Product knowledge graph construction method and device, computer equipment and storage medium

Abstract

The application relates to the technical field of artificial intelligence, and discloses a product knowledge graph construction method and device, computer equipment and a storage medium, wherein the method comprises the following steps: performing fusion processing on data, forming a preset structural field, performing space mapping, obtaining a first semantic vector, and finally constructing a product knowledge graph based on the first semantic vector.The application has the beneficial effect of realizing integrated analysis of data, improving the efficiency of constructing a knowledge graph, and further providing important support for subsequent intelligent query, data mining and a recommendation system, thereby helping an enterprise to quickly respond to market demand, optimize product combination and improve user experience.

Description

Technical Field

[0001] The present invention relates to the technical field of constructing product knowledge graphs, and in particular to a method, device, computer equipment and storage medium for constructing a product knowledge graph. Background Art

[0002] When building a product knowledge graph, data generally comes from structured data in the ERP system and / or unstructured text data on the e-commerce platform. Generally speaking, the data in the ERP system is usually highly standardized and structured, with clear field definitions and relationship models, while the unstructured data on the e-commerce platform is highly diverse, with complex information presentation forms, and often lacks unified standards.

[0003] Since there is no unified standard for unstructured data, enterprises face many challenges in effective data integration and analysis, which makes it difficult to build knowledge graphs, resulting in a significant reduction in the efficiency of knowledge extraction and representation. Summary of the Invention

[0004] Based on this, it is necessary to address the problem of constructing existing product knowledge graphs and propose a method, device, computer equipment and storage medium for constructing product knowledge graphs.

[0005] A method for constructing a product knowledge graph, the method comprising:

[0006] Get the product name and unstructured text data for each product;

[0007] Performing fusion processing on the unstructured text data to obtain preset structured fields;

[0008] Performing spatial mapping on the preset structural field to obtain a first semantic vector;

[0009] Based on the product name of each product and the first semantic vector, a key-value pair is formed;

[0010] Mining the association information between the products in the preset structural fields through a preset GNN network;

[0011] The association information between each product and the corresponding key-value pairs are imported into a preset graph database to obtain a product knowledge graph.

[0012] Furthermore, the step of fusing the unstructured text data to obtain a preset structured field includes:

[0013] Extracting total preset core attributes of the unstructured text data of each product;

[0014] Analyzing the preset core attributes through dependency syntax to construct a relationship chain of each of the products;

[0015] A preset structural field is generated based on the relationship chain.

[0016] Furthermore, the step of forming a key-value pair based on the product name of each product and the first semantic vector includes:

[0017] Obtaining an ERP field for each product, wherein the ERP field has the same structure description as the preset structured field;

[0018] Performing spatial mapping on the ERP field to obtain a second semantic vector;

[0019] A key-value pair is formed based on the product name, the first semantic vector, and the second semantic vector of each product.

[0020] Furthermore, before the step of mining the association information between the products in the preset structural field through the preset GNN network, the step further includes:

[0021] Obtain multiple basic categories under existing standards and use the basic categories as the clustering core;

[0022] Clustering each product using the K-means algorithm on the basic clustering core to obtain a target clustering core and a cluster cluster corresponding to each target clustering core;

[0023] Screening the preset structural fields based on the clusters to obtain a target preset structural field corresponding to each cluster;

[0024] The preset GNN network is used to mine the association information between products in each cluster in the target preset structural field.

[0025] Furthermore, after the step of importing the association information and corresponding key-value pairs between the products into a preset graph database to obtain a product knowledge graph, the method further includes:

[0026] Determine whether the preset recommendation instruction of the specified user is triggered;

[0027] If the preset recommendation instruction is triggered, obtain the user's product information of interest;

[0028] Based on the product information of interest and the product knowledge graph, similar products are recommended by associating the graphs.

[0029] Furthermore, after the step of importing the association information and corresponding key-value pairs between the products into a preset graph database to obtain a product knowledge graph, the method further includes:

[0030] Obtain the target product name and target unstructured text data of the product to be added;

[0031] Performing fusion processing on the target unstructured text data to obtain target preset structured fields;

[0032] Performing spatial mapping on the target preset structural field to obtain a first semantic vector of the target;

[0033] Based on the target product name and the target first semantic vector, forming a target key-value pair;

[0034] Mining the target association information between the product to be added and other products in the product knowledge graph in the target preset structural field through a preset GNN network;

[0035] Import the target key-value pairs and the target association information into the product knowledge graph to obtain an updated product knowledge graph.

[0036] Furthermore, after the step of obtaining the product name and unstructured text data of each product, the method further includes:

[0037] Determine whether the preset recommendation instruction of the specified user is triggered;

[0038] If the preset recommendation instruction is triggered, obtain the user's product information of interest;

[0039] Based on the product information of interest and the product knowledge graph, similar products are recommended based on the similarity of semantic vectors.

[0040] A device for constructing a product knowledge graph, comprising:

[0041] The acquisition module is used to obtain the product name and unstructured text data of each product;

[0042] A fusion module, configured to perform fusion processing on the unstructured text data to obtain a preset structured field;

[0043] A mapping module, configured to perform spatial mapping on the preset structural field to obtain a first semantic vector;

[0044] a forming module, configured to form a key-value pair based on the product name of each product and the first semantic vector;

[0045] A mining module, configured to mine the association information between the products in the preset structural fields through a preset GNN network;

[0046] The import module is used to import the association information between each product and the corresponding key-value pairs into a preset graph database to obtain a product knowledge graph.

[0047] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0048] Get the product name and unstructured text data for each product;

[0049] Performing fusion processing on the unstructured text data to obtain preset structured fields;

[0050] Performing spatial mapping on the preset structural field to obtain a first semantic vector;

[0051] Based on the product name of each product and the first semantic vector, a key-value pair is formed;

[0052] Mining the association information between the products in the preset structural fields through a preset GNN network;

[0053] The association information between each product and the corresponding key-value pairs are imported into a preset graph database to obtain a product knowledge graph.

[0054] A computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0055] Get the product name and unstructured text data for each product;

[0056] Performing fusion processing on the unstructured text data to obtain preset structured fields;

[0057] Performing spatial mapping on the preset structural field to obtain a first semantic vector;

[0058] Based on the product name of each product and the first semantic vector, a key-value pair is formed;

[0059] Mining the association information between the products in the preset structural fields through a preset GNN network;

[0060] The association information between each product and the corresponding key-value pairs are imported into a preset graph database to obtain a product knowledge graph.

[0061] The beneficial effects of the present invention are as follows: by fusing data to form preset structural fields and performing spatial mapping, a first semantic vector is obtained, and finally a product knowledge graph is constructed based on the first semantic vector, thereby realizing integrated analysis of data and improving the efficiency of constructing the knowledge graph, thereby providing important support for subsequent intelligent query, data mining and recommendation systems, thereby helping enterprises to quickly respond to market demand, optimize product portfolios and enhance user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] 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 only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] in:

[0064] Figure 1 This is an application environment diagram of a method for constructing a product knowledge graph in one embodiment;

[0065] Figure 2 A flowchart of constructing a product knowledge graph in one embodiment;

[0066] Figure 3 This is a structural block diagram of a device for constructing a product knowledge graph in one embodiment;

[0067] Figure 4 FIG. 1 is a structural block diagram of a computer device in one embodiment. DETAILED DESCRIPTION

[0068] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0069] Figure 1 This is a diagram of the application environment for building a product knowledge graph in an embodiment. Figure 1The product knowledge graph construction method is applied to a product knowledge graph construction system. The product knowledge graph construction system comprises a terminal 110 and a server 120. The terminal 110 and the server 120 are connected through a network. The terminal 110 can be a desktop terminal or a mobile terminal. The mobile terminal can be at least one of a mobile phone, a tablet computer, a notebook computer, and the like. The server 120 can be implemented by an independent server or a server cluster composed of multiple servers. The terminal 110 is configured to acquire unstructured text data. The server 120 is configured to construct a product knowledge graph.

[0070] As shown in FIG. 1, Figure 2 In one embodiment, a product knowledge graph construction method is provided. The method can be applied to a terminal or a server. In this embodiment, the method is applied to a server. The product knowledge graph construction method comprises the following steps:

[0071] S1: acquiring product names and unstructured text data of each product;

[0072] S2: performing fusion processing on the unstructured text data to obtain a preset structural field;

[0073] S3: performing spatial mapping on the preset structural field to obtain a first semantic vector;

[0074] S4: forming a key-value pair based on the product name and the first semantic vector of each product;

[0075] S5: mining association information between products in the preset structural field through a preset GNN network;

[0076] S6: importing the association information between products and the corresponding key-value pair into a preset graph database to obtain a product knowledge graph.

[0077] As described in step S1, the product name and unstructured text data of each product are acquired. The product name is a unique identifier for each product, which is used to distinguish different products and is also a basic element for constructing a knowledge graph. The unstructured text data can come from multiple channels, such as product manuals, user reviews, social media discussions, and e-commerce platform product descriptions, which can provide multi-dimensional understanding of product performance, characteristics, usage, and user feedback. In a specific embodiment, to improve efficiency and accuracy, automated scraping tools and API interfaces can be considered. This method can significantly reduce the need for manual intervention, especially when dealing with large amounts of data. Meanwhile, the integration of product names and unstructured text data provides basic data for subsequent analysis, making the data processing in the process of constructing a knowledge graph have sufficient basis.

[0078] As described in step S2 above, the unstructured text data is fused to obtain preset structured fields. After obtaining the product name and unstructured text data, the unstructured text data is fused to convert these relatively messy information into one or more usable structured table formats, which usually includes data cleaning, information extraction and feature selection. Specifically, natural language processing (NLP) technology can be used to analyze key features from unstructured text and extract some preset structured fields, such as material, function, application scenario, scope of application, user feedback, etc. These fields are helpful for further analysis and association. In the process of fusion processing, fusion can be achieved through named entity recognition (NER) and keyword extraction, so that important product-related information in the text can be identified and classified to form consistent structured fields, which not only reduces the irregularity of the data, but also lays the foundation for semantic analysis and spatial mapping in subsequent steps. At the same time, the processed structured fields help improve the quality and consistency of the data and provide necessary support for building a knowledge graph.

[0079] As described in step S3 above, the preset structural fields are spatially mapped to obtain a first semantic vector. The preset structural fields are spatially mapped to obtain a first semantic vector. Spatial mapping mainly refers to converting actual text, features or attributes into vector form so that the computer can understand and process them. This process usually uses a word embedding model (such as Word2Vec, GloVe or more advanced BERT, GPT, etc.) to map text data into a high-dimensional vector space. By mapping each structural field to a semantic vector, the semantic relationship between texts can be more intuitively represented, focusing on the contextual information reflected by each field. The corresponding high-dimensional vector can capture the semantic features and relationship information of the vocabulary. It will help the subsequent analysis process, such as strengthening the similarity calculation and relationship mining between products. By spatially mapping the structural fields, a set of vectors with rich semantic information can be formed, providing strong data support for the construction of the global knowledge graph.

[0080] As described in step S4 above, a key-value pair is formed based on the product name and the first semantic vector of each product. Each collected product name is combined with its corresponding first semantic vector to form a key-value pair. In the construction of the knowledge graph, the product name serves as the "key", and the mapped first semantic vector serves as its corresponding "value". Such a structural configuration enables the basic information of the product to be efficiently matched with its semantic features, and the establishment of key-value pairs can facilitate subsequent data query and retrieval. Through such a matching method, the relevant information of a product in the knowledge graph can be quickly located, which facilitates querying and comparing the similarities between products. At the same time, the formation of key-value pairs makes the data structured and easy to store and manage. In the context of constructing the knowledge graph, such a structure can help users access and analyze data more conveniently, and provide direct support for subsequent knowledge reasoning and intelligent recommendations.

[0081] As described in step S5 above, a preset GNN network is used to mine the association information between products in the preset structured fields. After the key-value pairs are formed, a preset graph neural network (GNN) is used to mine the association information between products in the preset structured fields. GNN is a deep learning model specifically for graph-structured data. It effectively captures the relationships and mutual influences between nodes (products). GNNs perform deep learning on complex relationships between products, analyzing their potential similarities and connections. Once a product knowledge graph is constructed, each product can be viewed as a node in the graph, with the associations as edges between nodes. By learning these nodes and their connections, GNNs can understand the relationship between any two nodes. Notably, GNNs can leverage information between nodes and their neighboring nodes for information transfer, thereby optimizing the mining of product relationships. This approach not only reveals explicit relationships but also uncovers potential implicit associations. Through this deep mining, companies can gain a more comprehensive understanding of the associations between products, making informed recommendations based on this information and enhancing their market competitiveness. At the same time, because GNN has excellent processing capabilities and flexibility, it can perform well in product graphs of different scales and structures, ensuring that the constructed knowledge graph has high practicality and accuracy.

[0082] As described in step S6 above, the association information between each product and the corresponding key-value pairs are imported into a preset graph database to obtain a product knowledge graph. The association information between each product and the corresponding key-value pairs obtained in this manner are imported into the preset graph database, ultimately forming a product knowledge graph. A graph database is a database specifically designed to process graph data models, enabling efficient storage and querying of nodes (products) and their relationships. By importing products, their attributes, and association information into a graph database, enterprises can achieve efficient knowledge management and fast dynamic querying. The preset graph database can be any of Neo4j, ArangoDB, and Amazon Neptune. During the import process, it is necessary to ensure consistency between the data format and the database model to ensure smooth data storage and retrieval. At this point, different types of data (e.g., node data, edge data, etc.) are integrated into the graph database to provide a comprehensive knowledge graph structure. The product knowledge graph not only displays detailed information about each product at the entity level, but also reveals connections between products at the relationship level. This dual-layered knowledge graph provides important support for subsequent intelligent query, data mining, and recommendation systems, helping enterprises quickly respond to market demand, optimize product portfolios, and enhance user experience.

[0083] In one embodiment, the step S2 of fusing the unstructured text data to obtain a preset structured field includes:

[0084] S201: Extracting total preset core attributes of the unstructured text data of each product;

[0085] S202: Analyze the preset core attributes through dependency syntax to construct a relationship chain for each of the products;

[0086] S203: Generate a preset structural field based on the relationship chain.

[0087] As described in steps S201-S203 above, the core attributes of each product are extracted from the collected unstructured text data. The key to this step is identifying and defining the core attributes required for a product. These core attributes are typically pre-determined based on prior research and experience, such as materials, specifications, functions, uses, and applicable scenarios. Extracting core attributes typically involves natural language processing (NLP) techniques, such as vocabulary identification and entity recognition. Entity recognition identifies important product-related information from text and categorizes it into different attributes. To this end, models such as named entity recognition (NER), keyword extraction algorithms, and deep learning-based text analysis models can be used. These technologies can automatically process large amounts of text data, reducing the need for human intervention. Dependency parsing techniques are used to understand and construct the relationship chains between individual products. The core of dependency parsing is to identify the dependencies between words in a sentence to understand its structure and grammar. This analysis not only focuses on the part of speech of words but also explores the functional relationships between words, including subject-verb relationships and relationships between attributives and central words. For the extracted core attributes, dependency parsing can help identify the relationships between them, building relationship chains such as "material → processing technology → application scenario." For example, if the core attributes are "aluminum" and "milling," dependency parsing can identify the relationship between these two concepts and determine how they are logically connected. To build a relationship chain, a dependency tree or graph must be processed, extracting edges (relationships) from the nodes (core attributes) and organizing them into a meaningful chain. This step results in a series of relationship chains based on core attributes, providing semantic association information for subsequent steps and ensuring a comprehensive representation of the relationships between product features and attributes. By systematizing the core attributes of each relationship chain and their relationships, a unified structural field format can be generated, such as a field name and value representation. Common structural fields include product name, material, processing technology, and application scenario, each of which will be matched to the extracted value. For example, if a relationship chain shows the relationship between "aluminum" and "milling," a field can be defined with "Material" set to "Aluminum" and "Processing Technology" set to "Milling." The generated structured fields lay the foundation for subsequent knowledge graph construction and data analysis, making it easier for machines to manipulate and understand this information. This step ensures the effective conversion of unstructured data into structured data, while reducing the risk of information loss and enhancing the data's value. The resulting structured fields will serve as the core component of the knowledge graph, supporting further analysis and association mining between products.

[0088] In one embodiment, the step S4 of forming a key-value pair based on the product name of each product and the first semantic vector includes:

[0089] S401: Acquire the ERP field of each product, wherein the structure description of the ERP field is the same as the preset structured field;

[0090] S402: Performing spatial mapping on the ERP field to obtain a second semantic vector;

[0091] S403: Forming a key-value pair based on the product name, the first semantic vector, and the second semantic vector of each product.

[0092] As described in steps S401-S403, relevant field data is extracted from the enterprise resource planning (ERP) system, which are components of structured representation of product information, with the same structure description as the previously defined preset structural fields. In order to obtain the ERP fields of each product, it is necessary to first define which fields are the most critical, which usually include product number, name, description, category, material, processing technology, etc. When obtaining these data, the database query language (such as SQL) can be used to extract the relevant fields from the ERP system as ERP fields. In addition, when obtaining the ERP fields, special attention should be paid to the integrity of the fields to ensure that each product has at least all the required fields. Only by accurately extracting and integrating these ERP fields can a consistent data basis be provided for the subsequent formation of key-value pairs. The ERP fields are spatially mapped to obtain a second semantic vector. The process of spatial mapping is to convert specific structured data into vector representation with semantic meaning, so that computers can effectively understand and process this information. Corresponding to the first semantic vector, the second semantic vector will be able to reflect the information characteristics of the ERP fields. Specifically, pre-trained models such as Word2Vec, GloVe, or more complex BERT can be used. By inputting the ERP field text into these models, the semantic representation of each field in high-dimensional space can be learned. For structured fields such as "material" or "processing method", if these fields will produce different understandings in different contexts, the model will be able to capture this well and generate appropriate vector representations. Ultimately, the second semantic vector generated through this mapping not only emphasizes the internal relationship of the entire data set, but also provides the necessary input for the subsequent formation of key-value pairs, making it possible to align the features between different data sources. By combining the product name, the first semantic vector and the second semantic vector, a key-value pair is formed, which integrates the representations of different data sources into a unified structure, facilitating storage, analysis and query in subsequent processes. Specifically, the product name will serve as the "key" (key), while the first semantic vector and the second semantic vector will jointly serve as the "value" (value). This form of structure provides an intuitive access method, which is helpful for subsequent rapid retrieval and comparison. For example, for a certain product, when the user queries, the corresponding semantic vector can be quickly located through the product name, and then the relevant information can be extracted and analyzed.

[0093] In some embodiments, since the first semantic vector and the second semantic vector may have repeated words, in a specific embodiment, the preset structural fields and the ERP fields can be pre-integrated, i.e. removing repeated words to obtain target fields, and then spatially mapping the target fields to obtain a third semantic vector. The third semantic vector and the product name form a key-value pair.

[0094] In one embodiment, before the step S5 of mining the association information between the products in the preset structural field using the preset GNN network, the method further includes:

[0095] S411: Obtain multiple basic categories under existing standards and use the basic categories as clustering cores;

[0096] S412: Clustering each product using the K-means algorithm on the basic clustering core to obtain a target clustering core and a cluster corresponding to each target clustering core;

[0097] S413: Screening the preset structural fields based on the clusters to obtain target preset structural fields corresponding to each cluster;

[0098] S414: Mining the association information between products in each cluster in the target preset structural field through a preset GNN network.

[0099] As described in steps S411-S414 above, multiple existing basic categories are collected and determined. Basic categories refer to fundamental classifications with common characteristics within a specific industry or product category, and are typically the standards used by companies for product management, classification, and analysis. To obtain these basic categories, reference can be made to industry standards, market research reports, relevant documents, or expert advice. For example, in the manufacturing industry, these basic categories may include "electronic products," "chemical products," "consumer goods," and so on. At this stage, it is crucial to ensure that the selected basic categories are accurate, comprehensive, and cover all product types. Factors such as industry dynamics and market demand fluctuations should also be considered. Basic categories will be considered core elements in cluster analysis. Specifically, each basic category can serve as a "cluster center" in subsequent clustering processes. For example, the basic categories are mapped into a pre-defined hyperspace, with the basic categories serving as the core. Cluster analysis of the collected products is performed using the K-means algorithm, a classic unsupervised learning algorithm that aims to divide data into K possible clusters by minimizing the distance between samples and their centroids. The clustering process divides products into groups based on the similarity of their features. When implementing K-means clustering, the number of clusters, K, must first be determined. This value can be based on previously determined basic categories. Guided by these basic categories, the algorithm assigns each product to the closest cluster core, iteratively updating the position of the cluster core until the product assignments cease to change or the set iteration limit is reached. Ultimately, the K-means algorithm outputs the target cluster core and its corresponding clusters. Each cluster contains products with similar features, facilitating subsequent analysis. The effectiveness of this step directly impacts the quality of the knowledge graph. By grouping products with strong similarities, it provides a more accurate basis for subsequent relationship mining and structural field screening. Pre-defined structural fields are then screened based on the previously determined clusters. By analyzing the characteristics and attributes of each cluster, target pre-defined structural fields suitable for that cluster can be identified. For example, if a cluster primarily consists of electronic products, relevant target structural fields might include "brand," "model," "technical parameters," and "usage occasion." This step not only helps simplify subsequent data processing but also improves the understanding and accuracy of product characteristics analysis in subsequent steps. When performing field screening, commonly used techniques may include correlation analysis and feature selection methods to ensure that the selected fields can fully reflect the common characteristics of products within the cluster. At the same time, through this process, the dimensionality of the data is reduced, relevant features are preferentially retained, and the association information between products in each cluster is mined through a preset graph neural network (GNN). The target preset structural fields after clustering serve as the input of the GNN for local association mining, using the powerful capabilities of the GNN to analyze complex product relationship networks to reveal potential correlations.During implementation, a graph structure representing the relationships between products within a cluster must be constructed. In this graph, nodes represent individual products, while edges represent various relationships between products, such as similarity, competition, and co-purchase. By leveraging the multi-level feature learning capabilities of GNNs, the characteristics of these nodes and their connections can be effectively captured. GNNs utilize information propagation mechanisms to aggregate information from a product's neighboring nodes. Through multiple rounds of iterative learning, they gain a deep understanding of the relationships between products. Common algorithms such as GCN (Graph Convolutional Network) and GAT (Graph Attention Network) can generate updated feature representations for each node, reflecting its relative position and importance within the cluster.

[0100] In one embodiment, after step S6 of importing the association information and corresponding key-value pairs between the products into a preset graph database to obtain a product knowledge graph, the method further includes:

[0101] S701: Determine whether a preset recommendation instruction of a specified user is triggered;

[0102] S702: If the preset recommendation instruction is triggered, obtain the user's product information of interest;

[0103] S703: Based on the product information of interest and the product knowledge graph, recommend similar products by associating the graph.

[0104] As described in steps S701-S703 above, it is determined whether the user's preset recommendation instructions have been triggered. These preset instructions can be rules set based on the user's access behavior, purchase history, search history, or other interactions. For example, when a user browses a product, the system may be designed to detect whether the user's browsing time meets certain triggering conditions. When these conditions are met, the system should be able to automatically identify and confirm that the recommendation instruction has been triggered. Furthermore, triggering conditions can include not only simple events, such as visiting a product page, but also more complex scenarios, such as users frequently searching for a certain product category during a specific time period or changes based on past behavior patterns. To monitor these conditions, technologies such as user behavior analysis and real-time data stream processing can be utilized to ensure timely capture of user needs. This allows for flexible response to changes in user behavior and timely provision of relevant product recommendations and prompts. This improves user experience and drives conversions, as personalized recommendations can make users feel understood and valued, thereby increasing satisfaction and purchase intention. When the preset recommendation instruction is triggered, information about the user's product interests is obtained. The user's product interests are typically based on data such as their historical behavior, preferences, and profile. The system analyzes this information using various dimensions to accurately capture the user's interests. Methods for acquiring user interests include, but are not limited to, several approaches: First, analyzing users' recent browsing history to understand the types of products they're currently interested in; second, combining their purchase history to identify products they tend to purchase; and third, leveraging user-provided preference data, such as completed surveys or preferences set within the app. The user interest data acquired at this stage is typically structured, such as product IDs, categories, or brands of interest. The system must ensure that user data from multiple channels is integrated to provide a comprehensive view, enabling deeper analysis of user interests. Furthermore, based on user interest information, subsequent recommendations can more accurately focus on products that meet user needs, thereby improving the effectiveness and satisfaction of the recommendation system. In some embodiments, product information of interest can also be the content the user is currently viewing. Based on the acquired product information of user interest and the product knowledge graph, related recommendations of similar products are made. The knowledge graph's rich product features and relationship network effectively identify similarities with products of interest. Specifically, the system utilizes characteristics of products of interest (such as category, brand, material, and application scenario) to match them with target products in the knowledge graph. Based on algorithms such as graph neural networks (GNN), it is possible to search for products in the graph that are adjacent to or have similar connections with the products that the user is interested in, and then discover a variety of recommendation options including similar products, alternative products, or related co-purchased products. This can not only improve user satisfaction, but also promote sales and increase conversion rates.This recommendation method, which combines user interests with product knowledge graphs, not only achieves personalization and intelligence, but also enhances the overall shopping experience and strengthens user stickiness.

[0105] In one embodiment, after step S6 of importing the association information and corresponding key-value pairs between the products into a preset graph database to obtain a product knowledge graph, the method further includes:

[0106] S711: Obtain the target product name and target unstructured text data of the product to be added;

[0107] S712: Performing fusion processing on the target unstructured text data to obtain a target preset structured field;

[0108] S713: Performing spatial mapping on the target preset structural field to obtain a target first semantic vector;

[0109] S714: Forming a target key-value pair based on the target product name and the target first semantic vector;

[0110] S715: mining target association information between the product to be added and other products in the product knowledge graph in the target preset structural field through a preset GNN network;

[0111] S716: Import the target key-value pair and the target association information into the product knowledge graph to obtain an updated product knowledge graph.

[0112] As described in steps S711-S716 above, the product knowledge graph is updated, the target product name and its related unstructured text data are obtained, and the new product is integrated to ensure its effective inclusion in the product knowledge graph. The target product's unstructured text data typically comes from a variety of sources, such as product manuals, user feedback, marketing materials, social media comments, or product reviews. This data contains important information about the target product, such as its functions, features, usage, and user experience, helping the system build a comprehensive understanding of the product. The target unstructured text data is fused and processed to extract pre-defined structured fields. This unstructured data is then converted into a fixed format for storage and management in the product knowledge graph. Pre-defined structured fields typically include key descriptive information such as product name, material, specifications, functions, and applicable scenarios. Natural language processing (NLP) techniques can be used to analyze potential key information in unstructured text through methods such as keyword extraction, named entity recognition, and text parsing. Spatial mapping involves converting useful text information into a high-dimensional, dense vector representation, enabling computers to better understand and process this information. In NLP, the mapping process is an efficient way to digitize text features. Common spatial mapping techniques include various word embedding methods such as Word2Vec and GloVe, as well as more complex context-free models such as BERT. The previously generated target first semantic vector is paired with the product name to form a structured data object. This object is conveniently stored in a graph database to support subsequent querying and analysis. A pre-defined graph neural network (GNN) is used to mine target association information between the target product and other products in the existing product knowledge graph. First, the system adds the target pre-defined structural fields of the newly added product as a node to the knowledge graph. Other related product nodes and their attributes are also included, forming a larger graph structure. In this graph, the target product is connected to other existing products through various relationships and edges. The GNN learns nodes in the graph through an information propagation mechanism. During this process, association information is transferred to the target product node. Through multiple iterations, the GNN updates the node's features, enhancing its understanding of the characteristics of surrounding nodes. This allows it to identify potential associations, such as products that could be cross-sold with the target product or that offer complementary functionality. Import the target key-value pairs generated earlier and the target association information mined by GNN into the existing product knowledge graph to form an updated graph. This step is a crucial part of the entire process, as it ensures that the relevant information of the new product can be seamlessly integrated into the system and connects all related product data.

[0113] In one embodiment, after the step S6 of obtaining the product name and unstructured text data of each product, the method further includes:

[0114] S721: Determine whether the preset recommendation instruction of the specified user is triggered;

[0115] S722: If the preset recommendation instruction is triggered, obtain the user's product information of interest;

[0116] S723: Based on the product information of interest and the product knowledge graph, similar products are recommended based on the similarity of semantic vectors.

[0117] As described in steps S701-S703 above, it is determined whether the user's preset recommendation instructions have been triggered. These preset instructions can be rules set based on the user's access behavior, purchase history, search history, or other interactions. For example, when a user browses a product, the system may be designed to detect whether the user's browsing time meets certain triggering conditions. When these conditions are met, the system should be able to automatically identify and confirm that the recommendation instruction has been triggered. Furthermore, triggering conditions can include not only simple events, such as visiting a product page, but also more complex scenarios, such as users frequently searching for a certain product category during a specific time period or changes based on past behavior patterns. To monitor these conditions, technologies such as user behavior analysis and real-time data stream processing can be utilized to ensure timely capture of user needs. This allows for flexible response to changes in user behavior and timely provision of relevant product recommendations and prompts. This improves user experience and drives conversions, as personalized recommendations can make users feel understood and valued, thereby increasing satisfaction and purchase intention. When the preset recommendation instruction is triggered, information about the user's product interests is obtained. The user's product interests are typically based on data such as their historical behavior, preferences, and profile. The system analyzes this information using various dimensions to accurately capture the user's interests. Methods for acquiring user interests include, but are not limited to, several approaches: First, analyzing users' recent browsing history to understand the types of products they're currently interested in; second, combining their purchase history to identify products they tend to purchase; and third, leveraging user-provided preference data, such as completed surveys or preferences set within the app. The user interest data acquired at this stage is typically structured, such as the IDs, categories, or brands of products of interest. The system must ensure that user data from multiple channels is integrated to provide a comprehensive view, enabling deeper analysis of user interests. Furthermore, based on user interest information, subsequent recommendations can more accurately focus on products that meet user needs, thereby improving the effectiveness and satisfaction of the recommendation system. In some embodiments, product information of interest can also refer to the content the user is currently viewing. Products of interest are converted into their corresponding semantic vectors, which are derived from the product knowledge graph. The product knowledge graph maintains the relationships and feature information between products, as well as their corresponding semantic vectors, allowing the similarity between different products to be determined by calculating their vector representations. Common similarity calculation methods include cosine similarity and Euclidean distance. Once the semantic vector of the product of interest is calculated, the system compares it with the semantic vectors of other products to identify the product with the highest similarity to the user's product of interest. In this way, new product recommendations can be discovered, helping users discover products they may be interested in but have not yet browsed.This semantic similarity-based recommendation method significantly improves the system's intelligence and accuracy, ensuring that users receive product recommendations that are highly relevant to their preferences during browsing, thereby increasing user satisfaction and conversion rates. Ultimately, personalized recommendations increase user purchase intent, driving sales growth.

[0118] In a specific embodiment, the first semantic vector, the second semantic vector, and the third semantic vector are all 768-dimensional semantic vectors, where a 768-dimensional semantic vector refers to a high-dimensional dense vector used to represent text (such as words, sentences, or paragraphs).

[0119] Figure 4 FIG1 shows an internal structure diagram of a computer device in an embodiment. The computer device can be a terminal or a server. Figure 4 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor may implement a method for constructing a product knowledge graph. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor may implement a method for constructing a product knowledge graph. Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0120] Reference Figure 3 The present invention also provides a device for constructing a product knowledge graph, the device comprising:

[0121] An acquisition module 902 is used to acquire the product name and unstructured text data of each product;

[0122] A fusion module 904 is used to perform fusion processing on the unstructured text data to obtain a preset structured field;

[0123] A mapping module 906 is configured to perform spatial mapping on the preset structural field to obtain a first semantic vector;

[0124] A forming module 908 is configured to form a key-value pair based on the product name of each product and the first semantic vector;

[0125] A mining module 910 is configured to mine association information between products in the preset structural field using a preset GNN network;

[0126] The import module 912 is used to import the association information between each product and the corresponding key-value pairs into a preset graph database to obtain a product knowledge graph.

[0127] In one embodiment, the fusion module 904 includes:

[0128] A preset core attribute extraction submodule, configured to extract the total preset core attributes of the unstructured text data of each product;

[0129] A preset core attribute analysis submodule, configured to analyze the preset core attributes through dependency syntax to construct a relationship chain of each of the products;

[0130] The preset structural field generation submodule is used to generate the preset structural field based on the relationship chain.

[0131] In one embodiment, forming module 908 includes:

[0132] An ERP field acquisition submodule, configured to acquire an ERP field for each product, wherein the ERP field has the same structure description as the preset structured field;

[0133] an ERP field mapping submodule, configured to perform spatial mapping on the ERP field to obtain a second semantic vector;

[0134] The key-value pair forming submodule is configured to form a key-value pair based on the product name, the first semantic vector, and the second semantic vector of each product.

[0135] In one embodiment, the device for constructing a product knowledge graph further includes:

[0136] Basic category acquisition module, used to obtain multiple basic categories under existing standards and use the basic categories as clustering cores;

[0137] A clustering module is used to cluster each product using the K-means algorithm on the basic clustering core to obtain a target clustering core and a cluster cluster corresponding to each target clustering core;

[0138] A screening module, configured to screen the preset structural fields based on the clusters to obtain a target preset structural field corresponding to each cluster;

[0139] The association information mining module is used to mine the association information between products in each cluster in the target preset structural field through a preset GNN network.

[0140] In one embodiment, the device for constructing a product knowledge graph further includes:

[0141] A preset recommendation instruction first determination module is used to determine whether the preset recommendation instruction of a specified user is triggered;

[0142] a first product information acquisition module for acquiring the user's product information of interest if the preset recommendation instruction is triggered;

[0143] The first similar product recommendation module is used to recommend similar products based on the association of the interest product information and the product knowledge graph.

[0144] In one embodiment, the device for constructing a product knowledge graph further includes:

[0145] A target unstructured text data acquisition module is used to acquire the target product name and target unstructured text data of the product to be added;

[0146] A target unstructured text data fusion module is used to perform fusion processing on the target unstructured text data to obtain a target preset structured field;

[0147] A target preset structural field mapping module is used to perform spatial mapping on the target preset structural field to obtain a target first semantic vector;

[0148] a target key-value pair forming module, configured to form a target key-value pair based on the target product name and the target first semantic vector;

[0149] A target association information mining module is used to mine the target association information between the product to be added and other products in the product knowledge graph in the target preset structural field through a preset GNN network;

[0150] Import the target key-value pairs and the target association information into the product knowledge graph to obtain an updated product knowledge graph.

[0151] In one embodiment, after the step of obtaining the product name and unstructured text data of each product, the method further includes:

[0152] A second preset recommendation instruction determination module is used to determine whether the preset recommendation instruction of a specified user is triggered;

[0153] A second module for obtaining product information of interest, configured to obtain product information of interest to the user if the preset recommendation instruction is triggered;

[0154] The second similar product recommendation module is used to recommend similar products based on the similarity of semantic vectors based on the information of the product of interest and the product knowledge graph.

[0155] In one embodiment, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0156] Get the product name and unstructured text data for each product;

[0157] Performing fusion processing on the unstructured text data to obtain preset structured fields;

[0158] Performing spatial mapping on the preset structural field to obtain a first semantic vector;

[0159] Based on the product name of each product and the first semantic vector, a key-value pair is formed;

[0160] Mining the association information between the products in the preset structural fields through a preset GNN network;

[0161] The association information between each product and the corresponding key-value pairs are imported into a preset graph database to obtain a product knowledge graph.

[0162] It realizes the integrated analysis of data, improves the efficiency of building knowledge graphs, and provides important support for subsequent intelligent query, data mining and recommendation systems, thereby helping enterprises to quickly respond to market demand, optimize product portfolios and improve user experience.

[0163] In one embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the processor performs the following steps:

[0164] Get the product name and unstructured text data for each product;

[0165] Performing fusion processing on the unstructured text data to obtain preset structured fields;

[0166] Performing spatial mapping on the preset structural field to obtain a first semantic vector;

[0167] Based on the product name of each product and the first semantic vector, a key-value pair is formed;

[0168] Mining the association information between the products in the preset structural fields through a preset GNN network;

[0169] The association information between each product and the corresponding key-value pairs are imported into a preset graph database to obtain a product knowledge graph.

[0170] It realizes the integrated analysis of data, improves the efficiency of building knowledge graphs, and provides important support for subsequent intelligent query, data mining and recommendation systems, thereby helping enterprises to quickly respond to market demand, optimize product portfolios and improve user experience.

[0171] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0172] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0173] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for constructing a product knowledge graph, characterized in that: The method comprises: Get the product name and unstructured text data for each product; Performing fusion processing on the unstructured text data to obtain preset structured fields; Performing spatial mapping on the preset structural field to obtain a first semantic vector; Based on the product name of each product and the first semantic vector, a key-value pair is formed; Mining the association information between the products in the preset structural fields through a preset GNN network; Import the association information between each product and the corresponding key-value pairs into a preset graph database to obtain a product knowledge graph; The step of fusing the unstructured text data to obtain preset structured fields includes: Extracting total preset core attributes of the unstructured text data of each product; Analyzing the preset core attributes through dependency syntax to construct a relationship chain of each of the products; generating a preset structural field based on the relationship chain; The step of forming a key-value pair based on the product name of each product and the first semantic vector includes: Obtaining an ERP field for each product, wherein the ERP field has the same structure description as the preset structured field; Performing spatial mapping on the ERP field to obtain a second semantic vector; Based on the product name, the first semantic vector, and the second semantic vector of each product, a key-value pair is formed; Before the step of mining the association information between the products in the preset structural field through the preset GNN network, the method further includes: Obtain multiple basic categories under existing standards and use the basic categories as the clustering core; Clustering each product on the cluster core using the K-means algorithm to obtain a target cluster core and a cluster cluster corresponding to each target cluster core; Screening the preset structural fields based on the clusters to obtain a target preset structural field corresponding to each cluster; The preset GNN network is used to mine the association information between products in each cluster in the target preset structural field.

2. The method for constructing a product knowledge graph according to claim 1, characterized in that: After the step of importing the association information and corresponding key-value pairs between the products into a preset graph database to obtain a product knowledge graph, the method further includes: Determine whether the preset recommendation instruction of the specified user is triggered; If the preset recommendation instruction is triggered, obtain the user's product information of interest; Based on the product information of interest and the product knowledge graph, similar products are recommended by associating the graphs.

3. The method for constructing a product knowledge graph according to claim 1, characterized in that: After the step of importing the association information and corresponding key-value pairs between the products into a preset graph database to obtain a product knowledge graph, the method further includes: Obtain the target product name and target unstructured text data of the product to be added; Performing fusion processing on the target unstructured text data to obtain target preset structured fields; Performing spatial mapping on the target preset structural field to obtain a first semantic vector of the target; Based on the target product name and the target first semantic vector, forming a target key-value pair; Mining the target association information between the product to be added and other products in the product knowledge graph in the target preset structural field through a preset GNN network; Import the target key-value pairs and the target association information into the product knowledge graph to obtain an updated product knowledge graph.

4. The method for constructing a product knowledge graph according to claim 1, characterized in that: After the step of obtaining the product name and unstructured text data of each product, the method further includes: Determine whether the preset recommendation instruction of the specified user is triggered; If the preset recommendation instruction is triggered, obtain the user's product information of interest; Based on the product information of interest and the product knowledge graph, similar products are recommended based on the similarity of semantic vectors.

5. A device for constructing a product knowledge graph, characterized in that: The device comprises: The acquisition module is used to obtain the product name and unstructured text data of each product; A fusion module, configured to perform fusion processing on the unstructured text data to obtain a preset structured field; A mapping module, configured to perform spatial mapping on the preset structural field to obtain a first semantic vector; a forming module, configured to form a key-value pair based on the product name of each product and the first semantic vector; A mining module, configured to mine the association information between the products in the preset structural fields through a preset GNN network; An import module is used to import the association information between each product and the corresponding key-value pairs into a preset graph database to obtain a product knowledge graph; The fusion module includes: A preset core attribute extraction submodule, configured to extract the total preset core attributes of the unstructured text data of each product; A preset core attribute analysis submodule, configured to analyze the preset core attributes through dependency syntax to construct a relationship chain of each of the products; A preset structural field generation submodule, configured to generate a preset structural field based on the relationship chain; The formation module includes: An ERP field acquisition submodule, configured to acquire an ERP field for each product, wherein the ERP field has the same structure description as the preset structured field; an ERP field mapping submodule, configured to perform spatial mapping on the ERP field to obtain a second semantic vector; a key-value pair forming submodule, configured to form a key-value pair based on the product name, the first semantic vector, and the second semantic vector of each product; Basic category acquisition module, used to obtain multiple basic categories under existing standards and use the basic categories as clustering cores; A clustering module is used to cluster each product on the cluster core using the K-means algorithm to obtain a target cluster core and a cluster cluster corresponding to each target cluster core; A screening module, configured to screen the preset structural fields based on the clusters to obtain a target preset structural field corresponding to each cluster; The association information mining module is used to mine the association information between products in each cluster in the target preset structural field through a preset GNN network.

6. A computer-readable storage medium, characterized in that A computer program is stored, and when the computer program is executed by a processor, the processor performs the steps of the method for constructing a product knowledge graph as described in any one of claims 1 to 4.

7. A computer device, characterized in that: The device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method for constructing a product knowledge graph as described in any one of claims 1 to 4.

Images