A policy-enterprise matching method and device based on large model

Abstract

The present invention discloses a policy-enterprise matching method and device based on a large model, and relates to the technical field of large models; the policy-enterprise matching method based on a large model includes: obtaining a policy training set; obtaining an enterprise training set; constructing policy and enterprise bipartite graph information according to the policy training set and the enterprise training set; training a recommendation model through the enterprise bipartite graph information to obtain a trained recommendation model; inputting the information to be matched into the trained recommendation model to obtain a list of recommendation results. The policy-enterprise matching method based on a large model of the present application realizes matching between policy texts by means of graph contrast learning. Contrastive learning can help the model understand the complex relationship between policies and enterprises and capture the underlying semantic information. The model learns the representation of nodes by comparing different views. By forcing the model to pay attention to important nodes and edges, contrastive learning can improve the model's ability to understand the graph structure.

Description

Technical Field

[0001] The invention discloses a policy-enterprise matching method based on a big model and a policy-enterprise matching device based on a big model, and relates to the technical field of big data. Background Art

[0002] Traditional similarity or matching metrics are usually based on our own specified calculation rules, such as Euclidean distance, cosine similarity and other methods based on certain features. These methods are usually linear or simple rules, which calculate similarity scores or matching scores, and then make recommendations based on the scores. The advantage of this method is that it is simple and direct, easy to understand and implement. Through clearly defined rules and measurement methods, recommendations can be made based on the scores obtained, such as selecting the policy with the highest similarity or the highest match to recommend to the enterprise. However, this method may not capture more complex nonlinear relationships and potential patterns in the data. They may ignore implicit information and semantic associations in the data, resulting in limited accuracy of the recommendation results; the amount of data for some policies is small, resulting in these high-quality policies not being recommended to more companies. Summary of the invention

[0003] In view of the problems of the prior art, the present invention provides a policy-enterprise matching method based on a large model, and the technical solution adopted is:

[0004] A policy-enterprise matching method based on a large model, the policy-enterprise matching method based on a large model comprising:

[0005] Get the policy training set;

[0006] Get the enterprise training set;

[0007] Constructing policy and enterprise bipartite graph information according to the policy training set and the enterprise training set;

[0008] The recommendation model is trained through the enterprise bipartite graph information to obtain a trained recommendation model;

[0009] The information to be matched is input into the trained recommendation model to obtain a list of recommendation results.

[0010] Optionally, acquiring a policy training set includes:

[0011] Get policy text data;

[0012] Through word embedding feature extraction technology, the policy text data is converted into vector representation, thereby obtaining a policy text vector training set;

[0013] The obtaining of the enterprise training set comprises:

[0014] Get enterprise text data;

[0015] Through word embedding feature extraction technology, enterprise text data is converted into vector representation to obtain the enterprise text vector training set.

[0016] Optionally, constructing a bipartite graph of policies and enterprises includes:

[0017] Get policy and enterprise bipartite graph node information ,in, For enterprise set, is the policy set, is the edge between sets;

[0018] Obtain two row vectors based on the policy and the enterprise bipartite graph and The similarity relationship between the two entities is used to construct the edge information in the bipartite graph of the enterprise.

[0019] Optionally, the recommendation model is a graph contrast learning model.

[0020] Optionally, the step of training the recommendation model by using the enterprise bipartite graph information to obtain a trained recommendation model includes:

[0021] Performing data enhancement twice on the enterprise bipartite graph information to obtain a first representation graph and a second representation graph corresponding to the enterprise bipartite graph information;

[0022] encoding the first representation graph using the recommendation model to obtain a first representation vector of the first representation graph;

[0023] encoding the second representation graph using the recommendation model to obtain a second representation vector of the second representation graph;

[0024] Calculating contrast loss using a plurality of samples consisting of the first representation vector and the second representation vector;

[0025] The parameters of the recommended model are updated according to the contrast loss of the recommended model and the training is repeated until the contrast loss meets the preset convergence condition, thereby obtaining a trained recommendation model.

[0026] Optionally, the first representation graph is obtained by edge loss and edge replacement, wherein a degree centrality-based method is used to assist in determining edge loss in an adjacency matrix, and the expression of the first representation graph is as follows:

[0027] ;

[0028] The second representation graph is obtained by edge loss and edge replacement, wherein a method based on degree centrality is used to assist in determining edge loss in the adjacency matrix. The expression of the second representation graph is as follows:

[0029] ;in,

[0030] and are the two augmented matrices generated, It is a mask matrix obtained based on the degree centrality method; is a permutation matrix and T represents the transpose operation.

[0031] Optionally, the contrast loss expression is as follows:

[0032] ;in,

[0033] For comparison loss, Compare losses for users, is the policy comparison loss.

[0034] Optionally, the expression of the user contrast loss is as follows:

[0035] ;in,

[0036] is the user comparison loss, sim is the similarity function, Enhanced view of the same user node, is the temperature hyperparameter.

[0037] The present application also provides a policy-enterprise matching device based on a large model, and the policy-enterprise matching device based on a large model includes:

[0038] A policy training set acquisition module, wherein the policy training set acquisition module is used to acquire a policy training set;

[0039] An enterprise training set acquisition module, wherein the enterprise training set acquisition module is used to acquire an enterprise training set;

[0040] A policy and enterprise bipartite graph information construction module, wherein the policy and enterprise bipartite graph information construction module is used to construct policy and enterprise bipartite graph information according to the policy training set and the enterprise training set;

[0041] A training module, wherein the training module is used to train the recommendation model through the enterprise bipartite graph information, thereby obtaining a trained recommendation model;

[0042] The recommendation result list acquisition module is used to input the information to be matched into the trained recommendation model to obtain the recommendation result list.

[0043] The large-model-based policy-enterprise matching method of this application achieves matching between policy texts through graph contrast learning. Contrastive learning can help the model understand the complex relationship between policies and enterprises and capture the underlying semantic information. The model learns the representation of nodes by comparing different views. By forcing the model to focus on important nodes and edges, contrastive learning can improve the model's ability to understand the graph structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0045] Figure 1 It is a flowchart of a policy-enterprise matching method based on a large model provided in one embodiment of the present invention.

[0046] Figure 2 It is a schematic diagram of data enhancement provided by an embodiment of the present invention.

[0047] Figure 3 It is a detailed flowchart of a policy-enterprise matching method based on a large model provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

[0049] like Figure 1 The policy-enterprise matching method based on the big model shown includes:

[0050] Step 1: Get the policy training set;

[0051] Step 2: Obtain enterprise training set;

[0052] Step 3: constructing policy and enterprise bipartite graph information according to the policy training set and the enterprise training set;

[0053] Step 4: Train the recommendation model using the enterprise bipartite graph information to obtain a trained recommendation model;

[0054] Step 5: Input the information to be matched into the trained recommendation model to obtain a list of recommendation results.

[0055] The large-model-based policy-enterprise matching method of this application achieves matching between policy texts through graph contrast learning. Contrastive learning can help the model understand the complex relationship between policies and enterprises and capture the underlying semantic information. The model learns the representation of nodes by comparing different views. By forcing the model to focus on important nodes and edges, contrastive learning can improve the model's ability to understand the graph structure.

[0056] In this embodiment, obtaining a policy training set includes:

[0057] Get policy text data;

[0058] Through word embedding feature extraction technology, the policy text data is converted into vector representation, thereby obtaining a policy text vector training set;

[0059] The obtaining of the enterprise training set comprises:

[0060] Get enterprise text data;

[0061] Through word embedding feature extraction technology, enterprise text data is converted into vector representation to obtain the enterprise text vector training set.

[0062] Specifically, policy information and enterprise information are both text information, among which policy text includes text data sets of applicable industries, laws and regulations, issuing agencies, etc.

[0063] Enterprise information includes enterprise name, number of employees, total profit, etc. Word embedding feature extraction technology is used to convert policy text and enterprise information into vector representation.

[0064] In this embodiment, the preprocessed text data is used as input. The corpus is a collection of all text data, which is used to train the model. The Word2Vec model is trained using the corpus, and the continuous bag of words model is used to predict the target word from the context. The word vector representation is obtained: after the model training is completed, the trained Word2Vec model can be used to obtain the word vector representation of the word. For each word, the corresponding vector can be obtained by querying the model.

[0065] In this embodiment, the recommendation model is a graph comparison learning model.

[0066] In this embodiment, constructing a bipartite graph of policies and enterprises includes:

[0067] Get policy and enterprise bipartite graph node information ,in, For enterprise set, is the policy set, is the edge between sets;

[0068] Obtain two row vectors based on the policy and the enterprise bipartite graph and The similarity relationship between the two entities is used to construct the edge information in the bipartite graph of the enterprise.

[0069] Specifically, the recommendation model of the graph convolutional network (GCN) can integrate the node information and topological structure of the policy-enterprise interaction graph. It can be modeled as a policy-enterprise bipartite graph ,in For enterprise set, is the policy set, is the edge between sets. It can be expressed as the policy-enterprise interaction matrix . Then the adjacency matrix is:

[0070] ;

[0071] The interaction matrix Scoring the cosine similarity between policies and enterprises. Since the word embedding model learns from a large amount of text data during the training process, the length of the word vector in the vector space is relatively consistent. We will get two row vectors and ( The weighted cosine similarity function (where is the relative length in the vector space) is defined as:

[0072] ;

[0073] in, Each dimension The weight of the weighted cosine similarity is a method for calculating the similarity between two vectors. Reflects each dimension Importance when calculating similarity.

[0074] In this embodiment, the step of training the recommendation model through the enterprise bipartite graph information to obtain the trained recommendation model includes:

[0075] Performing data enhancement twice on the enterprise bipartite graph information to obtain a first representation graph and a second representation graph corresponding to the enterprise bipartite graph information;

[0076] encoding the first representation graph using the recommendation model to obtain a first representation vector of the first representation graph;

[0077] encoding the second representation graph using the recommendation model to obtain a second representation vector of the second representation graph;

[0078] Calculating contrast loss using multiple samples composed of the first representation vector and the second representation vector; wherein enhanced views of the same user node are regarded as positive pairs, and views of different user nodes are regarded as negative pairs;

[0079] The parameters of the recommended model are updated according to the contrast loss of the recommended model and the training is repeated until the contrast loss meets the preset convergence condition, thereby obtaining a trained recommendation model.

[0080] Specifically, in our graph contrastive learning model, at each iteration, we perform two random augmentation functions and Sampling is performed, where is the set of all possible enhancement functions for the adjacency matrix Then, we generate two chart views (the first representation and the second representation), represented as and .

[0081] In this embodiment, considering that randomly discarding edges will make the graph structure more unstable, this randomness will lead to poor generalization ability. We use a combined enhancement scheme, using edge loss and edge replacement, and a method based on degree centrality to assist in determining edge loss in the adjacency matrix. Degree centrality is an indicator used to measure the number of connections between nodes in a graph. The higher the degree centrality of a node, the more connections it has with other nodes. Its expression is as follows:

[0082] ;in,

[0083] and are the two augmented matrices generated, Perform edge loss and edge replacement on the graph structure respectively. It is a mask matrix obtained based on the degree centrality method; is a permutation matrix, T represents the transpose operation, each element Indicates that in the permuted adjacency matrix, the original adjacency matrix The index of the element at row i and column j in the new position.

[0084] In this embodiment, the contrast loss expression is as follows:

[0085] ;in,

[0086] For comparison loss, Compare losses for users, is the policy comparison loss.

[0087] In this embodiment, the expression of the user contrast loss is as follows:

[0088] ;in,

[0089] is the user comparison loss, sim is the similarity function, Enhanced view of the same user node, is the temperature hyperparameter.

[0090] In this embodiment, we use a discriminator to compare two augmented samples and distinguish the embeddings of the same node in these two different views from other node embeddings. After data augmentation, the augmented views of the same user node are considered positive pairs, i.e. , views of different user nodes are regarded as negative pairs. A contrastive loss (InfoNCE) is adopted to maximize the consistency of positive pairs and minimize the consistency of negative pairs.

[0091] Similarly, the policy also uses contrastive loss The final contrast loss is a combination of the two losses. Contrastive learning in recommender systems usually adopts a joint learning strategy to train its model, rather than a pre-training and fine-tuning strategy. That is, the pre-training task and the downstream task are jointly optimized.

[0092] The method of this application alleviates the model robustness and data sparsity problems in policy-enterprise recommendation. Policy-enterprise matching can make recommendations based on specific policy goals and enterprise needs. It can more accurately recommend policies that meet the needs of enterprises. It can avoid recommending irrelevant or inappropriate policies to enterprises, and improve the effectiveness and practicality of recommendations.

[0093] In this application, contrastive learning can help the model understand the complex relationship between policies and enterprises and capture the underlying semantic information. The model learns the representation of nodes by comparing different views.

[0094] See also Figure 2 , Adaptive combination enhancement method: By performing edge loss and edge permutation on two graph topologies respectively, different views are generated to enrich the diversity of data. The method based on degree centrality can be used to assist in determining edge loss in the adjacency matrix. The method based on random loss may destroy the graph structure. Degree centrality is an indicator used to measure the number of connections between a node in a graph. The higher the degree centrality of a node, the more connections it has with other nodes.

[0095] The present application also provides a policy-enterprise matching device based on a large model, which includes a policy training set acquisition module, an enterprise training set acquisition module, a policy and enterprise bipartite graph information construction module, a training module, and a recommendation result list acquisition module, wherein:

[0096] The policy training set acquisition module is used to acquire the policy training set;

[0097] The enterprise training set acquisition module is used to acquire the enterprise training set;

[0098] The policy and enterprise bipartite graph information construction module is used to construct the policy and enterprise bipartite graph information according to the policy training set and the enterprise training set;

[0099] The training module is used to train the recommendation model through the enterprise bipartite graph information, so as to obtain the trained recommendation model;

[0100] The recommendation result list acquisition module is used to input the information to be matched into the trained recommendation model to obtain the recommendation result list.

[0101] Although the embodiments disclosed in the present invention are as above, the above contents are only embodiments adopted for facilitating the understanding of the present invention and are not intended to limit the present invention. Any technician in the technical field to which the present invention belongs can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in the present invention, but the patent protection scope of the present invention shall still be subject to the scope defined in the attached claims.

Claims

1. A policy-enterprise matching method based on a large model, characterized in that: The policy-enterprise matching method based on the big model includes: Get the policy training set; Get the enterprise training set; Constructing policy and enterprise bipartite graph information according to the policy training set and the enterprise training set; The recommendation model is trained through the enterprise bipartite graph information to obtain a trained recommendation model; Input the information to be matched into the trained recommendation model to obtain a list of recommendation results; The acquisition policy training set includes: Get policy text data; Through word embedding feature extraction technology, the policy text data is converted into vector representation, thereby obtaining a policy text vector training set; The obtaining of the enterprise training set comprises: Get enterprise text data; Through word embedding feature extraction technology, the enterprise text data is converted into vector representation, thereby obtaining the enterprise text vector training set; The construction of the policy and enterprise bipartite graph includes: Get policy and enterprise bipartite graph node information ,in, For enterprise set, is the policy set, is the edge between sets; Obtain two row vectors based on the policy and the enterprise bipartite graph and Similarity relations are used to construct edge information in the bipartite graph of enterprises; The recommendation model is a graph contrast learning model; The step of training the recommendation model through the enterprise bipartite graph information to obtain the trained recommendation model includes: Performing data enhancement twice on the enterprise bipartite graph information to obtain a first representation graph and a second representation graph corresponding to the enterprise bipartite graph information; encoding the first representation graph using the recommendation model to obtain a first representation vector of the first representation graph; encoding the second representation graph using the recommendation model to obtain a second representation vector of the second representation graph; Calculating contrast loss using a plurality of samples consisting of the first representation vector and the second representation vector; Updating the parameters of the recommended model according to the contrast loss of the recommended model and repeating the training until the contrast loss meets the preset convergence condition, thereby obtaining a trained recommendation model; The first representation graph is obtained by edge loss and edge replacement, wherein a method based on degree centrality is used to assist in determining edge loss in an adjacency matrix. The expression of the first representation graph is as follows: ； The second representation graph is obtained by edge loss and edge replacement, wherein a method based on degree centrality is used to assist in determining edge loss in the adjacency matrix. The expression of the second representation graph is as follows: ;in, and are the two augmented matrices generated, It is a mask matrix obtained based on the degree centrality method; is a permutation matrix and T represents the transpose operation.

2. The policy-enterprise matching method based on a large model as claimed in claim 1, characterized in that: The contrast loss expression is as follows: ;in, For comparison loss, Compare losses for users, is the policy comparison loss.

3. The policy-enterprise matching method based on a large model as claimed in claim 2, characterized in that: The expression of the user contrast loss is as follows: ;in, is the user contrast loss, sim is the similarity function, and is the enhanced view of the same user node. is the temperature hyperparameter.

4. A policy-enterprise matching device based on a large model, characterized in that: The policy enterprise matching device based on the big model includes: A policy training set acquisition module, wherein the policy training set acquisition module is used to acquire a policy training set; An enterprise training set acquisition module, wherein the enterprise training set acquisition module is used to acquire an enterprise training set; A policy and enterprise bipartite graph information construction module, wherein the policy and enterprise bipartite graph information construction module is used to construct policy and enterprise bipartite graph information according to the policy training set and the enterprise training set; A training module, wherein the training module is used to train the recommendation model through the enterprise bipartite graph information, thereby obtaining a trained recommendation model; A recommendation result list acquisition module, wherein the recommendation result list acquisition module is used to input the information to be matched into a trained recommendation model to obtain a recommendation result list; The acquisition policy training set includes: Get policy text data; Through word embedding feature extraction technology, the policy text data is converted into vector representation, thereby obtaining a policy text vector training set; The obtaining of the enterprise training set comprises: Get enterprise text data; Through word embedding feature extraction technology, the enterprise text data is converted into vector representation, thereby obtaining the enterprise text vector training set; The construction of the policy and enterprise bipartite graph includes: Get policy and enterprise bipartite graph node information ,in, For enterprise set, is the policy set, is the edge between sets; Obtain two row vectors based on the policy and the enterprise bipartite graph and Similarity relations are used to construct edge information in the bipartite graph of enterprises; The recommendation model is a graph contrast learning model; The step of training the recommendation model through the enterprise bipartite graph information to obtain the trained recommendation model includes: Performing data enhancement twice on the enterprise bipartite graph information to obtain a first representation graph and a second representation graph corresponding to the enterprise bipartite graph information; encoding the first representation graph using the recommendation model to obtain a first representation vector of the first representation graph; encoding the second representation graph using the recommendation model to obtain a second representation vector of the second representation graph; Calculating contrast loss using a plurality of samples consisting of the first representation vector and the second representation vector; Updating the parameters of the recommended model according to the contrast loss of the recommended model and repeating the training until the contrast loss meets the preset convergence condition, thereby obtaining a trained recommendation model; The first representation graph is obtained by edge loss and edge replacement, wherein a method based on degree centrality is used to assist in determining edge loss in an adjacency matrix. The expression of the first representation graph is as follows: ； The second representation graph is obtained by edge loss and edge replacement, wherein a method based on degree centrality is used to assist in determining edge loss in the adjacency matrix. The expression of the second representation graph is as follows: ;in, and are the two augmented matrices generated, It is a mask matrix obtained based on the degree centrality method; is a permutation matrix and T represents the transpose operation.

Images