A business cooperation relationship mining model training method, a friend recommendation method and system

Abstract

The application provides a business cooperation relationship mining model training method, a friend recommendation method and a system. First, multiple source data are integrated to construct structured enterprise labels through optical character recognition and natural language processing technology, and an enterprise relationship graph is constructed based on the labels, taking enterprises as nodes and various business relationships as edges. Then, node pairs with business transactions are selected from the graph as positive samples, and node pairs that do not have transactions but meet specific conditions are selected as negative samples to form a training set. The model based on inductive graph neural network encoder, multi-relationship graph attention layer, cross-relationship fusion layer and link prediction layer is used for training, and the parameters are optimized by marginal contrast loss function. In application, the trained model is used to calculate the cooperation probability between the target enterprise and the unknown enterprise, and the potential cooperation partners are recommended to the user through instant messaging. The application significantly improves the efficiency and intelligent level of business expansion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an enterprise cooperation relationship mining model training method, a friend recommendation method and system. BACKGROUND

[0002] In the current enterprise operation and expansion process, instant messaging (IM) software has become an indispensable collaboration tool. However, the built-in friend recommendation function of existing instant messaging software (such as WeChat for Enterprise, DingTalk, etc.) is mostly limited to the personal social dimension, such as relying on mobile phone address book synchronization, colleague relationship within the same organizational structure, geographical proximity or fuzzy search matching. This kind of recommendation mechanism essentially serves acquaintanceship or internal collaboration, and is seriously out of touch with the core business needs of enterprise users seeking external potential business partners and building business upstream and downstream relationships. On the other hand, traditional ways of business expansion, such as attending offline exhibitions, relying on industry association referrals, making telephone sales or advertising, generally have problems such as low efficiency, high cost, information asymmetry and difficulty in continuous scaling, especially the inability to effectively discover and utilize the large number of hidden business relationships in the market. This leads to the contradiction of "information island" and "information overload" coexisting for enterprises: valuable connections are hidden in a large amount of data and are difficult to discover, while a large amount of screening cost is wasted in invalid information. Therefore, the existing technology lacks a solution that can deeply integrate enterprise multi-dimensional data, intelligently analyze potential business associations between enterprises, and accurately and actively recommend through instant messaging, a high-frequency application scenario. SUMMARY

[0003] In view of this, the embodiments of the present application provide an enterprise cooperation relationship mining model training method, a friend recommendation method and system to eliminate or improve one or more defects in the prior art, and solve the problem of low efficiency, high cost and weak information processing capability in the prior art in the mining of potential business relationships in the business expansion scenario.

[0004] In one aspect, the present application provides an enterprise cooperation relationship mining model training method, which comprises the following steps:

[0005] Obtaining raw data containing enterprise registration information, transaction history data, instant messaging history data and enterprise upstream and downstream relationship data based on a multi-source database and cleaning and storing the raw data, the raw data including picture or text format;

[0006] text recognition and keyword extraction on the raw data based on optical character recognition and natural language processing, constructing structured enterprise labels for various characteristic attributes; mining transaction relationships according to the transaction data, mining communication relationships according to the instant messaging history data, and mining business association relationships according to the enterprise upstream and downstream relationship data; constructing an enterprise relationship graph, taking enterprises as nodes, taking the enterprise labels as node attributes, and taking the transaction relationships, the communication relationships and the business association relationships as edges;

[0007] In the enterprise relationship graph, node pairs with business transactions are constructed as positive sample pairs, and node pairs that have not occurred business transactions but meet preset sampling constraint conditions are constructed as negative sample pairs; the positive sample pairs and the negative sample pairs are constructed as a training sample set;

[0008] An initial enterprise relationship prediction network based on an inductive graph neural network encoder, a multi-relation graph attention layer, a cross-relation fusion layer and a link prediction layer is obtained; the inductive graph neural network encoder encodes features of the enterprise relationship graph to obtain embedding representations of nodes, and the embedding representations of nodes in the positive sample pairs or the negative sample pairs are connected and output a prediction of a cooperation probability score through the multi-relation graph attention layer, the cross-relation fusion layer and the link prediction layer; a marginal contrast loss is constructed to update parameters of the initial enterprise relationship prediction network to obtain a target enterprise cooperation relationship mining model.

[0009] In some embodiments, the enterprise labels include basic labels for marking industry categories and regional distribution of enterprises, operating feature labels for marking core business content, dynamic behavior labels for marking transaction modes and cooperation preferences, network relationship labels for marking enterprise community attributes, and predictive analysis labels for marking transaction risks.

[0010] In some embodiments, the business transactions include enterprise pairs that have occurred business transactions, enterprise user pairs that have established connections in instant messaging history, and enterprise relationship pairs that have added friends and have business interactions, and the business transactions are required to be within a preset time period; the preset sampling constraint conditions include industry constraints, regional constraints, enterprise label approximation constraints and topological proximity constraints for the enterprise relationship graph.

[0011] In some embodiments, the inductive graph neural network encoder adopts a GraphSAGE model; the multi-relation graph attention layer extracts embedding representations for business cooperation relationships, friend relationships, supply chain relationships and industry relationships; the cross-relation fusion layer includes a fusion layer and a fully connected layer; and the link prediction layer adopts a multi-layer perceptron dot product model.

[0012] In some embodiments, the calculation formula of the marginal contrast loss is:

[0013] ;

[0014] wherein, represents the marginal contrast loss, and is the positive sample pair, and is the negative sample pair, represents the predicted value of the cooperation probability between the negative sample pairs, represents the predicted value of the cooperation probability between the positive sample pairs; margin represents a margin parameter, and N represents the number of samples.

[0015] In another aspect, the present application also provides a business friend recommendation method, the method comprising the following steps:

[0016] Based on the multi-source database, the original data containing enterprise registration information, transaction history data, instant messaging history data and enterprise upstream and downstream relationship data are obtained and cleaned and stored, and the original data includes picture or text format;

[0017] Based on optical character recognition and natural language processing, the original data is subjected to text recognition and keyword extraction, and structured enterprise labels for various feature attributes are constructed; transaction relationships are mined according to the transaction data, communication relationships are mined according to the instant messaging history data, and business association relationships are mined according to the enterprise upstream and downstream relationship data; an enterprise relationship graph is constructed, taking enterprises as nodes, taking the enterprise labels as node attributes, and taking the transaction relationships, the communication relationships and the business association relationships as edges;

[0018] For a target enterprise node, the enterprise relationship graph is input into the target enterprise cooperation relationship mining model in the enterprise cooperation relationship mining model training method of any one of claims 1 to 5 to output the cooperation probability score between the target enterprise node and the enterprise nodes without business transactions;

[0019] The first set number of candidate enterprise nodes with the highest cooperation probability score are screened out, and the recommendation cards about the candidate enterprise nodes are sent to the target enterprise node at a set frequency through an instant messaging system.

[0020] In some embodiments, the method further comprises:

[0021] Monitoring the feedback results of the target enterprise node to the recommendation cards, the feedback results including: ignoring the card, adding the card, whether to accept information after adding, whether to interact after adding, and interaction frequency;

[0022] The feedback result is used to update the enterprise relationship graph, and the target enterprise cooperation relationship mining model is optimized and updated based on the updated enterprise relationship graph.

[0023] In another aspect, the present application also provides an enterprise cooperation relationship mining system, comprising a processor, a memory and a computer program / instruction stored on the memory, wherein the processor is configured to execute the computer program / instruction, and the system implements the steps of the above method when the computer program / instruction is executed.

[0024] In another aspect, the present application also provides a computer readable storage medium, which stores a computer program / instruction, and the computer program / instruction implements the steps of the above method when executed by a processor.

[0025] In another aspect, the present application also provides a computer program product, which comprises a computer program / instruction, and the computer program / instruction implements the steps of the above method when executed by a processor.

[0026] The enterprise cooperation relationship mining model training method, friend recommendation method and system of the present application integrate enterprise registration information, transaction history, instant messaging records and known upstream and downstream relationships, and construct an enterprise relationship graph rich in semantic information. The graph takes enterprises as nodes, the attributes of which are composed of structured labels extracted from original data by optical character recognition and natural language processing technology, and the edges represent various actual relationships such as transactions, communications and business associations. On this basis, positive and negative training sample pairs are constructed from the graph, and an inductive graph neural network model is trained. The multi-relation graph attention layer is used to accurately capture the differentiated importance of different types of business relationships, and the cross-relation fusion layer is used to aggregate these information. Finally, the link prediction layer outputs the cooperation probability. The training process uses a margin contrast loss function to widen the score gap between positive and negative sample pairs, thereby improving the discrimination ability of the model. The present application converts scattered enterprise data into a dynamic knowledge system that can be deeply learned by machines, and can actively and accurately identify partners with high cooperation potential from a large number of enterprises, greatly improving the efficiency and accuracy of business expansion.

[0027] Additional advantages, objects, and features of the application will be set forth in part by the description that follows, and will become apparent to those skilled in the art upon examination of the following figures and detailed description thereof or can be learned by practice of the application. The objects and other advantages of the application can be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

[0028] It will be understood by those skilled in the art that the objects and advantages of the present application can be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0029] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0030] Figure 1 A flowchart of an enterprise cooperation relationship mining model training method according to an embodiment of the application.

[0031] Figure 2 A flowchart of an enterprise friend recommendation method according to another embodiment of the application.

[0032] Figure 3 A logic diagram of an enterprise friend recommendation scheme according to another embodiment of the application. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions, and advantages of the application clearer, the following will further describe the application with reference to the embodiments and drawings. Here, the illustrative embodiments of the application and their descriptions are used to explain the application, but are not intended to limit the application.

[0034] It should also be noted that, in order to avoid obscuring the application due to unnecessary details, only structures and / or processing steps closely related to the scheme according to the application are shown in the drawings, and other details not closely related to the application are omitted.

[0035] It should be emphasized that the term “comprises / comprising” when used in this text indicates the presence of the stated features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0036] The friend recommendation function of existing instant messaging software in the enterprise scenario is mainly based on personal social networks (such as address books, colleague relationships), and its recommendation logic is seriously out of touch with the business expansion needs of enterprises. Enterprises still highly rely on offline channels, manual search, and other traditional ways to find potential partners, which has problems such as low efficiency, high cost, difficulty in sustainability, and inability to discover hidden business relationships. The present application aims to integrate multi-dimensional data of enterprises, intelligently analyze potential business associations between enterprises, and provide a technical scheme for accurate recommendation and active reach through instant messaging software, a high-frequency application, to solve the above problems.

[0037] Specifically, the present application provides an enterprise cooperation relationship mining model training method, as shown in Figure 1 The method comprises the following steps S101-S104:

[0038] Step S101: Obtain raw data containing enterprise registration information, transaction history data, instant messaging history data, and enterprise upstream and downstream relationship data based on multi-source databases and clean and store them, the raw data including pictures or text formats.

[0039] Step S102: Perform text recognition and keyword extraction on the raw data based on optical character recognition and natural language processing, and construct structured enterprise labels for various feature attributes; mine transaction relationships based on transaction data, communication relationships based on instant messaging history data, and business association relationships based on enterprise upstream and downstream relationship data; construct an enterprise relationship graph, taking enterprises as nodes, enterprise labels as node attributes, and transaction relationships, communication relationships, and business association relationships as edges.

[0040] Step S103: In the enterprise relationship graph, construct node pairs with business transactions as positive sample pairs, and construct node pairs that have not occurred business transactions but meet the preset sampling constraint conditions as negative sample pairs; construct the positive sample pairs and the negative sample pairs as a training sample set.

[0041] Step S104: Obtain an initial enterprise relationship prediction network based on an inductive graph neural network encoder, a multi-relation graph attention layer, a cross-relation fusion layer, and a link prediction layer; the inductive graph neural network encoder encodes the features of the enterprise relationship graph to obtain the embedding representation of each node, and the embedding representation of the nodes in the positive sample pair or the negative sample pair is connected and then outputted through the multi-relation graph attention layer, the cross-relation fusion layer, and the link prediction layer to predict the cooperation probability score; build a marginal contrast loss to update the parameters of the initial enterprise relationship prediction network, and obtain a target enterprise cooperation relationship mining model.

[0042] In step S101, in order to accurately predict the cooperation relationship between enterprises, multi-dimensional data is integrated, and information from different channels reflecting different aspects of enterprises is combined. Enterprise registration information usually comes from business databases and contains basic static attributes of enterprises, such as enterprise name, registered address, legal representative, registered capital, business scope, etc. This is the core of building an enterprise basic portrait. Transaction history data mainly comes from ERP (Enterprise Resource Planning) and CRM (Customer Relationship Management) systems of enterprises. This is the proof of the existence of substantial business transactions between enterprises, including purchase / sale orders, contracts, invoices, payment records, etc. Instant messaging history data comes from the internal of instant messaging (IM) platforms and records the communication behavior between enterprise employees, such as added friend relationships, chat frequency, group common members, etc. This reflects the degree of communication between enterprises. Enterprise upstream and downstream relationship data may come from enterprise self-labeling, supply chain platforms or public reports, and clearly indicates the supplier-customer relationship between enterprises.

[0043] After obtaining the original data, due to the diversity of data sources and the difference in formats, such as the business license is a picture, the contract is a text, and the transaction record is structured data, the original data needs to be cleaned, standardized, and aligned, and stored for subsequent analysis.

[0044] Step S102 is a key step for converting the original data into structured knowledge, and the technology of using a graph model to depict entities and their relationships. It can intuitively record the complex relationship network between enterprises.

[0045] The optical character recognition technology is used to analyze the picture data such as the business license of an enterprise, and the key text information (such as the scope of business) is converted into processable text data. Then, the natural language processing technology is used to deeply analyze the text, and operations including word segmentation, entity recognition, keyword extraction, and industry classification are performed, so as to construct multi-dimensional structured feature labels for each enterprise, such as “industry: intelligent manufacturing”, “main product: industrial robot”, “scale: medium” and the like. These labels together constitute the attribute vector of the node in the graph. On this basis, the system mines the association between enterprises from different dimensions to construct the edges of the graph: based on the transaction records (such as orders and contracts) in the ERP and CRM systems, the transaction relationship is mined; based on the friend relationship and interaction frequency of the instant messaging platform, the communication relationship is mined; according to the known supply chain directory or enterprise self-description data, the business association relationship is mined, such as the explicit upstream and downstream and investment relationship. Finally, all these information is imported into a graph database such as the Neo4j database, forming a large heterogeneous information network with enterprises as nodes and various business relationships as edges. This network not only statically depicts the attributes and direct associations of enterprises, but also the graph structure itself contains rich implicit business logic and community features that can be mined by machine learning models, laying a solid data foundation for subsequent intelligent prediction.

[0046] In some embodiments, the enterprise labels include basic labels for marking the regional distribution of the industry category of the enterprise, operating feature labels for marking the core business content, dynamic behavior labels for marking the transaction mode and cooperation preference, network relationship labels for marking the community attributes of the enterprise, and predictive analysis labels for marking the transaction risk.

[0047] The basic label is usually directly mapped from the industry code and registered address in the enterprise registration information, such as "industry: manufacturing" and "region: Yangtze River Delta". The operating feature label is mainly obtained by natural language processing technology on the "business scope" text in the business license, such as "core business: precision bearing manufacturing". The dynamic behavior label is the behavior mode obtained by analyzing historical transaction data and contract records, such as "transaction mode: high frequency small amount" and "cooperation preference: state-owned enterprise". The network relationship label is based on the constructed enterprise relationship graph, and the centrality and clustering coefficient of the node are analyzed by using graph calculation algorithm, or it is identified by community discovery algorithm, such as "network centrality: high" and "belonging community: automobile industry chain cluster". The predictive analysis label is a derivative index generated by inputting all the above labels and time series data into a special machine learning prediction model to infer the future state of the enterprise, such as "transaction risk: medium" and "growth potential: high". The label system collectively constitutes a three-dimensional and quantitative digital portrait of the enterprise.

[0048] In step S103, positive and negative sample pairs are constructed based on the enterprise relationship graph. The positive sample pair is the enterprise pair that has actually cooperated in the historical data. When constructing the negative sample pair, based on the preset sampling constraint condition, a difficult-to-distinguish negative sample is constructed, that is, those enterprises that look like they should cooperate but actually do not cooperate.

[0049] Specific business transactions include enterprise pairs that have conducted business transactions, enterprise user pairs that have established connections in instant messaging history, and enterprise relationship pairs that have added friends and have business interactions, and the business transactions are required within a preset time period. The preset sampling constraint conditions include industry constraint, regional constraint, enterprise label approximation constraint and topological proximity constraint of the enterprise relationship graph. Among them, the enterprise label approximation constraint requires the similarity of the enterprise labels between the negative sample pairs to be higher than a set value, and the topological proximity constraint requires the node distance between the negative samples in the enterprise relationship graph to be less than a set hop number.

[0050] In step S104, the role of the inductive graph neural network encoder is to read the structure and node attributes of the entire graph, and generate an embedded representation in the form of a vector for each enterprise node. This vector combines all the information of the enterprise and its network location features. Its inductive property means that it can generate a vector for new enterprises that have not been seen during training, ensuring the scalability of the model.

[0051] The multi-relation graph attention layer is the core of the model and can identify the importance of different relations. The multi-relation graph attention layer is a graph neural network component that simultaneously considers the relation type and node content: it first splits each edge into several subgraphs according to the relation name, then independently runs a multi-head attention mechanism within each relation to calculate the weight coefficient of who should pay more attention to whom, thereby obtaining a relation-specific node embedding. In this way, information in different semantic dimensions such as supply chain, transaction, friends, and the same industry can be modeled separately, and finally these vectors in the perspective of relations are weighted and summarized through a fusion layer, so that the model can both distinguish the difference between cooperation and chatting and comprehensively judge the possibility of business transactions between two companies. In some embodiments, the inductive graph neural network encoder adopts a GraphSAGE model; the multi-relation graph attention layer extracts embedding representations for business cooperation relations, friend relations, supply chain relations, and the same industry relations, respectively.

[0052] The cross-relation fusion layer fuses the multiple embedding representations output by the multi-relation graph attention layer for an enterprise node to form a more comprehensive and powerful node representation. The cross-relation fusion layer includes a fusion layer and a fully connected layer.

[0053] The link prediction layer receives the final vector representations of two enterprises, performs calculations, and outputs a cooperation probability score between 0 and 1. The link prediction layer adopts a multi-layer perceptron dot product model.

[0054] During the model training process, the margin contrast loss is the learning goal of the model. It forces the score of the positive sample pair to be significantly higher than that of the corresponding negative sample pair. The process of adjusting the model parameters is to constantly widen this score difference. For example, during training, the model will process both positive sample pairs and negative sample pairs. The loss function will calculate that if the model scores the positive sample as 0.9 and the negative sample as 0.7, although the order is correct, the score difference is not large enough, for example, the target score difference is 0.5, the model will be punished and adjust the internal parameters accordingly to strive to push the positive sample score to 0.95 and the negative sample score to 0.4 in the next round.

[0055] In some embodiments, the calculation formula of the margin contrast loss is:

[0056] ;

[0057] wherein, represents the margin contrast loss, and are positive sample pairs, and are negative sample pairs, represents the predicted value of the cooperation probability between the negative sample pairs, Pij represents a predicted value of a cooperation probability between a positive sample pair; margin represents a margin parameter, and N represents a sample quantity.

[0058] In another aspect, the present application also provides a business friend recommendation method, as shown in the method comprising the following steps S201-S204: Figure 2

[0059] Step S201: Obtain raw data containing business registration information, transaction history data, instant messaging history data and business upstream and downstream relationship data based on a multi-source database and clean and store the raw data, the raw data including pictures or text formats.

[0060] Step S202: Perform text recognition and keyword extraction on the raw data based on optical character recognition and natural language processing, and construct structured business labels for various feature attributes; mine transaction relationships according to transaction data, mine communication relationships according to instant messaging history data, and mine business association relationships according to business upstream and downstream relationship data; construct a business relationship graph, taking businesses as nodes, business labels as node attributes, and transaction relationships, communication relationships and business association relationships as edges.

[0061] Step S203: For a target business node, input the business relationship graph into the target business cooperation relationship mining model in the business cooperation relationship mining model training method described above to output a cooperation probability score between the target business node and a business node without business transactions.

[0062] Step S204: Screen out a first set number of candidate business nodes with the highest cooperation probability scores, and send a recommendation card about the candidate business nodes to the target business node through an instant messaging system according to a set frequency.

[0063] Part of steps S201 and S202 can refer to the description of steps S101 and S102, and the purpose is to establish a business relationship graph based on data mining for subsequent recognition processing.

[0064] ​Step S203 is the core reasoning and application stage of the present application. In this step, the trained target enterprise cooperation relationship mining model is deployed to the production environment for performing actual prediction tasks. Specifically, the system inputs the entire, real-time updated enterprise relationship graph together with the specified target enterprise node (i.e., the enterprise for which partners are to be found) into the model. The model, by virtue of its complex pattern recognition ability learned in the training stage, performs forward propagation calculation on the graph: first, it generates embedding representations for all nodes in the graph (including the target enterprise) that contain rich structural and attribute information through its internal graph neural network encoder; then, the model iterates through the node pairs formed by the target enterprise and all enterprise nodes in the graph that have no business transactions (i.e., enterprises that have not established direct transactions or friend relationships), and calculates their cooperation probability scores one by one via the multi-relation attention layer, fusion layer, and link prediction layer. This score is a continuous value between 0 and 1, quantifying the likelihood of the two enterprises establishing a cooperative relationship in the future from the model's perspective. This process realizes the intelligent and efficient screening of the most potential cooperation targets from the massive and complex enterprise network data.

[0065] Step S204 is a key step for converting intelligent prediction results into actual business value, focusing on recommendation strategies and user experience. This step first sorts all probability scores output by step S203 and selects the top first set number of enterprises with the highest scores, such as the top ten as candidate enterprise nodes. Subsequently, the system encapsulates the information of these candidate enterprises into a recommendation card and pushes it to the relevant employees of the target enterprise through the interface of the instant messaging system. The form can include a possible friend list or a specific recommendation message card. To ensure the effectiveness of the recommendation and avoid disturbing the user, this step introduces a set frequency to control the push frequency. This means that the recommendation is not a one-time bombardment, but follows a strategic rhythm, such as pushing on a weekly or monthly basis, or intelligently linking with user activity and business scenarios, such as the user searching for related products. This design not only ensures the proactive reach of business opportunities but also fully respects user experience, allowing the recommendation system to naturally integrate into the user's daily workflow, thereby efficiently and cost-effectively facilitating new business connections.

[0066] In some embodiments, the method further includes steps S205 and S206:

[0067] Step S205: Monitor the feedback results of the target enterprise node on the recommendation card, including ignoring the card, adding the card, whether accepting information after adding, whether interacting after adding, and interaction frequency.

[0068] Step S206: Update the enterprise relationship graph using the feedback results and optimize the target enterprise cooperation relationship mining model based on the updated enterprise relationship graph.

[0069] Step S205 is a key data collection link for building the self-evolution capability of the system. The core of this step is to finely monitor and quantify the full-link behavior of the user after receiving the recommendation. The feedback result recorded by the system is a signal sequence from shallow to deep, with increasing value: ignoring the business card is a direct negative feedback, indicating that this recommendation may not match the user's needs; adding the business card is a strong positive interest signal; and whether to accept the information after adding and the subsequent interaction frequency further measure the activity and quality after relationship establishment, and high-frequency deep interaction is the highest verification of the value of cooperative relationship. By systematically capturing these behaviors, the system converts the originally subjective and fuzzy user intention into structured and quantifiable data, providing important real-time feedback samples for subsequent model optimization.

[0070] Step S206 realizes the core closed-loop optimization logic of the system. In this step, the feedback results collected from S205 are used to dynamically update the enterprise relationship graph. For example, when two enterprise users successfully add as friends through recommendation, an instant messaging friend edge will be added between the corresponding two nodes in the graph; if they have business interactions afterwards, the weight or attribute of the edge will be enhanced. This real-time updated and more rich graph constitutes the new data basis for model optimization. Subsequently, the system can use these new positive samples (such as enterprise pairs that successfully establish connection and interact) and negative samples (such as recommendations that are continuously ignored) to perform incremental training or fine-tuning on the deployed target enterprise cooperative relationship mining model. This enables the model to continuously learn the latest user behavior patterns and constantly correct its prediction bias, so that the entire system forms an enhanced cycle of recommendation-feedback-learning-more accurate recommendation, and becomes more and more intelligent over time.

[0071] On the other hand, the present application also provides an enterprise cooperative relationship mining system, comprising a processor, a memory and a computer program / instructions stored on the memory, wherein the processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the above method.

[0072] On the other hand, the present application also provides a computer readable storage medium, which stores a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the above method are implemented.

[0073] On the other hand, the present application also provides a computer program product, comprising a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the above method are implemented.

[0074] The present application will be described below in conjunction with a specific embodiment:

[0075] The embodiments of the present year provide an intelligent and automated enterprise friend recommendation scheme, which aims to accurately predict and recommend enterprises most likely to have business transactions by analyzing internal and platform data, thereby improving business connection efficiency. Figure 3

[0076] The embodiments include the following steps:

[0077] 1) Data collection and integration: Collect and integrate data from multiple sources, including user-added enterprise friends, a large enterprise registration database (including detailed business scope), business transaction data (such as orders, contracts, invoices, etc.) between enterprises, and known upstream and downstream relationships of enterprises. Data comes from the ERP (Enterprise Resource Planning) and CRM (Customer Relationship Management) systems of enterprises, or is extracted, cleaned, and stored from related enterprise collaboration platforms; at the same time, relevant business scope is obtained from business license pictures.

[0078] 2) Enterprise portrait construction: Based on the above data, a multi-dimensional feature portrait is constructed for each enterprise, including industry labels, product / service keywords, business scale, relationship network characteristics, etc. Process text data, and extract keywords and industry classification and related features through natural language processing (NLP) technology.

[0079] Specifically, based on NLP natural language processing technology, keywords are extracted from the business scope text and industry classification is performed, multi-dimensional feature quantization is performed to form a structured enterprise label system, and enterprise state features in the enterprise label system are updated in real time based on business data.

[0080] The enterprise portrait construction method uses a multi-level label system to build, including: structured basic labels: industry classification, regional distribution, and other static attributes based on rule mapping; NLP extracted business feature labels: core business features are mined from business scope text through multi-algorithm fusion keyword extraction technology; dynamic behavior labels: transaction patterns, cooperation preferences, and other behavior characteristics based on business transaction data analysis; graph structure network labels: centrality, community attributes, and other network features calculated from enterprise relationship graphs; predictive analysis labels: growth, risk level, and other prediction indicators derived through machine learning models.

[0081] 3) Relationship graph construction: Use graph database technology to construct an enterprise knowledge graph with enterprises as nodes and various relationships as edges. Edge relationships can include upstream and downstream, cooperation, same industry, and added friends. For example, static relationships in the relationship graph include upstream and downstream associations, same industry relationships, and same region relationships. Dynamic relationships in the relationship graph include business transaction frequency and instant messaging interaction intensity. Displayed relationships in the relationship graph include added friend relationships, etc.

[0082] ​4) Machine learning model training: using historical business data as positive samples, train a prediction model. The model learns the feature patterns of business relationships between enterprises. The model includes an inductive graph neural network encoder, a multi-relation graph attention layer, a cross-relation fusion layer, and a link prediction layer.

[0083] In the relationship graph, construct positive and negative sample pairs. Positive samples are pairs of enterprise nodes that have actually cooperated in historical data, including pairs of enterprises that have business data, i.e., actual business records such as orders, contracts, and invoices; pairs of enterprise users that have successfully established connections in instant messaging systems; pairs of enterprises that have high-value interaction relationships after adding friends; and pairs of enterprises that have effective business interactions. At the same time, set a time window constraint to limit effective business relationships within the training time window. Negative sample pairs refer to pairs of enterprise nodes that have not cooperated in historical data but have potential correlation possibilities. Select pairs of enterprises in the same industry that have no business cooperation as same-industry negative samples, pairs of enterprises in the same region that have no cooperation as same-geographical negative samples, pairs of enterprises with high feature vector similarity based on enterprise portrait but no actual relationship as feature-similar negative samples, and pairs of enterprises in the graph that are less than a set number of hops apart but have no direct business relationship as topologically adjacent negative samples.

[0084] During model training, construct a margin contrast loss to make the cooperation probability score of positive samples significantly higher than that of negative samples.

[0085] The calculation formula of the margin contrast loss is:

[0086] ;

[0087] wherein, represents the margin contrast loss, and is a positive sample pair, and is a negative sample pair, represents the predicted value of the cooperation probability between negative sample pairs, represents the predicted value of the cooperation probability between positive sample pairs; margin represents the margin parameter, and N represents the number of samples.

[0088] 5) Potential partner prediction and recommendation: for a target enterprise, use the trained model to calculate the "cooperation probability" score among its unconnected enterprises, and select the top several enterprises with the highest scores.

[0089] 6) Presentation of recommendation results: in the interface of instant messaging software, such as the "Discovery" page and the "Friend Addition" page, present the recommendation results to the employees of the target enterprise in the form of "Recommended Business Cards".

[0090] The recommended information includes the user ID, name, and avatar of the recommended user, the enterprise ID, name, and business scope of the enterprise information; and the push location can be set in the possible enterprise friend push box.

[0091] 7) Recommendation frequency: The recommendation period and frequency are set through a timing task or chat room subscription, so as to avoid disturbing some unnecessary enterprise users. Specifically, the triggering time can be based on user activity and intelligent triggering of business scenarios, the recommendation frequency is controlled to avoid disturbing the user, and the recommendation content and form can be adjusted according to different user preferences.

[0092] 8) Closed-loop scenario collection: The subsequent operation behavior of the user who receives the "recommended business card" is monitored: the operation feedback of the user to the recommended enterprise business card is collected: for example, the card is ignored, the card is added, whether the opposite enterprise user receives after adding, whether interaction occurs after forming a friend relationship, and the related interaction frequency will be used as sample data to enrich the enterprise portrait and relationship graph, to realize the training data of the intelligent recommendation big model, form a relationship closed loop, and more and more accurate and complete recommendation results, and improve the success rate.

[0093] These behavior data are used as new samples and signals, and are returned to the data layer and the model layer. Positive feedback (such as adding and generating interaction) will strengthen the correctness of the existing model judgment. Negative feedback (such as ignoring) will help the model correct the wrong judgment. Thus, the system forms an enhanced closed loop of "recommendation, feedback, learning, and more accurate recommendation", so that the entire system becomes more and more intelligent over time, and the recommendation success rate becomes higher and higher.

[0094] Therefore, the embodiment associates, fuses, and cooperatively analyzes the four types of originally isolated data, including the private relationship chain related to the added enterprise friend, the static attributes of the enterprise in the enterprise database, the dynamic behavior of the enterprise recorded in the business data, and the known upstream and downstream relationships. By constructing an enterprise knowledge graph, not only the enterprise information is stored in the database, but also a graph rich in semantics is constructed, with enterprises as nodes and various types of relationships as edges. The types of edges include but are not limited to added friend relationship, upstream and downstream relationship, supply relationship, procurement relationship, same industry relationship, and same region relationship. This multi-dimensional heterogeneous graph and portrait are the basis for subsequent intelligent recommendation.

[0095] By constructing a positive and negative sample pair to execute a contrastive learning training model, the embedding features of the enterprise nodes in the graph are mined based on an inductive graph neural network model, the differentiated importance of different types of business relationships is accurately captured by a multi-relationship graph attention layer, and these information is aggregated by a cross-relationship fusion layer. Finally, the cooperation probability is output by a link prediction layer.

[0096] A recommended feedback business closed loop is constructed, for example, the operation behaviors (add, ignore, click, reject, pass, and the interaction frequency generated after passing the application, etc.) of the user to the recommended business card are recorded as new feedback data, which is used to update the model and the graph in real time, so that the subsequent recommended results are more accurate and efficient, and a complete closed loop is realized.

[0097] Further, the recommendation function is deeply integrated into the specific entrance of the IM message, and is presented in the form of a native, lightweight and easy-to-operate "recommended business card", which greatly reduces the use and connection cost of the user.

[0098] Corresponding to the above method, the application also provides a device / system, which comprises a computer device including a processor and a memory, the memory storing computer instructions, and the processor being configured to execute the computer instructions stored in the memory, so that the device / system implements the steps of the above method.

[0099] The embodiment of the application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the above edge computing server deployment method. The computer readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0100] In summary, the enterprise cooperation relationship mining model training method, the friend recommendation method and the system disclosed by the application construct an enterprise relationship graph rich in semantic information by integrating enterprise registration information, transaction history, instant messaging records and known upstream and downstream relationships and the like. The graph takes enterprises as nodes, the attributes of which are composed of structured labels extracted from original data by optical character recognition and natural language processing technology, and the edges represent various actual relationships such as transactions, communications and business associations. On this basis, positive and negative training sample pairs are constructed from the graph, and an inductive graph neural network model is trained, the differentiated importance of different types of business relationships is accurately captured by using a multi-relation graph attention layer, and these information is converged by a cross-relation fusion layer, and finally a cooperation probability is output by a link prediction layer. The training process adopts a margin contrast loss function, aiming to enlarge the score gap between positive and negative sample pairs, so as to improve the discrimination ability of the model. The scattered enterprise data is converted into a dynamic knowledge system that can be deeply learned by a machine, which can actively and accurately identify partners with high cooperation potential from a large number of enterprises, greatly improving the efficiency and accuracy of business expansion.

[0101] Those of ordinary skill in the art will appreciate that the various illustrative components, systems and methods described in connection with the embodiments disclosed herein can be implemented as hardware, software, or both. The particular implementation is dependent on the specific application and design constraints imposed on the overall system. Skilled persons can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application. When implemented in hardware, for example, the hardware can comprise an electronic circuit, an Application Specific Integrated Circuit (ASIC), a suitable firmware, a plug-in, a functional card, etc. When implemented in software, the elements of the application are the program or code segments to perform a specific task. The program or code segments can be stored in a machine-readable medium, or transmitted by a carrier wave as data signals over a transmission medium or communication link.

[0102] It is to be understood that the application is not limited to the particular configurations and processes described herein and shown in the drawings. For simplicity, detailed descriptions of known methods and apparatuses are omitted so as not to obscure the disclosure. In the above-described embodiments, several specific steps are described and illustrated as examples. However, the method processes of the present application are not limited to the specific steps described and illustrated, and the order of the steps can be changed, or other steps can be added, or replaced, or eliminated, depending on the application.

[0103] In the present application, features described and / or illustrated in relation to one embodiment can be used in the same or a similar way in one or more other embodiments, and / or combined with or instead of features of other embodiments.

[0104] The above description is only preferred embodiments of the present application, and is not intended to limit the present application. The embodiments of the present application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the present application.

Claims

1. A method for training a corporate partnership mining model, characterized in that, The method includes the following steps: The raw data, including enterprise registration information, transaction history data, instant messaging history data, and enterprise upstream and downstream relationship data, is obtained from a multi-source database, cleaned, and stored. The raw data includes image or text formats. Based on optical character recognition and natural language processing, text recognition and keyword extraction are performed on the raw data to construct structured enterprise tags for multiple feature attributes; transaction relationships are mined based on the transaction data, communication relationships are mined based on the instant messaging history data, and business relationships are mined based on the enterprise's upstream and downstream relationship data; an enterprise relationship graph is constructed, with enterprises as nodes, enterprise tags as node attributes, and transaction relationships, communication relationships, and business relationships as edges; In the enterprise relationship graph, node pairs with business interactions are constructed as positive sample pairs, and node pairs without business interactions but meeting preset sampling constraints are constructed as negative sample pairs; the positive sample pairs and the negative sample pairs are constructed as a training sample set; An initial enterprise relationship prediction network is obtained based on an inductive graph neural network encoder, a multi-relationship graph attention layer, a cross-relationship fusion layer, and a link prediction layer. The inductive graph neural network encoder performs feature encoding on the enterprise relationship graph to obtain the embedding representation of each node. The embedding representations of nodes in the positive sample pair or the negative sample pair are connected and then output as predicted cooperation probability scores through the multi-relationship graph attention layer, the cross-relationship fusion layer, and the link prediction layer. A marginal contrast loss is constructed to update the parameters of the initial enterprise relationship prediction network to obtain a target enterprise cooperation relationship mining model.

2. The business partnership mining model training method of claim 1, wherein, The enterprise tags include basic tags for marking the geographical distribution of the enterprise's industry category, operational characteristic tags for marking the core business content, dynamic behavioral tags for marking transaction patterns and cooperation preferences, network relationship tags for marking the enterprise's community attributes, and predictive analysis tags for marking transaction risks.

3. The business partnership mining model training method of claim 1, wherein, The business interactions include enterprise pairs that have already had business transactions, enterprise user pairs that have established connections in the instant messaging history, and enterprise relationship pairs that have added each other as friends and have business interactions. The business interactions are required to occur within a preset time period. The preset sampling constraints include industry-specific constraints, region-specific constraints, enterprise tag similarity constraints, and topological proximity constraints for the enterprise relationship graph.

4. The training method for the enterprise cooperation relationship mining model according to claim 1, characterized in that, The inductive graph neural network encoder uses the GraphSAGE model; the multi-relationship graph attention layer extracts embedding representations for business cooperation relationships, friend relationships, supply chain relationships, and industry peer relationships respectively; The cross-relationship fusion layer includes a fusion layer and a fully connected layer; the link prediction layer adopts a multilayer perceptron dot product model.

5. The training method for the enterprise cooperation relationship mining model according to claim 1, characterized in that, The formula for calculating the marginal comparison loss is: ； in, This represents the marginal contrast loss. and For the positive sample pair, and For the negative sample pair, This represents the predicted probability of cooperation between the negative sample pairs. The value represents the predicted probability of cooperation between the positive sample pairs; margin represents the marginal parameter, and N represents the number of samples.

6. A method for recommending business partners, characterized in that, The method includes the following steps: The raw data, including enterprise registration information, transaction history data, instant messaging history data, and enterprise upstream and downstream relationship data, is obtained from a multi-source database, cleaned, and stored. The raw data includes image or text formats. Based on optical character recognition and natural language processing, text recognition and keyword extraction are performed on the raw data to construct structured enterprise tags for multiple feature attributes; transaction relationships are mined based on the transaction data, communication relationships are mined based on the instant messaging history data, and business relationships are mined based on the enterprise's upstream and downstream relationship data; an enterprise relationship graph is constructed, with enterprises as nodes, enterprise tags as node attributes, and transaction relationships, communication relationships, and business relationships as edges; For the target enterprise node, the enterprise relationship graph is input into the target enterprise cooperation relationship mining model in the training method of enterprise cooperation relationship mining model according to any one of claims 1 to 5, so as to output the cooperation probability score between the target enterprise node and the enterprise node with no business dealings. A first set number of candidate enterprise nodes with the highest cooperation probability scores are selected, and recommended business cards of the candidate enterprise nodes are sent to the target enterprise node at a set frequency through an instant messaging system.

7. The enterprise friend recommendation method according to claim 6, characterized in that, The method further includes: Monitor the feedback results of the target enterprise node to the recommended business card. The feedback results include: ignoring the business card, adding the business card, whether to accept the information after adding the business card, whether to interact after adding the business card, and the frequency of interaction. The feedback results are used to update the enterprise relationship graph, and the target enterprise cooperation relationship mining model is optimized and updated based on the updated enterprise relationship graph.

8. A system for mining enterprise partnerships, comprising a processor, a memory, and a computer program / instructions stored in the memory, characterized in that, The processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.

Images