Supplier recommendation system and method based on supply chain network demand and preference mining

Abstract

This invention provides a supplier recommendation method and system based on supply chain network demand and preference mining, comprising: Step S1: using a bipartite graph network structure to describe the supply chain business relationship between buyers and suppliers; Step S2: mining the buyer's product demand and supplier preferences from two types of text information: supply chain inquiry forms and quotation forms; Step S3: using a bipartite graph collaborative filtering algorithm to recall candidate recommended suppliers from the bipartite graph network structure used to describe the supply chain business relationship between buyers and suppliers, and then performing precise filtering based on the buyer's product demand and supplier preferences to select recommended suppliers.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce platforms, and specifically, to a supplier recommendation system and method based on supply chain network demand and preference mining, and more specifically, to an e-commerce platform supplier recommendation method and system based on product matching and supply chain network analysis. Background Art

[0002] Supply and demand matching is a core problem in the supply chain field. In traditional supply chain business scenarios, matching between supply and demand is mostly achieved through precise matching between texts and manual experience. However, the traditional text-based precise matching method only considers the superficial matching degree of text characters, without considering the deep semantic information contained in the text, and the matching based on manual experience is difficult to consider the global knowledge and information of the supply chain due to individual differences. It is precisely because of the above limitations, uncertainties, and operational risks of these two methods that the needs and preferences of buyers are often difficult to be fully explored, resulting in less than ideal supply and demand matching results.

[0003] Patent document CN106021590B (application number: 201610395086.1) discloses a multi-ant colony parallel combination clustering method based on hypergraph, which belongs to the field of data mining technology. The present invention aims at the fast and accurate clustering of high-dimensional massive data, combines the improved single ant colony clustering algorithm based on information entropy with the hypergraph, and its main idea is that different ant colonies use the improved single ant colony clustering algorithm based on information entropy to work independently and in parallel, and combine the clustering results through the hypergraph to obtain different similarity matrices, and then exchange information and cooperate with each other, and use the similarity matrix to correct the single ant colony clustering process; after repeated iterations, the optimal clustering result is finally obtained. The present invention also discloses a B2B platform supplier recommendation method and system. The present invention can further improve the algorithm search efficiency when facing high-dimensional massive data, and effectively overcome the defect that the prior art is prone to fall into the local optimal solution too early, and improve the accuracy of the clustering results.

[0004] In order to fully grasp the global knowledge and information of the supply chain, fully explore the needs and preferences of supply chain buyers, break through the limitations of traditional matching methods, and improve the effect of supply and demand matching, the present invention provides a supplier recommendation system and method based on supply chain network demand and preference mining. Summary of the invention

[0005] In view of the defects in the prior art, the object of the present invention is to provide a supplier recommendation system and method based on supply chain network demand and preference mining.

[0006] A supplier recommendation method based on supply chain network demand and preference mining provided by the present invention includes:

[0007] Step S1: Using a bipartite graph network structure to describe the supply chain business relationship between buyers and suppliers;

[0008] Step S2: mining buyers’ commodity demands and supplier preferences from two types of text information, supply chain inquiry forms and quotation forms;

[0009] Step S3: The bipartite graph network structure used to describe the supply chain business relationship between buyers and suppliers is used to recall candidate recommended suppliers using a bipartite graph collaborative filtering algorithm, and then precise filtering is performed based on the buyer's product demand and supplier preference to select recommended suppliers.

[0010] Preferably, the bipartite graph network structure includes: a set of buyer nodes and a set of supplier nodes, and an edge connection is established when a buyer node and a supplier node have a supply chain business interaction relationship;

[0011] The bipartite graph network structure is expressed as G=(U,I,E), where U represents the buyer node set, I represents the supplier node set, and E represents the edge set; assuming that the number of buyer nodes |U|=m, and the number of supplier nodes |I|=n, the total number of bipartite graph network nodes is r=m+n; if the buyer u_p is associated with the supplier i_q, the value of G(p,q) is the number of associated orders; otherwise, G(p,q)=0.

[0012] Preferably, the step S2 adopts:

[0013] Step S2.1: Convert the inquiry form into the buyer's inquiry text embedding vector, convert the quotation form into the supplier's quotation text embedding vector, and calculate the vector similarity to obtain the supply product that is closest to the buyer's needs;

[0014] Step S2.2: Use the ALBERT model to semantically encode the text information of the technical terms in the inquiry to obtain a semantic vector, and then input the semantic vector as a feature into the TextCNN multi-label classification model. Use the TextCNN multi-label classification model to predict the probability distribution of buyers' preferences for supplier company qualifications, and finally realize intelligent mining of supplier preferences of different buyers.

[0015] Preferably, the step S2.1 adopts:

[0016] Step S2.1.1: preprocessing the inquiry form, and segmenting the preprocessed inquiry form to obtain a segmentation sequence;

[0017] Step S2.1.2: Count the TF-IDF scores of each word segment and obtain the word-text weight matrix after normalization;

[0018] Step S2.1.3: Use the trained Word2vec word embedding model to encode the word segmentation sequence to obtain word embedding;

[0019] Step S2.1.4: Perform weighted summation based on word embedding and word-text weight matrix to obtain the buyer inquiry text embedding vector.

[0020] Preferably, the step S3 adopts:

[0021] Step S3.1: Assume that the buyer set is P, the supplier set is S, and the number of suppliers to be recalled is k. Use the bipartite graph collaborative filtering algorithm to filter out the target buyer P. i The k suppliers with the most similar historical purchasing characteristics are selected as the candidate supplier set S_1 to be recommended;

[0022] Step S3.2: Filter out the suppliers in the set S_1 to obtain a supplier subset S_2 whose historical quotation texts and the target buyer p's inquiry texts meet the preset requirements;

[0023] Step S3.3: Filter the suppliers in set S_2 to find a subset S_3 of suppliers whose qualification labels meet the buyer's preference for the qualifications of cooperative suppliers;

[0024] Step S3.4: Set S_3 is used as the final result set of personalized supplier recommendations for buyer p.

[0025] Preferably, the step S3.1 adopts:

[0026] Step S3.1.1: Initialize the matrix M, the historical preference matrix D, and the relative error tolerance ε;

[0027] Step S3.1.2: Let IR(i,j)=1 / n,(i=1,...,m,j=1,...,n);

[0028] Step S3.1.3: When ||αM·IR+(1-α)D-IR||≥ε is satisfied, IR=αM·IR+(1-α)D; repeat step S3.1.3 until ||αM·IR+(1-α)D-IR||≥ε is not satisfied, and the target purchaser P is obtained. i The k suppliers with the most similar historical purchasing characteristics are selected as the candidate supplier set S_1 to be recommended;

[0029] Among them, αM·IR represents the potential preference brought by network communication; (1-α)D represents the observed historical preference; α represents the parameter measuring the attenuation level of the network; D represents the buyer's historical preference matrix; M represents the number of two different buyers choosing the same supplier; IR represents the buyer's potential preference matrix.

[0030] A supplier recommendation system based on supply chain network demand and preference mining provided by the present invention includes:

[0031] Module M1: Use a bipartite graph network structure to describe the supply chain business relationship between buyers and suppliers;

[0032] Module M2: Mining buyers’ commodity demands and supplier preferences from two types of text information: supply chain inquiry forms and quotation forms;

[0033] Module M3: The bipartite graph network structure used to describe the supply chain business relationship between buyers and suppliers is used to recall candidate recommended suppliers using the bipartite graph collaborative filtering algorithm, and then precise filtering is performed based on the buyer's product demand and supplier preference to screen out recommended suppliers.

[0034] Preferably, the bipartite graph network structure includes: a set of buyer nodes and a set of supplier nodes, and an edge connection is established when a buyer node and a supplier node have a supply chain business interaction relationship;

[0035] The bipartite graph network structure is expressed as G=(U,I,E), where U represents the buyer node set, I represents the supplier node set, and E represents the edge set; assuming that the number of buyer nodes |U|=m, and the number of supplier nodes |I|=n, the total number of bipartite graph network nodes is r=m+n; if the buyer u_p is associated with the supplier i_q, the value of G(p,q) is the number of associated orders; otherwise, G(p,q)=0.

[0036] Preferably, the module M2 adopts:

[0037] Module M2.1: Convert the inquiry form into the buyer's inquiry text embedding vector, convert the quotation form into the supplier's quotation text embedding vector, and calculate the vector similarity to obtain the supply product that is closest to the buyer's needs;

[0038] Module M2.2: Use the ALBERT model to semantically encode the text information of the technical terms in the inquiry to obtain a semantic vector, and then input the semantic vector as a feature into the TextCNN multi-label classification model. Use the TextCNN multi-label classification model to predict the probability distribution of buyers' preferences for supplier company qualifications, and finally realize intelligent mining of supplier preferences of different buyers;

[0039] The module M2.1 adopts:

[0040] Module M2.1.1: Preprocess the inquiry form and segment the preprocessed inquiry form to obtain a segmentation sequence;

[0041] Module M2.1.2: Count the TF-IDF scores of each word segment and obtain the word-text weight matrix after normalization.

[0042] Module M2.1.3: Use the trained Word2vec word embedding model to encode the word sequence to obtain word embedding;

[0043] Module M2.1.4: Obtain the buyer inquiry text embedding vector based on word embedding and word-text weight matrix through weighted summation.

[0044] Preferably, the module M3 adopts:

[0045] Module M3.1: Assume that the buyer set is P, the supplier set is S, and the number of suppliers to be recalled is k. Use the bipartite graph collaborative filtering algorithm to select the target buyer P. i The k suppliers with the most similar historical purchasing characteristics are selected as the candidate supplier set S_1 to be recommended;

[0046] Module M3.2: Filter out the suppliers in set S_1 to obtain a subset of suppliers S_2 whose historical quotation texts and the target buyer p's inquiry texts meet the preset requirements;

[0047] Module M3.3: Filter the suppliers in set S_2 to obtain a subset S_3 of suppliers whose qualification labels meet the buyer's preference for cooperative supplier qualifications;

[0048] Module M3.4: Set S_3 is used as the final result set of personalized supplier recommendations for buyer p;

[0049] The module M3.1 adopts:

[0050] Module M3.1.1: Initialize the matrix M, the historical preference matrix D, and the relative error tolerance ε;

[0051] Module M3.1.2: Let IR(i,j)=1 / n,(i=1,...,m,j=1,...,n);

[0052] Module M3.1.3: When ||αM·IR+(1-α)D-IR||≥ε is satisfied, IR=αM·IR+(1-α)D; Repeat triggering module M3.1.3 until ||αM·IR+(1-α)D-IR||≥ε is no longer satisfied, and the target purchaser P is obtained. i The k suppliers with the most similar historical purchasing characteristics are selected as the candidate supplier set S_1 to be recommended;

[0053] Among them, αM·IR represents the potential preference brought by network communication; (1-α)D represents the observed historical preference; α represents the parameter measuring the attenuation level of the network; D represents the buyer's historical preference matrix; M represents the number of two different buyers choosing the same supplier; IR represents the buyer's potential preference matrix.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. The present invention adopts a recommendation method that combines collaborative filtering and semantic matching of a bipartite graph network. On the one hand, the collaborative filtering of the bipartite graph network can make full use of the bipartite graph relationship between buyers and suppliers, use the bipartite graph to propagate user preferences, and mine the user's potential preferences, thereby improving the recall and novelty of the recommendation results; on the other hand, semantic matching is separated from the individual user category, which can solve the problems of sparse buyer behavior and cold start;

[0056] 2. The present invention achieves a complete description of supply chain procurement behavior by adopting a bipartite graph network structure;

[0057] 3. The present invention realizes the technical capability of recalling suppliers with similar supply chain behaviors by adopting the bipartite graph collaborative filtering algorithm;

[0058] 4. The present invention realizes the technical capability of mining buyers’ purchasing preferences by adopting a semantic matching model based on deep learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0060] Figure 1 Schematic diagram of the bipartite graph network structure applied to supply chain scenarios.

[0061] Figure 2 Schematic diagram of the demand and preference mining unit framework.

[0062] Figure 3 This is an example diagram of the word-text weight matrix.

[0063] Figure 4 Develop process for commodity demand.

[0064] Figure 5 Mining processes for supplier preferences.

[0065] Figure 6 This is a partial screenshot of the iterative calculation program for the buyer's potential preference matrix.

[0066] Figure 7 Flowchart of the personalized supplier recommendation process. DETAILED DESCRIPTION

[0067] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0068] Example 1

[0069] The present invention provides a supplier recommendation method based on supply chain network demand and preference mining, comprising:

[0070] Step S1: Using a bipartite graph network structure to describe the supply chain business relationship between buyers and suppliers;

[0071] Step S2: mining buyers’ commodity demands and supplier preferences from two types of text information, supply chain inquiry forms and quotation forms;

[0072] Step S3: The bipartite graph network structure used to describe the supply chain business relationship between buyers and suppliers is used to recall candidate recommended suppliers using a bipartite graph collaborative filtering algorithm, and then precise filtering is performed based on the buyer's product demand and supplier preference to select recommended suppliers.

[0073] Specifically, the step S1 adopts:

[0074] The ultimate goal of the present invention is to recommend suitable suppliers to buyers in the supply chain. Since the overlap between buyers and suppliers is low in the supply chain business data, a bipartite graph network structure is used to describe the supply chain business relationship between buyers and suppliers. In addition to depicting the network structure characteristics in the traditional sense, the bipartite graph network also helps to achieve accurate collaborative recommendations for users.

[0075] like Figure 1 As shown, the bipartite graph network consists of two types of node groups and the connections between the two types of nodes. The structural feature of the bipartite graph network is that there is no edge between nodes of the same type, and the two vertices to which each edge is attached belong to different types of nodes. In the present invention, the former type of node represents the buyer node, and the latter type of node represents the supplier node. When the buyer node and the supplier node have a supply chain business interaction relationship, an edge connection is established, so the buyer or supplier nodes with common connection nodes are indirectly connected. The connection status between the buyer domain and the supplier domain in the network reflects the tendency to establish a connection between the two, which can be regarded as the tendency of the buyer node to be mapped on the supplier node.

[0076] The bipartite graph network is represented by a graph structure as G = (U, I, E), where U is the buyer node set, I is the supplier node set, and E is the edge set. Assuming that the number of buyer nodes |U| = m, the number of supplier nodes |I| = n, the total number of nodes in the bipartite graph network is r = m + n. If buyer u_p is associated with supplier i_q, the value of G(p, q) is the number of associated orders; otherwise, G(p, q) = 0.

[0077] Specifically, Figure 2 As shown, the step S2 adopts:

[0078] Step S2.1: Convert the inquiry form into the buyer's inquiry text embedding vector, convert the quotation form into the supplier's quotation text embedding vector, and calculate the vector similarity to obtain the supply product that is closest to the buyer's needs;

[0079] Step S2.2: Use the ALBERT model to semantically encode the text information of the technical terms in the inquiry to obtain a semantic vector, and then input the semantic vector as a feature into the TextCNN multi-label classification model. Use the TextCNN multi-label classification model to predict the probability distribution of buyers' preferences for supplier company qualifications, and finally realize intelligent mining of supplier preferences of different buyers.

[0080] Among them, the step S2.1 adopts: the Word2Vec word embedding technology weighted by the TF-IDF (term frequency-inverse document frequency) statistical method is used to train the word embedding model in the supply chain field. The model training samples come from the fields such as "material name", "material category", "material brand", "technical attributes" and "technical terms" in the historical inquiry data of the company platform. First, based on the long-term accumulated industrial product field vocabulary, the jieba word segmentation tool is used to segment the text in the training sample. After word segmentation, the TF-IDF statistical method is used to calculate the importance of each word, and after normalization operation, the word-text weight matrix is ​​obtained, such as Figure 3 As shown in Figure 1, on the other hand, the Word2vec word embedding model is used to train the word segmentation sequence after word segmentation to obtain the word embedding model in the supply chain field. Finally, the supply chain text can be converted into a vector by combining the word weight matrix and the word embedding model; the semantic matching degree between the texts is obtained by calculating the similarity between the vectors (cosine similarity), as shown in Figure 1. Figure 4 shown.

[0081] More specifically, the step S2.1 adopts:

[0082] Step S2.1.1: preprocessing the inquiry form, and segmenting the preprocessed inquiry form to obtain a segmentation sequence;

[0083] Step S2.1.2: Count the TF-IDF scores of each word segment and obtain the word-text weight matrix after normalization;

[0084] Step S2.1.3: Use the trained Word2vec word embedding model to encode the word segmentation sequence to obtain word embedding;

[0085] Step S2.1.4: Perform weighted summation based on word embedding and word-text weight matrix to obtain the buyer inquiry text embedding vector.

[0086] Specifically, suppose W is a weight matrix composed of text I and word J. For a certain text i, its weight matrix is The word embedding of the jth word is e j , M is the total number of words contained in text i, defining the semantic information of text i, that is, the semantic embedding vector s i for:

[0087]

[0088] Among them, the step S2.2 adopts: an end-to-end training method consisting of ALBERT semantic encoding and TextCNN multi-label classification is selected to learn the buyer's qualification requirements for cooperative suppliers from the technical terms information in the inquiry form, thereby realizing the buyer's supplier selection preference prediction. In terms of training samples, the present invention adopts the supervised learning training sample format of "technical terms text-company qualification label". The forward process of the entire model can be described as: first, use the ALBERT model to semantically encode the technical terms text information to obtain a semantic vector, and then input the semantic vector as a feature into the TextCNN model, and use the TextCNN multi-label classification model to predict the probability distribution of the buyer's preference for the supplier company qualifications, and finally realize the intelligent mining of the supplier preferences of different buyers, such as Figure 5 shown.

[0089] The step S3 adopts: using a bipartite graph collaborative filtering algorithm to recall candidate recommended suppliers, and then using product demand matching and supplier preference matching technology to accurately screen out the final recommended supplier from the candidate recommended suppliers.

[0090] Among them, bipartite graph collaborative filtering is a recommendation algorithm that uses graph structure data. It can make full use of the established bipartite graph network of the supply chain between buyers and suppliers to make relevant recommendations. Bipartite graph collaborative filtering contains two key characteristics: propagation and attenuation. Its basic idea is to use bipartite graphs to propagate user preferences and use parameter α to measure the attenuation level of the network. Assume that the buyer's potential preference matrix is ​​IR_(m×n), D_(m×n) represents the buyer's historical preference matrix, where m and n represent the total number of buyers and suppliers respectively, IR(i,j) represents the potential preference value of buyer i for supplier j, D(i,j) represents the historical preference observation value, and the sum of each row of IR and D is 1; Assume that the matrix M_(m×m) is the buyer correlation matrix, and M(i,j) represents the number of buyers i and j who choose the same supplier. The core assumption of the bipartite graph collaborative filtering algorithm is that the potential preference of user i is equal to the weighted average of its historical preference and the potential preference propagation value obtained based on its neighbors, that is:

[0091] IR=aM·IR+(1-α)D

[0092] Among them, αM·IR measures the potential preference brought by network communication, (1-α)D represents the observed historical preference, and the formula can be solved by iteration. The solution process is as follows: Figure 6 shown.

[0093] The overall operation process recommended for suppliers is as follows Figure 7 As shown. Assuming the target buyer is p and the supplier set is S, first use the above-mentioned bipartite graph collaborative filtering method to screen out suppliers with the most similar historical purchasing characteristics to buyer p as the candidate supplier set S_1 to be recommended; the next step is to screen out the supplier subset S_2 in the set S_1 whose historical quotation text is more similar to the supplier p's inquiry text; further, screen out the supplier subset S_3 in the set S_2 whose qualification labels meet the buyer's preference for the qualification of cooperative suppliers; finally, set S_3 is used as the final result set of personalized supplier recommendation for buyer p.

[0094] The present invention provides a supplier recommendation system based on supply chain network demand and preference mining, comprising:

[0095] Module M1: Use a bipartite graph network structure to describe the supply chain business relationship between buyers and suppliers;

[0096] Module M2: Mining buyers’ commodity demands and supplier preferences from two types of text information: supply chain inquiry forms and quotation forms;

[0097] Module M3: The bipartite graph network structure used to describe the supply chain business relationship between buyers and suppliers is used to recall candidate recommended suppliers using the bipartite graph collaborative filtering algorithm, and then precise filtering is performed based on the buyer's product demand and supplier preference to screen out recommended suppliers.

[0098] Specifically, the module M1 adopts:

[0099] The ultimate goal of the present invention is to recommend suitable suppliers to buyers in the supply chain. Since the overlap between buyers and suppliers is low in the supply chain business data, a bipartite graph network structure is used to describe the supply chain business relationship between buyers and suppliers. In addition to depicting the network structure characteristics in the traditional sense, the bipartite graph network also helps to achieve accurate collaborative recommendations for users.

[0100] like Figure 1 As shown, the bipartite graph network consists of two types of node groups and the connections between the two types of nodes. The structural feature of the bipartite graph network is that there is no edge between nodes of the same type, and the two vertices to which each edge is attached belong to different types of nodes. In the present invention, the former type of node represents the buyer node, and the latter type of node represents the supplier node. When the buyer node and the supplier node have a supply chain business interaction relationship, an edge connection is established, so the buyer or supplier nodes with common connection nodes are indirectly connected. The connection status between the buyer domain and the supplier domain in the network reflects the tendency to establish a connection between the two, which can be regarded as the tendency of the buyer node to be mapped on the supplier node.

[0101] The bipartite graph network is represented by a graph structure as G = (U, I, E), where U is the buyer node set, I is the supplier node set, and E is the edge set. Assuming that the number of buyer nodes |U| = m, the number of supplier nodes |I| = n, the total number of nodes in the bipartite graph network is r = m + n. If buyer u_p is associated with supplier i_q, the value of G(p, q) is the number of associated orders; otherwise, G(p, q) = 0.

[0102] Specifically, Figure 2 As shown, the module M2 adopts:

[0103] Module M2.1: Convert the inquiry form into the buyer's inquiry text embedding vector, convert the quotation form into the supplier's quotation text embedding vector, and calculate the vector similarity to obtain the supply product that is closest to the buyer's needs;

[0104] Module M2.2: Use the ALBERT model to semantically encode the text information of the technical terms in the inquiry to obtain a semantic vector, and then input the semantic vector as a feature into the TextCNN multi-label classification model. Use the TextCNN multi-label classification model to predict the probability distribution of buyers' preferences for supplier company qualifications, and finally realize intelligent mining of supplier preferences of different buyers.

[0105] Among them, the module M2.1 adopts: the Word2Vec word embedding technology weighted by the TF-IDF (term frequency-inverse document frequency) statistical method is used to train the word embedding model in the supply chain field. The model training samples come from the fields such as "material name", "material category", "material brand", "technical attributes" and "technical terms" in the historical inquiry data of the company platform. First, based on the long-term accumulated industrial product field vocabulary, the jieba word segmentation tool is used to segment the text in the training sample. After word segmentation, the TF-IDF statistical method is used to calculate the importance of each word, and after normalization, the word-text weight matrix is ​​obtained, such as Figure 3 As shown in Figure 1, on the other hand, the Word2vec word embedding model is used to train the word segmentation sequence after word segmentation to obtain the word embedding model in the supply chain field. Finally, the supply chain text can be converted into a vector by combining the word weight matrix and the word embedding model; the semantic matching degree between the texts is obtained by calculating the similarity between the vectors (cosine similarity), as shown in Figure 1. Figure 4 shown.

[0106] More specifically, the module M2.1 adopts:

[0107] Module M2.1.1: Preprocess the inquiry form and segment the preprocessed inquiry form to obtain a segmentation sequence;

[0108] Module M2.1.2: Count the TF-IDF scores of each word segment and obtain the word-text weight matrix after normalization.

[0109] Module M2.1.3: Use the trained Word2vec word embedding model to encode the word sequence to obtain word embedding;

[0110] Module M2.1.4: Obtain the buyer inquiry text embedding vector based on word embedding and word-text weight matrix through weighted summation.

[0111] Specifically, suppose W is a weight matrix composed of text I and word J. For a certain text i, its weight matrix is The word embedding of the jth word is e j , M is the total number of words contained in text i, defining the semantic information of text i, that is, the semantic embedding vector s i for:

[0112]

[0113] Among them, the module M2.2 adopts: an end-to-end training method consisting of ALBERT semantic encoding and TextCNN multi-label classification is selected to learn the buyer's qualification requirements for cooperative suppliers from the technical terms information in the inquiry form, thereby realizing the buyer's supplier selection preference prediction. In terms of training samples, the present invention adopts the supervised learning training sample format of "technical terms text-company qualification label". The forward process of the entire model can be described as follows: first, the ALBERT model is used to semantically encode the technical terms text information to obtain a semantic vector, and then the semantic vector is input as a feature into the TextCNN model, and the TextCNN multi-label classification model is used to predict the probability distribution of the buyer's preference for the supplier company qualifications, and finally the intelligent mining of the supplier preferences of different buyers is realized, such as Figure 5 shown.

[0114] The module M3 adopts: using a bipartite graph collaborative filtering algorithm to recall candidate recommended suppliers, and then using product demand matching and supplier preference matching technology to accurately screen out the final recommended suppliers from the candidate recommended suppliers.

[0115] Among them, bipartite graph collaborative filtering is a recommendation algorithm that uses graph structure data. It can make full use of the established bipartite graph network of the supply chain between buyers and suppliers to make relevant recommendations. Bipartite graph collaborative filtering contains two key characteristics: propagation and attenuation. Its basic idea is to use bipartite graphs to propagate user preferences and use parameter α to measure the attenuation level of the network. Assume that the buyer's potential preference matrix is ​​IR_(m×n), D_(m×n) represents the buyer's historical preference matrix, where m and n represent the total number of buyers and suppliers respectively, IR(i,j) represents the potential preference value of buyer i for supplier j, D(i,j) represents the historical preference observation value, and the sum of each row of IR and D is 1; Assume that the matrix M_(m×m) is the buyer correlation matrix, and M(i,j) represents the number of buyers i and j who choose the same supplier. The core assumption of the bipartite graph collaborative filtering algorithm is that the potential preference of user i is equal to the weighted average of its historical preference and the potential preference propagation value obtained based on its neighbors, that is:

[0116] IR=αM·IR+(1-α)D

[0117] Among them, αM·IR measures the potential preference brought by network communication, (1-α)D represents the observed historical preference, and the formula can be solved by iteration. The solution process is as follows: Figure 6 shown.

[0118] The overall operation process recommended for suppliers is as follows Figure 7As shown. Assuming the target buyer is p and the supplier set is S, first use the above-mentioned bipartite graph collaborative filtering method to screen out suppliers with the most similar historical purchasing characteristics to buyer p as the candidate supplier set S_1 to be recommended; the next step is to screen out the supplier subset S_2 in the set S_1 whose historical quotation text is more similar to the supplier p's inquiry text; further, screen out the supplier subset S_3 in the set S_2 whose qualification labels meet the buyer's preference for the qualification of cooperative suppliers; finally, set S_3 is used as the final result set of personalized supplier recommendation for buyer p.

[0119] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.

[0120] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A supplier recommendation method based on supply chain network demand and preference mining, characterized in that: include: Step S1: Use a bipartite graph network structure to describe the supply chain business relationship between buyers and suppliers; Step S2: mining buyers’ commodity demands and supplier preferences from two types of text information, supply chain inquiry forms and quotation forms; Step S3: Using the bipartite graph network structure used to describe the supply chain business relationship between buyers and suppliers, recall candidate recommended suppliers using the bipartite graph collaborative filtering algorithm, and then perform precise filtering based on the buyer's commodity demand and supplier preference to select recommended suppliers; The bipartite graph network structure consists of two types of node groups and connections between the two types of nodes; The bipartite graph network structure includes: a set of buyer nodes and a set of supplier nodes. When a buyer node and a supplier node have a supply chain business interaction relationship, an edge connection is established. Therefore, buyer nodes or supplier nodes with common connection nodes are indirectly connected. The connection status between the buyer domain and the supplier domain in the bipartite graph network reflects the tendency of establishing a connection between the two, which can be regarded as the tendency of the buyer node to be mapped on the supplier node. The bipartite graph network structure is represented by G=(U,I,E), where U represents the buyer node set, I represents the supplier node set, and E represents the edge set; assuming that the number of buyer nodes |U|=m, and the number of supplier nodes |I|=n, the total number of bipartite graph network nodes is r=m+n; if the buyer u_p is associated with the supplier i_q, the value of G(p,q) is the number of associated orders; otherwise, G(p,q)=0; The step S2 adopts: Step S2.1: Convert the inquiry form into the buyer's inquiry text embedding vector, convert the quotation form into the supplier's quotation text embedding vector, and calculate the vector similarity to obtain the supply product that is closest to the buyer's needs; Step S2.2: Use the ALBERT model to semantically encode the text information of the technical terms in the inquiry to obtain a semantic vector, and then input the semantic vector as a feature into the TextCNN multi-label classification model. Use the TextCNN multi-label classification model to predict the probability distribution of buyers' preferences for supplier company qualifications, and finally realize intelligent mining of supplier preferences of different buyers; The step S2.1 adopts: using the TF-IDF statistical method weighted Word2Vec word embedding technology to train the word embedding model in the supply chain field; the word embedding model training samples come from the company platform's historical inquiry data, including material name, material category, material brand, technical attributes and technical terms fields; firstly, based on the long-term accumulated industrial product field vocabulary, the jieba word segmentation tool is used to segment the text in the training sample; after segmentation, the TF-IDF statistical method is used to calculate the importance of each word, and the word-text weight matrix is ​​obtained after normalization operation, and the Word2vec word embedding model is used to train the segmented word sequence after segmentation to obtain the supply chain field word embedding model; finally, the supply chain text can be converted into a vector by combining the word weight matrix and the word embedding model; the semantic matching degree between the texts is obtained by calculating the similarity between the vectors; More specifically, the step S2.1 adopts: Step S2.1.1: preprocessing the inquiry form, and segmenting the preprocessed inquiry form to obtain a segmentation sequence; Step S2.1.2: Count the TF-IDF scores of each word segment and obtain the word-text weight matrix after normalization; Step S2.1.3: Use the trained Word2vec word embedding model to encode the word sequence to obtain word embedding; Step S2.1.4: Obtain the buyer inquiry text embedding vector by performing weighted summation based on word embedding and word-text weight matrix; Assumptions By text and words The weight matrix formed by Its weight matrix is , among which The word embedding of a word is , For Text The total number of words contained in the definition text The amount of semantic information, that is, the semantic embedding vector for: Step S2.2 adopts: an end-to-end training method consisting of ALBERT semantic encoding and TextCNN multi-label classification is used to learn the buyer's qualification requirements for cooperative suppliers from the technical terms information of the inquiry form, so as to realize the buyer's supplier selection preference prediction; in terms of training samples, the supervised learning training sample format of "technical terms text-company qualification label" is adopted; the forward process of the entire model can be described as follows: first, the ALBERT model is used to semantically encode the technical terms text information to obtain a semantic vector, and then the semantic vector is input as a feature into the TextCNN model, and the TextCNN multi-label classification model is used to predict the probability distribution of the buyer's preference for the supplier company qualifications, and finally the supplier preferences of different buyers are intelligently mined; The step S3 adopts: Bipartite graph collaborative filtering is a recommendation algorithm that uses graph structured data and makes recommendations based on the established buyer-supplier supply chain bipartite graph network. Bipartite graph collaborative filtering contains two key features: propagation and attenuation. It uses bipartite graphs to propagate user preferences and uses parameter α to measure the attenuation level of the network. Assume that the buyer's potential preference matrix is ​​IR_(m×n), D_(m×n) represents the buyer's historical preference matrix, where m and n represent the total number of buyers and suppliers, respectively, IR(i,j) represents the potential preference value of buyer i for supplier j, D(i,j) represents the historical preference observation value, and the sum of each row of IR and D is 1. Assume that the matrix M_(m×m) is the buyer correlation matrix, and M(i,j) represents the number of buyers i and j who choose the same supplier. The core assumption of the bipartite graph collaborative filtering algorithm is that the potential preference of user i is equal to the weighted average of its historical preference and the potential preference propagation value obtained based on its neighbors, that is, the formula: Among them, αM·IR measures the potential preference brought by network communication, (1-α)D represents the observed historical preference, and the formula is solved by iteration; Step S3.1: Assume that the buyer set is P, the supplier set is S, and the number of suppliers to be recalled is k. Use the bipartite graph collaborative filtering algorithm to filter out suppliers that are related to the target buyer. The k suppliers with the most similar historical purchasing characteristics are selected as the candidate supplier set S_1 to be recommended; Step S3.2: Filter out the suppliers in the set S_1 to obtain a supplier subset S_2 whose historical quotation texts and the target buyer p's inquiry texts meet the preset requirements; Step S3.3: Filter the suppliers in set S_2 to find a subset S_3 of suppliers whose qualification labels meet the buyer's preference for the qualifications of cooperative suppliers; Step S3.4: Set S_3 is used as the final result set of personalized supplier recommendations for buyer p; The step S3.1 adopts: Step S3.1.1: Initialize the matrix M, the historical preference matrix D, and the relative error tolerance ; Step S3.1.2: Let ; Step S3.1.3: When the When ; Repeat triggering step S3.1.3 until it is not satisfied , obtain the target buyers The k suppliers with the most similar historical purchasing characteristics are selected as the candidate supplier set S_1 to be recommended; in, Represents the potential preference brought by network communication; represents the observed historical preferences; represents the parameter for measuring the attenuation level of the network; D represents the buyer's historical preference matrix; M represents the number of times two different buyers choose the same supplier; Represents the potential preference matrix of buyers.

2. A supplier recommendation system based on supply chain network demand and preference mining, characterized by: include: Module M1: Use a bipartite graph network structure to describe the supply chain business relationship between buyers and suppliers; Module M2: Mining buyers’ commodity demands and supplier preferences from two types of text information: supply chain inquiry forms and quotation forms; Module M3: The bipartite graph network structure used to describe the supply chain business relationship between buyers and suppliers is used to recall candidate recommended suppliers using the bipartite graph collaborative filtering algorithm, and then precise filtering is performed based on the buyer's commodity demand and supplier preference to select the recommended suppliers; The bipartite graph network structure consists of two types of node groups and connections between the two types of nodes; The bipartite graph network structure includes: a set of buyer nodes and a set of supplier nodes. When a buyer node and a supplier node have a supply chain business interaction relationship, an edge connection is established. Therefore, buyer nodes or supplier nodes with common connection nodes are indirectly connected. The connection status between the buyer domain and the supplier domain in the bipartite graph network reflects the tendency of establishing a connection between the two, which can be regarded as the tendency of the buyer node to be mapped on the supplier node. The bipartite graph network structure is represented by G=(U,I,E), where U represents the buyer node set, I represents the supplier node set, and E represents the edge set; assuming that the number of buyer nodes |U|=m, and the number of supplier nodes |I|=n, the total number of bipartite graph network nodes is r=m+n; if the buyer u_p is associated with the supplier i_q, the value of G(p,q) is the number of associated orders; otherwise, G(p,q)=0; The module M2 adopts: Module M2.1: Convert the inquiry form into the buyer's inquiry text embedding vector, convert the quotation form into the supplier's quotation text embedding vector, and calculate the vector similarity to obtain the supply product that is closest to the buyer's needs; Module M2.2: Use the ALBERT model to semantically encode the text information of the technical terms in the inquiry to obtain a semantic vector, and then input the semantic vector as a feature into the TextCNN multi-label classification model. Use the TextCNN multi-label classification model to predict the probability distribution of buyers' preferences for supplier company qualifications, and finally realize intelligent mining of supplier preferences of different buyers; The module M2.1 adopts: the TF-IDF statistical method weighted Word2Vec word embedding technology is used to train the word embedding model in the supply chain field; the word embedding model training samples are derived from the company platform's historical inquiry data, including material name, material category, material brand, technical attributes and technical terms fields; firstly, based on the long-term accumulated industrial product field vocabulary, the jieba word segmentation tool is used to segment the text in the training sample; after segmentation, the TF-IDF statistical method is used to calculate the importance of each word, and the word-text weight matrix is ​​obtained after normalization operation; on the other hand, the Word2vec word embedding model is used to train the segmented word sequence after segmentation to obtain the supply chain field word embedding model; finally, the supply chain text can be converted into a vector by combining the word weight matrix and the word embedding model; the semantic matching degree between the texts is obtained by calculating the similarity between the vectors; More specifically, the module M2.1 adopts: Module M2.1.1: Preprocess the inquiry form and segment the preprocessed inquiry form to obtain a segmentation sequence; Module M2.1.2: Count the TF-IDF scores of each word segment and obtain the word-text weight matrix after normalization. Module M2.1.3: Use the trained Word2vec word embedding model to encode the word sequence to obtain word embedding; Module M2.1.4: Obtain the buyer's inquiry text embedding vector by weighted summation based on word embedding and word-text weight matrix; Assumptions By text and words The weight matrix formed by Its weight matrix is , among which The word embedding of a word is , For Text The total number of words contained in the definition text The amount of semantic information, that is, the semantic embedding vector for: Module M2.2 adopts: an end-to-end training method consisting of ALBERT semantic encoding and TextCNN multi-label classification is used to learn the buyer's qualification requirements for cooperative suppliers from the technical terms information of the inquiry form, so as to realize the buyer's supplier selection preference prediction; in terms of training samples, the supervised learning training sample format of "technical terms text-company qualification label" is adopted; the forward process of the entire model can be described as follows: first, the ALBERT model is used to semantically encode the technical terms text information to obtain a semantic vector, and then the semantic vector is input as a feature into the TextCNN model, and the TextCNN multi-label classification model is used to predict the probability distribution of the buyer's preference for the supplier company qualifications, and finally the supplier preferences of different buyers are intelligently mined; The module M3 adopts: Bipartite graph collaborative filtering is a recommendation algorithm that uses graph structured data and makes recommendations based on the established buyer-supplier supply chain bipartite graph network. Bipartite graph collaborative filtering contains two key features: propagation and attenuation. It uses bipartite graphs to propagate user preferences and uses parameter α to measure the attenuation level of the network. Assume that the buyer's potential preference matrix is ​​IR_(m×n), D_(m×n) represents the buyer's historical preference matrix, where m and n represent the total number of buyers and suppliers, respectively, IR(i,j) represents the potential preference value of buyer i for supplier j, D(i,j) represents the historical preference observation value, and the sum of each row of IR and D is 1. Assume that the matrix M_(m×m) is the buyer correlation matrix, and M(i,j) represents the number of buyers i and j who choose the same supplier. The core assumption of the bipartite graph collaborative filtering algorithm is that the potential preference of user i is equal to the weighted average of its historical preference and the potential preference propagation value obtained based on its neighbors, that is, the formula: Among them, αM·IR measures the potential preference brought by network communication, (1-α)D represents the observed historical preference, and the formula is solved by iteration; Module M3.1: Assuming the complete set of buyers is P, the complete set of suppliers is S, and the number of suppliers to be recalled is k, use the bipartite graph collaborative filtering algorithm to screen out suppliers with target buyers. The k suppliers with the most similar historical purchasing characteristics are selected as the candidate supplier set S_1 to be recommended; Module M3.2: Filter out the suppliers in set S_1 to obtain a subset of suppliers S_2 whose historical quotation texts and the target buyer p's inquiry texts meet the preset requirements; Module M3.3: Filter the suppliers in set S_2 to obtain a subset S_3 of suppliers whose qualification labels meet the buyer's preference for cooperative supplier qualifications; Module M3.4: Set S_3 is used as the final result set of personalized supplier recommendations for buyer p; The module M3.1 adopts: Module M3.1.1: Initialization matrix M, historical preference matrix D, and relative error tolerance ; Module M3.1.2: Order ; Module M3.1.3: When satisfied When ; Repeat triggering module M3.1.3 until it is not satisfied , obtain the target buyers The k suppliers with the most similar historical purchasing characteristics are selected as the candidate supplier set S_1 to be recommended; in, Represents the potential preference brought by network communication; represents the observed historical preferences; represents the parameter for measuring the attenuation level of the network; D represents the buyer's historical preference matrix; M represents the number of times two different buyers choose the same supplier; Represents the potential preference matrix of buyers.

Images