness, outputs recommended suppliers, and compares the recommended suppliers with the actual suppliers.

[0140] If the recommended supplier is inconsistent with the actual adopted supplier, the supplier will be marked as a misjudged supplier;

[0141] If the recommended supplier is consistent with the actual adopted supplier, the supplier will be marked as the adopted supplier;

[0142] Obtain the number of misjudged suppliers in the supplier macro models constructed using different algorithms, as well as the number of recommended suppliers output by the supplier macro models;

[0143] The number of misjudged suppliers is compared with the number of recommended suppliers output by the suppliers to obtain the misjudgment ratio.

[0144] Calculate the supplier model constructed by different algorithms, output recommended suppliers and misjudgment ratios multiple times, and sum and average the misjudgment ratios output multiple times to obtain the mean misjudgment ratio.

[0145] Compare the mean false positive ratio with the false positive threshold to determine the target model algorithm;

[0146] If the mean false positive ratio is higher than the false positive threshold, the algorithm corresponding to the supplier's large model is obtained and marked as a non-target model algorithm;

[0147] If the mean false positive ratio is lower than the false positive threshold, the algorithm corresponding to the supplier's large model is obtained, and the model constructed by the corresponding algorithm is marked as the target model algorithm;

[0148] It should be noted that there are multiple target model algorithms;

[0149] Step 5: Based on the target model algorithm, perform numerical analysis on the output results of the target model algorithm to obtain the training enhancement ratio Xlq of the target model algorithm. Based on the training enhancement ratio Xlq of the target model algorithm, dynamically adjust the number of supplier large model training times of the target model algorithm. Based on the dynamically adjusted target algorithm model, determine the supplier large model corresponding to the recommended model algorithm and determine the recommended supplier.

[0150] Obtain the number of adopted suppliers output by the target model algorithm and the number of recommended suppliers output by the target model algorithm;

[0151] The number of adopted suppliers output by the target model algorithm is compared with the number of recommended suppliers output by the target model algorithm to obtain the adoption-supply ratio;

[0152] Obtain the adoption supply ratios of all target model algorithms, sum the adoption supply ratios of all target algorithms, and obtain the total adoption ratio value;

[0153] The adoption supply ratio is compared with the total adoption ratio to obtain the adoption weight ratio, which is marked as Cnq;

[0154] Obtain the target model algorithm output, the supplier similarity corresponding to the misjudged suppliers, and the similarity corresponding to the actually adopted suppliers;

[0155] The supplier similarity corresponding to the misjudged supplier is subtracted from the similarity corresponding to the actually adopted supplier to obtain the misjudged similarity difference;

[0156] Obtain the misjudgment similarity differences outputted multiple times by the target model algorithm, sum and average the misjudgment similarity differences outputted multiple times to obtain the misjudgment similarity mean, and mark the misjudgment similarity mean as Wp;

[0157] Obtain the misjudged suppliers output by the target model algorithm, and compare and analyze the misjudged suppliers with the supplier categories output by the hierarchical clustering algorithm;

[0158] If a supplier is mistakenly identified as the sole supplier, no action will be taken;

[0159] If the supplier is misjudged as a similar supplier, obtain the similarity characterization value Ts of the misjudged supplier;

[0160] Based on the adopted weight ratio Cnq, the misjudgment similarity mean Wp, and the misjudgment supplier similarity representation value Ts, the training enhancement ratio Xlq of the target model algorithm is calculated;

[0161] By formula: Obtain the training enhancement ratio Xlq of the target model algorithm, where α = 0.561, β = 0.487, z is the total number of misjudged suppliers output by the target model algorithm, c = 1, 2, ..., z; if the misjudged supplier is a similar supplier, then γ = 1, otherwise, γ = 0;

[0162] It should be noted that the adoption weight ratio Cnq highlights the relative advantages or disadvantages of a single target model algorithm in the overall set. When the adoption weight ratio Cnq is high, it means that the adoption supply ratio of the target model algorithm is relatively prominent among all target model algorithms, and it accounts for a large proportion of the overall recommendation effect. Therefore, the target algorithm with a higher adoption weight ratio Cnq should require fewer additional training times.

[0163] The mean false similarity value Wp quantifies the difference in similarity between misjudged suppliers and actual suppliers. A higher mean false similarity value Wp indicates a higher average level of false similarity differences across multiple outputs of the target model algorithm. This indicates that the model has a larger error in judging supplier similarity and has significant room for improvement, requiring increased model training times.

[0164] If a supplier is misjudged as similar, a high Ts value indicates that the supplier is highly important or unique within its category. When calculating the Ts value, combined with other parameters, a high Ts value will cause the model to focus more on learning and adjusting the relevant features of this type of supplier, thereby increasing the number of training reinforcements.

[0165] Obtain the number of supplier large model training times Mxc of the target model algorithm, calculate the additional training times of the target model algorithm, and mark the additional training times of the target model algorithm as Ew;

[0166] By formula: Get the additional training times of the target model algorithm;

[0167] Based on the additional training times of the target model algorithm, the training times of the supplier large model constructed by the target model algorithm are dynamically adjusted;

[0168] Based on the dynamically adjusted target model algorithm and the mean error ratio of the target model algorithm, all the mean error ratios of the target model algorithms are arranged in ascending order, and the target model algorithm corresponding to the minimum mean error ratio is selected as the recommended model algorithm;

[0169] Output recommended suppliers based on the supplier model built by the recommendation model algorithm;

[0170] The technical solution of this embodiment is: numerically analyze the supplier similarity of the same type of suppliers in the same category to obtain the central similarity difference Zx, obtain the historical supply times of suppliers in the same category, and perform numerical calculations to obtain the same type characterization value Ts, sort the same type of suppliers based on the same type characterization value Ts, construct a model training data group, and use different algorithms to construct a supplier big model based on the model training data group, output recommended suppliers based on the supplier big model, and perform comparative analysis to determine the target model algorithm, and based on the target model algorithm, perform numerical analysis on the output result of the target model algorithm to obtain the training enhancement ratio Xlq of the target model algorithm, and dynamically adjust the training times of the supplier big model of the target model algorithm, and determine the supplier big model corresponding to the recommended model algorithm based on the dynamically adjusted target algorithm model, and determine the recommended supplier, and continuously optimize the recommendation results through complex calculations and model training, thereby greatly improving the accuracy and reliability of the recommended suppliers.

[0171] Example 3: Please refer to Figure 3 As shown, the supplier recommendation system based on the large model of the present invention includes the following device modules:

[0172] Data collection module: obtain historical supplier data and store it in the supplier data group, perform similarity analysis on the supplier data group, and build a supplier similarity matrix ;

[0173] Classification judgment module: Based on the supplier similarity matrix, a supplier classification model is constructed to identify the same suppliers in different categories;

[0174] Data analysis module: This module performs a numerical analysis on the supplier similarity of suppliers of the same type in the same category to obtain the central similarity difference Zx. The module also obtains the historical supply times of suppliers of the same category and performs numerical calculations to obtain the same-category supply ratio TI. The module then analyzes the central similarity difference Zx and the same-category supply ratio TI to obtain the same-category characterization value Ts. Suppliers of the same type are ranked based on the same-category characterization value Ts to construct a model training data set.

[0175] Model building module: Based on the model training data set, different algorithms are used to build a supplier model. Based on the supplier model output, recommended suppliers are recommended, and comparative analysis is performed to determine the target model algorithm.

[0176] Model determination module: Based on the target model algorithm, the output results of the target model algorithm are numerically analyzed to obtain the training enhancement ratio Xlq of the target model algorithm. Based on the training enhancement ratio Xlq of the target model algorithm, the number of supplier large model training times of the target model algorithm is dynamically adjusted. Based on the dynamically adjusted target algorithm model, the supplier large model corresponding to the recommended model algorithm is determined, and the recommended supplier is determined.

[0177] Example 4: Please refer to Figure 3 As shown, an embodiment of the present invention provides a computer device 3, comprising a memory 302, a processor 301 and a computer program 303 stored in the memory 302;

[0178] Processor 301 uses a central processing unit (CPU) to handle basic and data-intensive tasks such as building a supplier similarity matrix. During the construction process, a large amount of supplier data needs to be extracted, sorted, and analyzed, and the CPU can quickly complete these operations. The digital signal processor (DSP) focuses on processing time-related data processing tasks. For example, when analyzing supplier delivery timeliness data, the DSP can extract timeliness features and perform quantitative analysis. Application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), and other programmable logic devices can be highly customized based on the complex logic and special computing requirements of the supplier recommendation method of the present invention. When running model algorithms such as graph convolutional networks (GCNs), graph attention networks (GATs), and multi-layer perceptrons (MLPs), these customized processors can optimize the model structure and data flow at the hardware level, significantly improving the computing speed and accuracy, thereby ensuring the efficient operation of the entire supplier recommendation process.

[0179] The memory 302 serves as a storage medium for storing information related to basic software such as the operating system, application programs, and boot loaders. It also stores the program code of the supplier recommendation system involved in the present invention. These program codes, when called by the processor 301, drive the execution of the entire supplier recommendation process. At the same time, the internal memory also stores a large amount of historical supplier data obtained from the enterprise procurement management system. These data cover detailed information such as the categories of goods supplied by the supplier, the timeliness of each supply, the number of supplies, etc. After the data acquisition module obtains the data, the data will first be stored in the internal memory, and then read and updated during the operation of the classification judgment module, data analysis module, etc. For example, when constructing a supplier similarity matrix, it is necessary to read the supplier's product category data from the internal memory for similarity calculation; when analyzing the supply situation of the same supplier, it is necessary to extract the corresponding supply frequency and timeliness data from the internal memory for analysis and statistics.

[0180] Through the collaborative work of computer equipment 3, storage media and devices, computing support, data storage environment and function implementation mechanism are provided for the supplier recommendation system based on the large model.

[0181] The above is a detailed description of an embodiment of the present invention. However, the content is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A supplier recommendation method based on a large model, characterized by: The following steps are involved: Step 1: Obtain historical supplier data and store it in a supplier data group, perform similarity analysis on the supplier data group, and build a supplier similarity matrix ; Step 2: Based on the supplier similarity matrix , build a supplier classification model to identify the same suppliers in different categories; Step 3: Perform a numerical analysis on the supplier similarity of the same type of suppliers in the same category to obtain the central similarity difference Zx. Obtain the historical supply times of suppliers in the same category and perform numerical calculations to obtain the same type characterization value Ts. Rank the same type of suppliers based on the same type characterization value Ts and construct a model training data set. The same type characterization value Ts is obtained as follows: Obtain the historical supply counts of the same type of supplier within the same category, and the historical supply counts of all suppliers of the same type; The historical supply times of the same supplier are compared with the historical supply times of all suppliers of the same type to obtain the same type supply ratio, which is marked as T1; Arrange the supplier similarities of the same type of suppliers in the same category from largest to smallest, and place the sorted suppliers as elements in the similarity sequence; Get the total number of elements in the same similarity sequence, mark the total number of elements as Zs, and perform parity judgment on the total number of elements; If the total number of elements Zs is an odd number, get the similarity sequence of the same type. The similarity of suppliers is used as the central similarity; If the total number of elements Zs is an even number, obtain the similarity between the (Zs / 2)th and (Zs / 2+1)th suppliers in the similarity sequence, sum and average the similarities between the (Zs / 2)th and (Zs / 2+1)th suppliers in the similarity sequence, and obtain the central similarity. Based on the similarity sequence of the same type, calculate the difference between each element and the center similarity of the similarity sequence to obtain the center similarity difference, and mark the center similarity difference as Zx; Based on the center similarity difference Zx and the similar supply ratio TI, calculate the similar characterization value Ts; By formula: Get the same type characterization value Ts, f and g are preset proportional coefficients; Step 4: Based on the model training data set, use different algorithms to build a supplier model, output recommended suppliers based on the supplier model, and conduct comparative analysis to determine the target model algorithm; Step 5: Based on the target model algorithm, perform numerical analysis on the output results of the target model algorithm to obtain the training enhancement ratio Xlq of the target model algorithm. Based on the training enhancement ratio Xlq of the target model algorithm, dynamically adjust the number of training times of the supplier large model of the target model algorithm. Based on the dynamically adjusted target algorithm model, determine the supplier large model corresponding to the recommended model algorithm and determine the recommended supplier.

2. The supplier recommendation method based on a large model according to claim 1, characterized in that: The Supplier Similarity Matrix The way to obtain is: Based on the enterprise's internal procurement management system, the historical transaction record data between the enterprise and different suppliers is obtained, and the historical transaction record data of different suppliers is used as the supplier data set a i ; The historical transaction records include the supplier's product category and timeliness. i is the supplier's ID, i = 1, 2, ..., n, where n is the number of suppliers. Put all supplier data sets into supplier data group, and construct supplier data group A, A={a1, a2, ...a n }; Extract supplier data set a i The goods category data in the i , calculate the supplier similarity G of different supplier data sets ij ; By formula: Get the supplier similarity G of the supplier data set ij ; where j=i+1,i+2,...,n, Represents the supplier data set c i with c j The number of intersections, Represents the supplier data set c i with c j The number of unions, X 12 The supplier similarity G between two suppliers numbered 1 and 2 ij ; Based on the supplier similarity G in different supplier data ij , construct supplier similarity matrix S ij ; Construct the supplier similarity matrix S ij = ; The supplier similarity matrix S ij Perform preprocessing to remove the supplier similarity matrix S ij The meaningless data in the matrix, including the matrix diagonal and the upper triangle data, mark the pre-processed supplier similarity matrix as .

3. The supplier recommendation method based on a large model according to claim 1, characterized in that: The methods for obtaining the same suppliers of different categories are as follows: Supplier Similarity Matrix Use hierarchical clustering algorithm to identify unique suppliers and homogeneous suppliers; The supplier classification model is constructed as follows: S1. Calculate the similarity distance d(i,j) using the formula: Get the similarity distance d(i,j), and build a similarity distance matrix based on the similarity distance d(i,j); S2. Input the supplier matrix into the hierarchical clustering algorithm. The hierarchical clustering algorithm establishes different clusters based on the similarity distance d(i, j), calculates the distance between each cluster, and iterates each cluster. S3. Analyze the clusters output by the hierarchical clustering algorithm. If there is only one supplier in the cluster, mark the corresponding supplier as the only supplier. If there are multiple suppliers in the cluster, mark the suppliers in the cluster as the same type of suppliers.

4. The supplier recommendation method based on a large model according to claim 1, characterized in that: The model training data set is constructed as follows: Sort the elements of the similarity sequence according to the order of the similarity characterization value Ts from large to small; Get all unique suppliers and put them into the unique supplier sequence; Obtain similarity sequences and unique supplier sequences of all categories that have been sorted, and build a model training data set.

5. The supplier recommendation method based on a large model according to claim 1, characterized in that: The target model algorithm is constructed as follows: Based on the model training data set, different algorithms are used to build a large supplier model; A supplier model built based on different algorithms takes into account the type of goods and timeliness, outputs recommended suppliers, and compares the recommended suppliers with the actual suppliers. If the recommended supplier is inconsistent with the actual adopted supplier, the supplier will be marked as a misjudged supplier; If the recommended supplier is consistent with the actual adopted supplier, the supplier will be marked as the adopted supplier; Obtain the number of misjudged suppliers in the supplier macro models constructed using different algorithms, as well as the number of recommended suppliers output by the supplier macro models; The number of misjudged suppliers is compared with the number of recommended suppliers output by the suppliers to obtain the misjudgment ratio. Calculate the supplier model constructed by different algorithms, output recommended suppliers and misjudgment ratios multiple times, and sum and average the misjudgment ratios output multiple times to obtain the mean misjudgment ratio. If the mean false positive ratio is lower than the false positive threshold, the algorithm corresponding to the supplier's large model is obtained, and the model constructed by the corresponding algorithm is marked as the target model algorithm.

6. The supplier recommendation method based on a large model according to claim 1, characterized in that: The method for dynamically adjusting the number of supplier large model training times of the target model algorithm is as follows: Obtain the number of supplier large model training times Mxc of the target model algorithm, calculate the additional training times of the target model algorithm, and mark the additional training times of the target model algorithm as Ew; Calculate the additional training times Ew of the target model algorithm based on the additional training times Ew of the target model algorithm and the training enhancement ratio Xlq of the target model algorithm; By formula: Get the additional training times of the target model algorithm; Based on the additional training times of the target model algorithm, the training times of the supplier large model constructed by the target model algorithm are dynamically adjusted.

7. The supplier recommendation method based on a large model according to claim 6, characterized in that: The training enhancement ratio Xlq of the target model algorithm is obtained as follows: Obtain the number of adopted suppliers output by the target model algorithm and the number of recommended suppliers output by the target model algorithm; The number of adopted suppliers output by the target model algorithm is compared with the number of recommended suppliers output by the target model algorithm to obtain the adoption-supply ratio; Obtain the adoption supply ratios of all target model algorithms, sum the adoption supply ratios of all target algorithms, and obtain the total adoption ratio value; The adoption supply ratio is compared with the total adoption ratio to obtain the adoption weight ratio, which is marked as Cnq; Obtain the target model algorithm output misjudged suppliers, the corresponding supplier similarity, and the similarity corresponding to the actually adopted suppliers; The supplier similarity corresponding to the misjudged supplier is subtracted from the similarity corresponding to the actually adopted supplier to obtain the misjudged similarity difference; Obtain the misjudgment similarity differences outputted multiple times by the target model algorithm, sum and average the misjudgment similarity differences outputted multiple times to obtain the misjudgment similarity mean, and mark the misjudgment similarity mean as Wp; Obtain the misjudged suppliers output by the target model algorithm, and compare and analyze the misjudged suppliers with the supplier categories output by the hierarchical clustering algorithm; If the supplier is misjudged as a similar supplier, obtain the similarity characterization value Ts of the misjudged supplier; Based on the adopted weight ratio Cnq, the misjudgment similarity mean Wp, and the misjudgment supplier similarity representation value Ts, the training enhancement ratio Xlq of the target model algorithm is calculated; By formula: Obtain the training enhancement ratio Xlq of the target model algorithm, where α and β are preset proportional coefficients, z is the total number of misjudged suppliers output by the target model algorithm, c=1, 2, ..., z; if the misjudged supplier is a similar supplier, then γ=1, otherwise, γ=0.

8. The supplier recommendation method based on a large model according to claim 1, characterized in that: The method for determining the supplier model corresponding to the recommendation model algorithm is as follows: Based on the dynamically adjusted target model algorithm and the mean error ratio of the target model algorithm, all the mean error ratios of the target model algorithms are arranged in ascending order, and the target model algorithm corresponding to the minimum mean error ratio is selected as the recommended model algorithm; The recommended suppliers are obtained as follows: Based on the supplier model constructed by the recommendation model algorithm, the recommended suppliers are output.

9. A supplier recommendation system based on a large model, used to implement a supplier recommendation method based on a large model according to any one of claims 1 to 8, characterized in that: Includes the following modules: Data collection module: obtain historical supplier data and store it in the supplier data group, perform similarity analysis on the supplier data group, and build a supplier similarity matrix ; Classification and discrimination module: based on supplier similarity matrix , build a supplier classification model to identify the same suppliers in different categories; Data analysis module: This module performs numerical analysis on the supplier similarity of suppliers of the same type in the same category to obtain the central similarity difference Zx. It also obtains the historical supply count of suppliers of the same category and performs numerical calculations to obtain the similarity characterization value Ts. Suppliers of the same type are ranked based on the similarity characterization value Ts to construct a model training data set. Model building module: Based on the model training data set, different algorithms are used to build a supplier model. Based on the supplier model output, recommended suppliers are recommended, and comparative analysis is performed to determine the target model algorithm. Model determination module: Based on the target model algorithm, perform numerical analysis on the output results of the target model algorithm to obtain the training enhancement ratio Xlq of the target model algorithm. Based on the training enhancement ratio Xlq of the target model algorithm, dynamically adjust the number of training times of the supplier large model of the target model algorithm. Based on the dynamically adjusted target algorithm model, determine the supplier large model corresponding to the recommended model algorithm and determine the recommended supplier.

Images