Multi-level mixed search method and device, equipment and medium

Abstract

The invention relates to a multi-level mixed search method and device, equipment and a medium, and the method comprises the steps: obtaining a first hierarchical navigation chart which is used for representing the proximity relation between corresponding nodes of commodities in a plurality of online shops of an e-commerce platform, each node comprises a high-dimensional semantic vector of the corresponding commodity and an attribute scalar which indicates the attribution category of the corresponding commodity in the e-commerce platform hierarchical classification system; corresponding to each attribution category indicated by the attribute scalar, respectively constructing a second hierarchical navigation map which corresponds to each attribution category and represents a proximity relationship between corresponding nodes of the commodity corresponding to the attribution category; responding to a commodity search event, and determining a query category and a query vector corresponding to query information carried by the event; and calling the second hierarchical navigation map with the same attribution category as the query category as a target navigation map, and retrieving nodes matched with the query vector semantics from the target navigation map. According to the invention, the search efficiency and accuracy of the e-commerce platform for providing the commodity search service can be improved.

Description

Technical Field

[0001] The present application relates to e-commerce information technology, and in particular to a multi-level hybrid search method and its device, equipment, and medium. Background Art

[0002] With the rapid development of e-commerce, e-commerce platforms need to provide centralized product search services for their multiple online stores. The core challenge of this service is how to quickly and accurately retrieve the product information required by users from massive data. These data include both image-type product images and text-type product descriptions, or a combination of the two. In order to achieve efficient product search, it is usually necessary to extract high-dimensional semantic vectors based on these data and combine them with corresponding algorithms for semantic matching to obtain search results.

[0003] In the prior art, algorithms for implementing semantic matching mainly use KNN (K Nearest Neighbors) or ANN (Approximate Nearest Neighbor) methods. The KNN algorithm finds the k nearest neighbors to a query vector by calculating the distance between vectors, but as the amount of data increases, its computational complexity grows quadratically, resulting in low efficiency. As an approximate substitute for KNN, the ANN algorithm improves search efficiency by building an index, allowing approximate nearest neighbors to be found quickly at the expense of a certain degree of accuracy. As an efficient ANN implementation, HNSW (Hierarchical Navigable Small World Graph) accelerates the retrieval process by building a hierarchical graph structure, significantly improving the retrieval efficiency on large-scale data sets.

[0004] However, although the HNSW algorithm performs well in processing vector data, it is insufficient in processing scalar data. In practical applications, users rely not only on vector data (such as high-dimensional semantic vectors based on product images) but also on scalar data (such as store logos) when searching for products. Traditional HNSW technology does not have the ability to process such scalar data, which limits its application in product search services on e-commerce platforms.

[0005] Specifically, the traditional HNSW algorithm faces two major challenges when processing scalar data: pre-filtering and post-filtering. The pre-filtering method first filters out qualified products based on scalar data, and then performs vector retrieval on these products. This method cannot fully utilize the speed advantage of the ANN algorithm. The post-filtering method first performs ANN retrieval and then filters the retrieval results by scalar data. This method requires retrieving more products in the first step, and due to the existence of post-filtering, it will have a negative impact on the recall rate of the retrieval. Both methods fail to effectively combine vector and scalar data, resulting in a compromise between search efficiency and accuracy.

[0006] Therefore, the product search service based on the traditional HNSW algorithm has some obvious shortcomings, including but not limited to: low retrieval efficiency, low recall rate, and inability to simultaneously take into account the filtering requirements of vector and scalar data. These problems limit the ability of e-commerce platforms to concurrently provide fast and accurate product search services at all levels within their hierarchical classification system, especially when faced with massive data and complex query requirements involving scalars and vectors. Summary of the invention

[0007] The purpose of the present application is to solve the above-mentioned problems and to provide a multi-level hybrid search method and its corresponding device, equipment, non-volatile readable storage medium, and computer program product.

[0008] According to one aspect of the present application, a multi-level hybrid search method is provided, comprising:

[0009] Obtaining a first hierarchical navigation graph for representing proximity relationships between corresponding nodes of products in a plurality of online stores on an e-commerce platform, wherein the nodes include high-dimensional semantic vectors for describing corresponding products and attribute scalars for indicating the categories to which the corresponding products belong in a hierarchical classification system of the e-commerce platform;

[0010] Corresponding to each attribution category indicated by the attribute scalar, respectively constructing a second hierarchical navigation graph corresponding to each attribution category and indicating the neighbor relationship between the corresponding nodes of the commodities corresponding to the attribution category;

[0011] In response to a product search event, determine the query category and query vector corresponding to the query information carried by the event;

[0012] A second hierarchical navigation graph with the same belonging category as the query category is called as a target navigation graph, and a corresponding node of the commodity that matches the semantics of the query vector is retrieved from the target navigation graph.

[0013] According to another aspect of the present application, a multi-level hybrid search device is provided, comprising:

[0014] a base graph calling module, configured to obtain a first hierarchical navigation graph for representing proximity relationships between corresponding nodes of products in a plurality of online stores on the e-commerce platform, wherein the nodes include high-dimensional semantic vectors for describing the corresponding products and attribute scalars for indicating the categories to which the corresponding products belong in the hierarchical classification system of the e-commerce platform;

[0015] A sub-graph construction module is set to each attribution category indicated by the corresponding attribute scalar, and constructs a second hierarchical navigation graph corresponding to each attribution category, which represents the neighbor relationship between the corresponding nodes of the commodities corresponding to the attribution category;

[0016] An event response module, configured to respond to a product search event and determine a query category and a query vector corresponding to the query information carried by the event;

[0017] The response execution module is configured to call a second hierarchical navigation graph with the same belonging category as the query category as a target navigation graph, and retrieve corresponding nodes of commodities that match the semantics of the query vector from the target navigation graph.

[0018] According to another aspect of the present application, a multi-level hybrid search device is provided, comprising a central processing unit and a memory, wherein the central processing unit is used to call and run a computer program stored in the memory to execute the steps of the method described in the present application.

[0019] According to another aspect of the present application, a non-volatile readable storage medium is provided, which stores a computer program implemented according to the multi-level hybrid search method in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the method are executed.

[0020] According to another aspect of the present application, a computer program product is provided, comprising a computer program / instruction, wherein the computer program / instruction implements the steps of the method when executed by a processor.

[0021] Compared with the prior art, this application effectively solves the challenges faced by e-commerce platforms in processing massive commodity data through its innovative technical solutions, especially in achieving multi-level mixed commodity searches. It not only improves the flexibility of search and the ability to handle complex queries, but also significantly improves retrieval efficiency and accuracy. The technical solution of this application is particularly suitable for scenarios such as e-commerce platforms where there are a large number of online stores and massive commodity data. It can ensure that when users search for commodities, whether based on images, text, or a combination of the two, relevant results can be returned quickly and accurately to meet users' needs for comprehensive search results. In this way, this application not only optimizes the user's shopping experience, but also provides an efficient commodity retrieval solution for e-commerce platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This application is an exemplary network architecture that can be used to deploy and implement product search services;

[0023] Figure 2 This is an example of the graph structure of the first hierarchical navigation graph shown in this application;

[0024] Figure 3 This is an example of a graph structure of a fusion graph obtained by merging the first layered navigation graph and the second layered navigation graph of the present application;

[0025] Figure 4 A flowchart of an embodiment of a multi-level hybrid search method of the present application;

[0026] Figure 5 This is a principle block diagram of the multi-level hybrid search device of the present application;

[0027] Figure 6 This is a schematic diagram of the structure of a multi-level hybrid search device used in this application. DETAILED DESCRIPTION

[0028] like Figure 1 As shown, the network architecture used by the present application to implement the product search service includes a front-end server 81 and a terminal device 80. The front-end server can be a search service server in the service cluster of the e-commerce platform, which undertakes the core function of providing product search services to the terminal device. The front-end server 81 runs a computer program product implemented according to the multi-level hybrid search method of the present application. The program product can process product search requests from terminal devices and return a relatively accurate product recommendation list based on the technical solution of the present application. The terminal device 80 is the interface used by users of the e-commerce platform to trigger product search events. Users input query information through these devices to start the search process.

[0029] Under the network architecture of the present application, when a user enters a search keyword and / or uploads a picture in the search page of an e-commerce platform or a specific online store to trigger a product search event, it can also carry or hide scalar data used to indicate the corresponding query category in the hierarchical classification system of the e-commerce platform. For example, when entering the search keyword, the user can specify the store logo of the online store as the query category, or enter the product category as the query category. The query category is essentially a scalar data, and a complex query is activated accordingly. The user's product search event is first transmitted to the front-end server, and the server immediately responds to this event request. According to the corresponding query category and query vector specified by the user in his query information, the corresponding first layered navigation graph and / or second layered navigation graph are called as the target navigation graph to perform a product search that matches the query vector. The target navigation graph reflects the proximity relationship between the corresponding nodes of each product, so that the server can quickly retrieve the product nodes that match the user's query vector in the target navigation graph.

[0030] In one embodiment, when the user does not provide a query category in the search keyword, if the user is in the state of visiting an online store, the store logo of the online store can be assumed as the query category specified by the user. In another embodiment, when the user does not provide a query category in the search keyword, and it is not possible to identify whether the user is visiting an online store, it can be assumed that the user is performing a product search at the level of the entire e-commerce platform.

[0031] The product search service of the present application specifically takes into account the hierarchical classification system architecture of the e-commerce platform. In an exemplary hierarchical classification system of an e-commerce platform, the e-commerce platform itself serves as the root level, and each of its online stores constitutes the next level, which is regarded as an attribution category. Each online store itself can also be regarded as an independent root level, and the first-level product category in the product classification system deployed in the online store forms the next level of the online store, which is regarded as the corresponding attribution category. By analogy, when there is a multi-level structure in the product classification system of the online store, the relationship between the parent level and the child level is the same as the affiliation between the e-commerce platform and the online store. This hierarchical structure makes the organization and retrieval of products more orderly, and also provides application flexibility for the technical solution of the present application. The technical solution of the present application can cross these levels to achieve multi-level product search from the e-commerce platform level to the store level and then to the product classification level, meeting the search needs of users at different levels.

[0032] In this application, the hierarchical navigable small-world graph algorithm is applied to organize and manage the product data in the hierarchical classification system of the e-commerce platform. By converting the product data in each online store of the e-commerce platform into the node data of the corresponding node, a first hierarchical navigation graph at the full platform level is constructed, such as Figure 2In the example of the graph structure shown, the graph depicts in detail the proximity relationship between the corresponding nodes of each product from A to F. On this basis, according to the categories of the products, such as different store logos, a second hierarchical navigation graph corresponding to each category is further constructed. These second hierarchical navigation graphs not only inherit the high-dimensional semantic vectors and attribute scalars of the first hierarchical navigation graph, but also form a more sophisticated navigation structure based on the specific attributes of the products in each store.

[0033] In the hierarchical navigation graph, product nodes are distributed to multiple sub-navigation layers, and the position of each node is determined by its high-dimensional semantic vector and attribute scalar. The edge connection relationship between nodes and the mapping relationship between different sub-navigation layers are determined based on the data distance between product nodes. This structured approach not only optimizes the organization of product data, but also provides a basis for performing efficient search.

[0034] In some embodiments, the second layered navigation graph can be merged into the first layered navigation graph to achieve data compression to reduce system storage and computing overhead, and the merged navigation graph contains not only the edge connection relationship of the first layered navigation graph, but also the edge connection relationship of the second layered navigation graph. Figure 3 In the example of the graph structure shown in Figure 2 On the basis of the edge connection relationship of node C, the edge connection relationship between node C and nodes D and E is also added to reflect their edge connection relationship in the second hierarchical navigation graph.

[0035] When a user triggers a complex query request on an e-commerce platform and provides a query category and a query vector, the product search service of the present application can apply an approximate nearest neighbor algorithm on the second layered navigation graph corresponding to the query category. This method is based on the query vector, and by quickly locating the product nodes that match the semantics of the query vector in the target navigation graph, the amount of data faced by the search is greatly reduced. Such a search strategy not only improves the search efficiency, but also improves the accuracy of the search results, especially when facing an e-commerce platform containing massive product data, this advantage is particularly obvious. In this way, the technical solution of the present application can provide users with a product search service experience that is both fast and accurate.

[0036] When the e-commerce platform is a cross-border e-commerce platform, it runs the corresponding online store services through independent sites. The independent sites, that is, the online stores, are discretely distributed on the network, and the data between them are usually not interoperable. However, the e-commerce platform itself, as the service provider of product search services, needs to realize centralized calling of products on the entire platform. When users search for products in various online stores, the e-commerce platform often needs to respond centrally. In this case, the technical solution of the present application can reflect the characteristics of taking into account the independence of online stores and the centralization of the entire platform, maximize the use of limited server resources, avoid discrete and repeated deployment, and provide balanced comprehensive search services.

[0037] See also Figure 4 According to a multi-level hybrid search method provided by the present application, it can be implemented as a computer program product installed in a server and run. After the computer program product runs, a commodity search system is constructed to provide corresponding commodity search services for users of the e-commerce platform. In some embodiments, the following steps are included:

[0038] Step S3100: obtaining a first hierarchical navigation graph for representing the proximity relationship between corresponding nodes of commodities in a plurality of online stores of the e-commerce platform, wherein the nodes include a high-dimensional semantic vector for describing the corresponding commodity and an attribute scalar for indicating the category to which the corresponding commodity belongs in the hierarchical classification system of the e-commerce platform;

[0039] In this embodiment, the products in the multiple online stores in the e-commerce platform are usually the products that have been put on the shelves. A first hierarchical navigation graph is pre-built and can be directly called in this step. The first hierarchical navigation graph is used to represent the proximity relationship between the product nodes of the multiple online stores in the e-commerce platform.

[0040] When constructing the first hierarchical navigation graph, we first collect the product data corresponding to the products of each online store in the e-commerce platform. These product data include not only the image and description text of the product, but also the attribute scalar indicating the category to which the product belongs in the hierarchical classification system of the e-commerce platform. The attribute scalar is quantitative information used to distinguish the hierarchical category to which the product belongs. For example, it can be a store ID or a product category identifier. This information helps to determine the specific location of the product in the hierarchical classification structure of the e-commerce platform. In order to reflect the advantages of structured storage, the product data is converted into corresponding node data according to the data structure of the nodes in the hierarchical navigation graph.

[0041] Next, feature extraction is performed on the data of each product to generate a high-dimensional semantic vector. In this application, a pre-trained deep learning model can be used to extract features of product images and / or features of product description texts, depending on the amount and modality of the data content used to generate the high-dimensional semantic vector. The extracted features are integrated into a high-dimensional semantic vector that can capture the key semantic information of the product.

[0042] After obtaining the high-dimensional semantic vectors of each product, the hierarchical navigable small-world graph algorithm is used to convert and store these high-dimensional semantic vectors into the node data of the corresponding node. The hierarchical navigable small-world graph algorithm determines the sub-navigation layer to which the node belongs and the edge connection relationship between the nodes based on the data distance between the high-dimensional semantic vectors of the nodes, thereby reflecting the similarity between the products in the graph structure.

[0043] Finally, by merging multiple sub-navigation layers, a first-layer navigation map covering all products on the e-commerce platform is constructed. The first-layer navigation map not only contains the high-dimensional semantic information of the products, but also integrates the attribute scalars of the products, providing a comprehensive, multi-dimensional navigation framework for subsequent searches.

[0044] At the data storage level, the hierarchical navigation graph of the present application adopts a unified and structured data storage solution, which can encapsulate all node-related data in a standardized node data structure. Specifically, the data structure of each node not only contains the corresponding attributes of the high-dimensional semantic vector of the product; it also includes the attribute scalar, which is one or a group of numerical values ​​that identify the specific position of the product in the hierarchical classification system of the e-commerce platform, such as store ID or product category; and edge connection relationship attributes, which define the proximity or similarity between nodes. Through these edge connection relationships, the hierarchical navigation graph can show the complex relationship network between products. This encapsulation method is not only applicable to the first hierarchical navigation graph of the present application, but also to the second hierarchical navigation graph and the graph structure after the fusion of the two, making data storage and retrieval more efficient, but also facilitating multi-dimensional analysis and processing of products, providing a solid data foundation for fast and accurate product search.

[0045] Step S3200: for each attribution category indicated by the corresponding attribute scalar, respectively construct a second hierarchical navigation graph corresponding to each attribution category, which indicates the neighbor relationship between the corresponding nodes of the commodities corresponding to the attribution category;

[0046] In an e-commerce platform, products are divided into different categories according to their attribute scalars, for example, into different stores or product categories. For each category, a specific second hierarchical navigation graph can be constructed to represent the proximity relationship between product nodes in the same category. A customized search view can be provided for each specific category, thereby optimizing the efficiency and accuracy of product searches within the category.

[0047] In specific implementation, the target attribution category for which the second hierarchical navigation graph needs to be constructed can be determined first according to the attribute scalar. The first hierarchical navigation graph is filtered according to the target attribution category to filter out the commodity nodes belonging to the target attribution category to form a target node set. For example, if the attribute scalar is a store ID of a specific store, all commodity corresponding nodes belonging to the specific store can be filtered out; if the attribute scalar is a specific commodity category in a specific online store, all commodity corresponding nodes belonging to the specific commodity category of the specific online store are filtered out to form a target node set.

[0048] Next, using the hierarchical navigable small-world graph algorithm, the sub-navigation layers to which these nodes belong and the edge connection relationships between the nodes are determined based on the data distance between the high-dimensional semantic vectors of the node data in the target node set. This ensures that in each second-layer navigation graph, the organizational structure of the product nodes can reflect their similarity in the vector space, so that the search operation can be performed more accurately between similar products.

[0049] Finally, by merging multiple sub-navigation layers, a second hierarchical navigation graph covering products of a specific category is constructed. The second hierarchical navigation graph not only inherits the node data of the first hierarchical navigation graph, including the high-dimensional semantic vectors and attribute scalars therein, but also forms a more sophisticated navigation structure based on the specific attribute data of the products in each category, including but not limited to any associated data such as product page links, product prices, and product inventory. For example, if an e-commerce platform contains multiple stores, each of which sells products of different categories, then each store can correspond to a second hierarchical navigation graph, which is specifically used for searching and retrieving products in the store.

[0050] Through a second hierarchical navigation graph customized for each category, a flexible and efficient way is provided to process and respond to product search requests for a specific category within a specific hierarchy. It enables the search service to quickly locate products in a specific store or product category while maintaining a comprehensive consideration of the high-dimensional semantic information and hierarchical structure of the product, thereby improving the efficiency and accuracy of the search.

[0051] It should be pointed out that, since the second layered navigation graph is reconstructed according to the corresponding attribution category, the edge connection relationship between the nodes is reset according to the high-dimensional semantic vector of the node, so there is no causal relationship between the edge connection relationship between the nodes in the second layered navigation graph and the edge connection relationship between the nodes in the first layered navigation graph. In one embodiment, the second layered navigation graph can be merged with the first layered navigation graph so that the node data in the final layered navigation graph can simultaneously include the representation of the edge connection relationship in the first layered navigation graph and the second layered navigation graph, and the effect of efficient access can be obtained while compressing the data.

[0052] Step S3300: respond to a product search event and determine a query category and a query vector corresponding to the query information carried by the event;

[0053] After a user on an e-commerce platform triggers a product search event, the front-end server responds to the product search event and determines the query information. When a user initiates a product search request on an e-commerce platform, whether by entering keywords, uploading pictures, or selecting specific filtering conditions, the request is considered a product search event. The query information carried by the event consists of two main parts: query category and query vector.

[0054] The query category is directly associated with the level and category specified by the user, and it indicates which level and category the user wants to search in. For example, the user may want to search for products in a specific store or a specific product category, in which case the query category is the store ID of the store or the category ID of the product. The query category can be obtained by extracting the query information carried by the product search event, which can be derived from the user's direct input or automatically determined based on the user's current browsing context.

[0055] The query vector is generated based on the specific content of the user's query, and it contains the semantic information of the user's query. If the user's query is text-based, the query vector may extract keywords from the content of the user input in the query information through text analysis technology, such as natural language processing technology, and then extract its deep semantic information to form a query vector. If the user's query is image-based, the query vector can be generated by extracting visual features from the pictures in the query information uploaded by the user through image recognition technology, such as deep learning models. If the user provides both text and pictures corresponding to the query vector, a deep learning model suitable for extracting multimodal semantic information can be used to extract semantic features from the text and pictures to obtain the corresponding query vector. In either case, the query vector converts the user query into a mathematical representation that can be matched and retrieved in a layered navigation graph.

[0056] After obtaining the query category and query vector, the second hierarchical navigation graph matching the query category is called as the target navigation graph, which will be used in the subsequent retrieval process, that is, to retrieve the commodity nodes that match the semantics of the query vector in the navigation graph.

[0057] Step S3400: call a second hierarchical navigation graph with the same belonging category as the query category as a target navigation graph, and retrieve corresponding nodes of commodities that match the semantics of the query vector from the target navigation graph.

[0058] To obtain search results, the corresponding second hierarchical navigation graph is selected as the target navigation graph according to the query category determined by the user query. This target navigation graph specifically represents the proximity relationship between commodity nodes that carry the same belonging category as the query category, so the search can be limited to the hierarchical category specified by the user. For example, if the query category is a specific online store or commodity category, then the second hierarchical navigation graph corresponding to the online store or commodity category will be selected and determined.

[0059] In different embodiments, the matching relationship between the query category and the attribution category that matches it can be either an exact matching relationship or a fuzzy matching relationship; it can be determined based on rule matching or semantic matching. In addition, in one embodiment, based on a given query category, multiple second hierarchical navigation graphs with attribution categories similar to the query category can be determined as target navigation graphs through semantic matching, and then the query vector is used to determine the semantically matched commodity corresponding nodes therein, and all the matched commodity corresponding nodes are used to respond to the commodity search event.

[0060] In the target navigation graph, a search operation based on the query vector can be performed to determine the corresponding nodes of the relevant products. Specifically, the screening is completed by calculating the data distance between the query vector and the high-dimensional semantic vector of each product node in the target navigation graph. After determining the data distance between the query vector and each product node, the data distance is filtered using a preset threshold to identify the set of nodes closest to the query vector in the vector space. These nodes represent the products that are most semantically relevant to the user's query vector.

[0061] In order to improve the efficiency and accuracy of retrieval, a variety of search algorithms can be used, such as the approximate nearest neighbor (ANN) algorithm, especially the hierarchical navigable small world graph (HNSW) algorithm, which is suitable for quickly locating the most relevant commodity corresponding nodes based on the target navigation graph. These algorithms can reduce the amount of calculation and speed up the search while maintaining a high recall rate.

[0062] For the products corresponding to each node in the search results, various attribute data can be further obtained to construct a product recommendation list, which contains the corresponding nodes of the products that match the query vector semantics. This recommendation list is returned to the user as the search result, and the response to the product search event can be completed. In this way, users can quickly obtain the product information most relevant to their query, thereby optimizing the search experience and improving the service efficiency of the e-commerce platform.

[0063] According to the above embodiments, it can be known that the present application effectively solves the technical problems existing in the background technology, realizes the mixed search of commodities at any level in the hierarchical classification system of the e-commerce platform, and has the ability to provide concurrent search services through the layered navigation graphs of each level. Specifically, the present application can achieve many advantages over the prior art, including but not limited to:

[0064] First of all, the present application has comprehensive advantages in realizing multi-level hybrid search of commodities. The present application prepares a first hierarchical navigation graph, which represents the proximity relationship between commodity nodes in multiple online stores in an e-commerce platform, wherein each node contains not only a high-dimensional semantic vector describing the commodity, but also an attribute scalar indicating the category to which the commodity belongs. This design enables the search service to consider the semantic information of the commodity while also considering the position of the commodity in the hierarchical classification system of the e-commerce platform. Based on this, the present application provides a hybrid search service for commodities by constructing a second hierarchical navigation graph corresponding to each belonging category, and inheriting data such as high-dimensional semantic vectors and attribute scalars through these second hierarchical navigation graphs. This structured method allows the use of corresponding hierarchical navigation graphs at different levels to concurrently respond to commodity search events, so that the commodity search service can simultaneously process vector data and scalar data, thereby responding to the user's query needs more comprehensively. In this way, the present application not only improves the flexibility of search, but also enhances the processing ability of complex queries. Users can search for commodities at any level, whether based on images, texts, or a combination of the two, and can be implemented through vector matching.

[0065] Secondly, this application embodies the advantage of efficiency. By responding to product search events and determining the query category and query vector, the second layered navigation graph with the same query category can be directly called as the target navigation graph, thereby performing a search based on the target navigation graph and quickly locating the relevant product nodes. This method avoids repeated searches and filtering across the entire e-commerce platform, greatly reduces the amount of calculation, and improves the retrieval speed. In addition, since the navigation graph of each level is constructed based on the attribute scalar of the product, it is possible to quickly perform searches on the corresponding level without traversing the entire e-commerce platform data set, thereby achieving rapid retrieval of products.

[0066] Furthermore, the present application embodies the advantages of high recall and high accuracy. In the target navigation graph, relatively accurate matching can be performed based on the semantic similarity between the query vector and the product node, which ensures the relevance of the retrieval results. Since each product node contains a high-dimensional semantic vector, the semantic relationship between the products can be accurately captured, thereby improving the accuracy of the retrieval. At the same time, by independently constructing a navigation graph at each level, all possible search results can be more comprehensively covered, reducing the possibility of missed detection and improving the recall rate. This method of performing hybrid search by combining vector and scalar data enables the search service to provide more relevant results while maintaining high accuracy, making up for the shortcomings of the traditional hierarchical navigable small world graph algorithm for performing product searches and meeting the user's demand for comprehensive search results.

[0067] Based on any embodiment of the method of the present application, before obtaining a first hierarchical navigation graph for representing the proximity relationship between corresponding nodes of commodities in a plurality of online stores of the e-commerce platform, the method includes:

[0068] Step S1100: obtaining product data of each online store in the e-commerce platform, and performing feature extraction on the product image and / or product description text of each product in the product data, generating a corresponding high-dimensional semantic vector and adding it to the corresponding product data, wherein the product data also includes an attribute scalar for indicating the category to which the corresponding product belongs in the hierarchical classification system of the e-commerce platform;

[0069] Each product on the e-commerce platform has its corresponding product data, which includes not only some metadata such as name, price and inventory status, but also key data for subsequent search and matching processes - product images and product description text. These data are the basis for feature extraction, and they will be used to generate high-dimensional semantic vectors that can represent product features.

[0070] Feature extraction is the process of converting unstructured data (such as images and text) into numerical representations that can be used for machine learning and search algorithms. For product images, computer vision techniques, such as deep convolutional neural networks (CNNs), can be used to extract visual features of images, such as color, texture, and shape. For product description texts, natural language processing techniques, such as word embedding or BERT models, may be used to extract semantic features of text. These features capture the key visual and textual information of the product and encode it into vectors in a high-dimensional space, i.e., high-dimensional semantic vectors used to represent the product. In some embodiments, a deep learning model that implements multimodal feature extraction capabilities can also be used. After feature extraction of product images and product description texts, their respective features are fused and processed into the same high-dimensional semantic vector to achieve accurate and comprehensive semantic representation of the product.

[0071] The generated high-dimensional semantic vectors are then added to the corresponding product data and subsequently inherited to the corresponding node data, and stored together with other metadata of the product. These high-dimensional semantic vectors are the core of realizing similarity comparison between products, because they can accurately represent the characteristics of products in high-dimensional space and allow the similarity between products to be evaluated by calculating the data distance between vectors.

[0072] In addition, each product data also contains attribute scalars, which are quantitative information indicating the category to which the product belongs in the hierarchical classification system of the e-commerce platform. The attribute scalar can be the category ID, brand or store logo of the product, etc. It is suitable for the subsequent construction of the second layered navigation map and the execution of hierarchical category-based searches.

[0073] Step S1200: Adapting the hierarchical navigable small-world graph algorithm, performing structured conversion on each commodity data to obtain node data of the corresponding node, and determining the sub-navigation layer to which each node belongs and the edge connection relationship between each node according to the data distance between the high-dimensional semantic vectors of each node data;

[0074] After obtaining the product data of each online store in the e-commerce platform and generating the corresponding high-dimensional semantic vectors, the product data is structured and converted into node data in the Hierarchical Navigable Small World Graph (HNSW) algorithm. The HNSW algorithm is an efficient approximate nearest neighbor search algorithm that accelerates the retrieval process by building a hierarchical graph structure and is particularly suitable for processing high-dimensional data.

[0075] Specifically, each product data is converted into a node in HNSW, which contains the high-dimensional semantic vector and attribute scalar of the product. Attribute scalars, such as store ID or product category, provide each node with a location in the hierarchical classification system of the e-commerce platform. The construction of node data is the basis for implementing product search services, because it allows effective search and matching based on the query vector and query category of the product in subsequent steps.

[0076] When calculating the data distance of high-dimensional semantic vectors between nodes in node data based on the HNSW algorithm, distance metrics (such as Euclidean distance or cosine similarity) can be used to determine the data distance between nodes to represent the similarity or difference between them. Based on these calculation results, the sub-navigation layer to which each node belongs and the edge connection relationship between nodes are determined. In HNSW, each node is not only connected to other nodes in the same sub-navigation layer, but may also be connected to nodes in different sub-navigation layers, forming a complex, multi-dimensional connection network.

[0077] The HNSW (Hierarchical Navigation Small World) algorithm optimizes approximate nearest neighbor search by constructing a hierarchical graph structure. In the present application, the HNSW algorithm is used to determine the number of sub-navigation layers and the specific sub-navigation layer where each node is located. Specifically, the HNSW algorithm constructs multiple sub-navigation layers by decreasing dimensional information, where each layer corresponds to a different dimension and search accuracy. The number of sub-navigation layers is determined by preset parameters, which are usually set based on the size of the data set and the search requirements. Lower-level sub-navigation layers have higher dimensions and finer search granularity, while higher-level sub-navigation layers have lower dimensions and a coarser search range.

[0078] When determining the sub-navigation layer where a node is located, the HNSW algorithm clusters or classifies the nodes according to the data distance between their high-dimensional semantic vectors, and assigns similar nodes to the same sub-navigation layer. This allocation strategy allows the algorithm to search at different levels with different precisions, thereby balancing search efficiency and result accuracy. Lower-level sub-navigation graphs are suitable for precise searches because of their high dimensionality, which can capture more subtle similarities; while higher-level sub-navigation graphs speed up searches by ignoring some details, which is suitable for rough searches. In this way, the HNSW algorithm can effectively locate nodes in multi-layer sub-navigation graphs, achieving fast and accurate product searches.

[0079] It should be pointed out that in the HNSW algorithm, each node originally has a complete high-dimensional semantic vector, which is the same in all levels of sub-navigation layers. The "dimension" referred to in the present application above actually refers to the dimension of the vector space considered in different sub-navigation layers, rather than the change in the dimension of the node's high-dimensional semantic vector itself.

[0080] Specifically, the relationship of decreasing dimensions of multiple layers in the HNSW algorithm does not mean reducing the dimensions of the inherent high-dimensional semantic vectors of nodes, but building different graph structures on different sub-navigation layers. The node data in each level of the graph structure contains complete high-dimensional semantic vectors, but they differ in space division and search strategies. In the sub-navigation graphs at lower levels (close to the bottom layer), the algorithm will consider information of all dimensions to calculate the similarity between nodes; while in the sub-navigation graphs at higher levels (close to the top layer), the algorithm may adopt a rougher strategy, such as considering only part of the dimensions of the vector or adopting looser matching conditions to quickly narrow the search scope.

[0081] This hierarchical design allows the HNSW algorithm to quickly eliminate a large number of irrelevant nodes at the top level and then perform a more detailed search at the bottom level, thus ensuring search efficiency while maintaining the accuracy of search results.

[0082] It can be seen that in this application, the high-dimensional semantic vector of each node is the same in all sub-navigation layers, and the node will not change the dimension of its vector due to the different levels. The difference between different sub-navigation layers lies in the connection structure and search strategy between them, not the dimension of the node vector. In this way, no matter at which level, the algorithm can use the complete high-dimensional semantic vector to perform the search, and use the hierarchical structure to optimize the search process.

[0083] Step S1300: merge multiple sub-navigation layers into a first hierarchical navigation map covering all products on the e-commerce platform.

[0084] Integrate the data of multiple sub-navigation layers into a complete hierarchical navigation graph to obtain the first hierarchical navigation graph, which realizes the graph structure representation of all products on the entire e-commerce platform, making it easier to conduct efficient product searches. When merging sub-navigation layers, the node and edge connection relationships in each sub-navigation layer are retained, and their corresponding positions are found in the new first hierarchical navigation graph. In this way, each node is not only connected to its neighboring nodes in the same sub-navigation layer, but may also be connected to nodes in other sub-navigation layers, forming a multi-dimensional, interconnected complex network. This network can capture the subtle differences and similarities between products, providing rich information for subsequent searches.

[0085] The merged navigation graph not only contains the high-dimensional semantic information of the products, but also integrates the attribute scalar information of the products, so that the search can be performed in multiple dimensions. For example, if the attribute scalar is the category ID of the product, then the product nodes of the same category may be clustered together in the merged navigation graph to form category-specific clusters. This structured information organization method enables the quick location of products of a specific category or with specific attributes when performing a search.

[0086] In addition, the process of merging sub-navigation layers can also include the optimization of the graph structure to ensure the connectivity and search efficiency of the graph. For example, it includes adjusting the edge connection relationship between nodes to reduce redundant edges, or strengthening certain edge connection relationships to improve navigation efficiency during search. In this way, the merged navigation graph can provide a comprehensive and efficient search framework for products on the entire e-commerce platform, allowing users to obtain fast and accurate results no matter which level or category they search in.

[0087] In this embodiment, a multi-layer navigation graph structure is obtained through the above construction process, in which each node is placed in a specific sub-navigation layer according to its high-dimensional semantic vector, and the edge connection relationship between the nodes reflects their relative positions in the vector space. This structured method not only optimizes the organization of commodity data, but also provides a basis for performing efficient search, and can quickly locate relevant commodity nodes when users initiate search requests. In addition, the application of the HNSW algorithm enables the commodity search service to process massive commodity data while maintaining high efficiency and high accuracy of the search. For e-commerce platforms, relevant and accurate results can be quickly returned when users initiate search requests, which can improve user experience and satisfaction.

[0088] On the basis of any embodiment of the method of the present application, corresponding to each attribution category indicated by the attribute scalar, a second hierarchical navigation graph corresponding to each attribution category and indicating the neighbor relationship between the corresponding nodes of the commodity corresponding to the attribution category is constructed, including:

[0089] Step S3210: determine the target attribution category for which the second hierarchical navigation graph needs to be constructed, filter the corresponding attribute scalars of the first hierarchical navigation graph according to the target attribution category, and obtain a target node set containing the target attribution category in the node data;

[0090] In this embodiment, a corresponding second hierarchical navigation graph can be constructed for each attribution category covered by the attribute scalar. To this end, the target attribution category for which the second hierarchical navigation graph needs to be constructed is first determined, and the node data belonging to the target attribution category is filtered out from the first hierarchical navigation graph to form a target node set.

[0091] Specifically, the target attribution category can be to take each attribution category covered by the attribute scalar as the target attribution category one by one, or to select one of all the attribution categories covered by the attribute scalar in the actual usage scenario. Dynamic determination of the target attribution category can be based on a variety of factors, such as user search preferences, marketing strategies of e-commerce platforms, or classification standards of goods. After determining the target attribution category, the first layered navigation graph is filtered to retain only those nodes whose attribute scalars match the target attribution category. This filtering process can be achieved by comparing the attribute scalar of the node with the identifier of the target attribution category, ensuring that only nodes belonging to the target attribution category are selected into the target node set.

[0092] For example, if the target category is a specific store ID, the attribute scalar of each node in the first layered navigation graph can be checked to see if it matches the store ID. Matching nodes are included in the target node set, while non-matching nodes are excluded. Based on this, it is ensured that the target node set only contains corresponding nodes of products directly related to the target category, laying the foundation for building a targeted second layered navigation graph.

[0093] Step S3220, adapting the hierarchical navigable small-world graph algorithm, determining the sub-navigation layer to which each node in the target node set belongs and the edge connection relationship between each node in the target node set according to the data distance between the high-dimensional semantic vectors of the data of each node in the target node set;

[0094] After determining the target node set, the hierarchical navigable small-world graph algorithm can be reapplied to determine the specific layout of these nodes in the second hierarchical navigation graph. To this end, the data distance between the high-dimensional semantic vectors of each node data in the target node set is first calculated, and the sub-navigation layer to which each node belongs and the edge connection relationship between nodes are determined accordingly.

[0095] In this step, similar to the construction of the first hierarchical navigation graph, the hierarchical navigable small-world graph algorithm will be used to construct a sub-navigation layer in the target node set and determine the connection relationship between the nodes. Specifically, the algorithm divides the target node set into multiple sub-navigation layers according to the data distance between the high-dimensional semantic vectors. The nodes in each sub-navigation layer are organized together according to the data distance between their high-dimensional semantic vectors, and their edge connection relationship is determined to form a local connection within the layer.

[0096] For example, if the high-dimensional semantic vectors of two product nodes are very close in the vector space, indicating that the two products are semantically similar, then the two nodes will be connected in the sub-navigation layer. This edge connection relationship reflects the similarity between products and helps to quickly locate similar products during the search process.

[0097] In addition, the hierarchical structure allows the algorithm to adopt different search strategies at different levels. At the lower level, the algorithm can adopt a more sophisticated search strategy, considering information from all dimensions to calculate the similarity between nodes; while at the higher level, the algorithm may adopt a coarser strategy, ignoring some details to speed up the search. This multi-resolution search strategy enables the algorithm to provide high-precision search results while maintaining high efficiency.

[0098] When implementing the hierarchical navigable small-world graph algorithm, a variety of specific embodiments can be used. For example, the number of sub-navigation layers and the connection density of nodes in each layer can be controlled by adjusting the parameters of the algorithm, or the calculation method of the distance between nodes can be customized according to specific business needs. These embodiments demonstrate the flexibility and adaptability of the algorithm, enabling it to be optimized according to different application scenarios.

[0099] Step S3230: merge multiple sub-navigation layers into a second hierarchical navigation map covering the target category of products.

[0100] The process of merging sub-navigation layers is actually to aggregate the node and edge connection relationships in each sub-navigation layer into a unified graph structure. In this process, it is necessary to ensure that the high-dimensional semantic vector and attribute scalar of each node are accurately reflected in the final second-layer navigation graph. Doing so can maintain the proximity and similarity information between product nodes, while making the entire graph structure more compact and efficient.

[0101] During the merging process, a variety of technical means can be used to optimize the graph structure. For example, the edge connection relationship between nodes can be adjusted to strengthen the connection between nodes with similar high-dimensional semantic vectors, while weakening or removing the connection between nodes that are farther away. Such optimization helps improve search efficiency because similar product nodes are easier to locate and retrieve quickly.

[0102] In addition, in one embodiment, merging sub-navigation layers may further compress and simplify the graph structure, for example, by merging nodes with highly similar data distances (differences less than a preset threshold) in the same sub-navigation layer or removing isolated nodes to reduce the complexity of the graph. These operations help reduce storage requirements and improve search speed without significantly affecting the accuracy of search results.

[0103] Through this embodiment, the merged second layered navigation graph will provide a structured search framework for commodities within a specific category, so that when searching for commodities within the category, commodity nodes that match the semantics of the query vector can be quickly located. This targeted navigation graph can improve the commodity search efficiency and user experience of the e-commerce platform when facing user search requests for commodities in a specific category. In this way, the second layered navigation graph not only improves the accuracy of the search, but also optimizes the performance of the search, allowing the e-commerce platform to provide users with faster and more accurate commodity search services.

[0104] On the basis of any embodiment of the method of the present application, corresponding to each attribution category indicated by the attribute scalar, after constructing a second hierarchical navigation graph corresponding to each attribution category and indicating the neighbor relationship between the corresponding nodes of the commodity corresponding to the attribution category, the method includes:

[0105] Step S5100: de-duplication node data of nodes of the second hierarchical navigation graph and node data of nodes of the first hierarchical navigation graph and merge them into the same node data;

[0106] The node data in the second hierarchical navigation graph is deduplicated and merged with the node data in the first hierarchical navigation graph to form a unified node data set. The purpose of deduplication and merging is to avoid data redundancy and ensure that each product node appears only once in the final data set while retaining all relevant information in different hierarchical navigation graphs.

[0107] Specifically, the de-duplication merging process first identifies the nodes of the same product in the two navigation graphs, which can be implemented based on the unique identifier of the node, such as the product ID in the node data or other attribute scalars that can identify the uniqueness of the node. Once the same nodes between different layered navigation graphs are identified, the data of these nodes can be merged, including their high-dimensional semantic vectors and attribute scalars. During the merging process, ensure that all edge connection relationships of the nodes in the first layered navigation graph and the second layered navigation graph are retained, so that these connection relationships can be used for effective navigation in the subsequent search process.

[0108] For example, if a product node is connected to multiple nodes in the first hierarchical navigation graph and is also connected to different nodes in the second hierarchical navigation graph, then in the merged node data, the connection relationships in both navigation graphs will be represented. The advantage of this is that the merged node data can fully reflect the proximity relationship of product nodes at different levels, providing a solid data foundation for accurate product search.

[0109] This deduplication and merging method can effectively integrate information from different levels of navigation graphs, while reducing the need for data storage space and improving data processing efficiency. This is especially important for e-commerce platforms, as they need to process large amounts of product data and ensure that this data can be retrieved and used quickly and accurately. The merged node data will be used to build a fusion graph, which will provide all the information of the first and second layer navigation graphs, providing strong data support for the product search service of the e-commerce platform.

[0110] Step S5200: In the merged node data, the edge connection relationships of the corresponding nodes in the first layered navigation graph and the second layered navigation graph are respectively represented by different edge connection relationship attributes, so as to realize data compression of the first layered navigation graph and the second layered navigation graph to obtain a fusion graph, and the fusion graph is used to provide the first layered navigation graph and the second layered navigation graph.

[0111] In the merged node data, the edge connection relationship between the corresponding nodes in the first hierarchical navigation graph and the second hierarchical navigation graph is distinguished, and the purpose is to merge the information in each hierarchical navigation graph into a fusion graph through data compression technology, such as Figure 3 As shown in Figure 2, the graph can reflect both the global relationship between commodity nodes and the local relationship within a specific level.

[0112] In the fusion graph, the edge connection relationship of each node is represented by different attributes, which can clearly indicate whether the connection between nodes belongs to the first layer navigation graph or the second layer navigation graph. This design allows you to quickly switch between the global and local levels according to the needs of the query when performing a search, thereby improving the flexibility and efficiency of the search.

[0113] For example, if two nodes are connected in the first layered navigation graph because of similar high-dimensional semantic vectors, and in the second layered navigation graph because they belong to the same category, the fusion graph will mark these two connection relationships with different attributes. During the search process, these attributes can be used to quickly identify the similarity or category association between nodes, thereby providing more accurate search results.

[0114] In addition, the design of the fusion graph also takes into account the need for data compression. By merging the repeated information in the two navigation graphs and storing the node and edge connection relationships in a compact form, the fusion graph reduces the storage space required. This data compression method not only improves storage efficiency, but also speeds up the reading and processing of graph data, which is particularly beneficial for e-commerce platforms that process large-scale commodity data.

[0115] In practical applications, the fusion graph can be used in a variety of scenarios. For example, when responding to a search request for a product belonging to a specific category, the second layer navigation graph part of the fusion graph can be used directly for a quick search. When a cross-category or global search is required, the first layer navigation graph part of the fusion graph can be used for a wide search. The flexible search mechanism makes the fusion graph a powerful tool for product search services on e-commerce platforms.

[0116] This embodiment realizes efficient management and retrieval of product data on the e-commerce platform by de-duplicating and merging the data of the second hierarchical navigation graph with the data of the first hierarchical navigation graph and constructing a fusion graph. Compared with other embodiments, the unique technical advantages of this embodiment are mainly reflected in the following aspects:

[0117] First, this embodiment effectively avoids data redundancy through deduplication and merging operations, ensuring that each product node appears only once in the final data set, while retaining all relevant information of the node in the navigation graphs at different levels. This data integration method not only reduces the demand for storage space, but also improves the efficiency of data processing, and can respond to product search requests more quickly.

[0118] Secondly, the design of the fusion graph allows the edge connection relationship of each node to be represented by different attributes, which can clearly indicate whether the connection between nodes belongs to the first layer navigation graph or the second layer navigation graph. This design allows you to quickly switch between the global and local levels according to the needs of the query when performing a search, thereby improving the flexibility and efficiency of the search.

[0119] Furthermore, the data compression technology of the integrated graph not only improves storage efficiency, but also speeds up the reading and processing of graph data. This is especially beneficial for e-commerce platforms that process large-scale commodity data, because it can significantly reduce the computing and storage burden when maintaining and querying large-scale graph data.

[0120] Finally, the flexible search mechanism of the fusion graph enables e-commerce platforms to quickly locate products in a specific category or conduct a global search across categories based on user query requirements. This mechanism not only improves the accuracy of search results, but also optimizes the user's shopping experience, allowing users to find the products they need more quickly.

[0121] Based on any embodiment of the method of the present application, responding to a product search event and determining a query category and a query vector corresponding to the query information carried by the event include:

[0122] Step S3310: respond to a product search event, parse the event to obtain query information carried therein, wherein the data content of the query information includes a first content corresponding to a category to which the product belongs, and a second content corresponding to a data content for generating a high-dimensional semantic vector of the product;

[0123] When a user initiates a search request, the corresponding product search event needs to be parsed to obtain the query information. The query information usually contains two main parts: a first content for indicating the category to which the product belongs, and a second content for generating a high-dimensional semantic vector for the product.

[0124] The first content is the information related to the category of the product in the user's query, which can be a category identifier explicitly specified by the user, such as the brand, type or store of the product. For example, the user can enter a specific online store name, logo or select a specific product category in the search box, and this information is the first content. By parsing the filter conditions entered or selected by the user, the specific information related to the product category is extracted to limit the scope of subsequent searches.

[0125] The second content is information related to the product features in the user query, which may include pictures uploaded by the user and / or text descriptions entered. For example, a user may upload a product picture or enter a descriptive text of the product, and this information is the second content.

[0126] Step S3320: formatting the first content to convert the first content into a query category;

[0127] Considering the possible ambiguity when the user inputs information, the first content usually needs to be formatted to be converted into a recognizable query category. In this way, the first content input by the user can be converted to ensure that the product category that the user is interested in can be understood and accurately located based on the first content.

[0128] The formatting process includes multiple specific embodiments. For example, if the user inputs a category identifier in text form, text cleaning can be performed to remove useless characters or spaces and convert the text into a standardized format. For a category identifier of a numerical type, such as a category ID of a product, the legality of the input can be verified and converted into a category identifier for internal use. In some cases, the user may select a category through a graphical user interface, in which case the predefined category identifier returned by the interface component can be used directly.

[0129] For example, when a user enters "sports shoes" in the search box and selects the subcategory "running shoes", the text string "running shoes" is converted into a corresponding category identifier, such as the corresponding ID or code in the database. If the user searches by selecting the category navigation on the store page, the category identifier of the category can be directly obtained based on the user's selection. These converted query categories will be used to guide the subsequent search process to ensure that the search results are limited to the categories specified by the user.

[0130] Step S3330: calling a feature extraction model for generating a high-dimensional semantic vector of a product, extracting deep semantic information from the second content, and obtaining a corresponding query vector.

[0131] In this step, the feature extraction model is called to process the second content in the user query, that is, information related to product features, such as uploaded pictures and / or input text descriptions. The purpose of the feature extraction model is to extract deep semantic information from the second content and generate a high-dimensional semantic vector that can represent the user's query intent.

[0132] Feature extraction models can be implemented using a variety of techniques. For example, for image data, a convolutional neural network (CNN) in deep learning technology can be used, which can extract visual features such as shape, texture, and color from images. For text data, natural language processing techniques such as word embedding or Transformer-based models can be used, which can convert text into continuous vector representations and capture the semantic information of the text.

[0133] In actual operation, the feature extraction model receives the user's uploaded image or input text as input, and outputs a high-dimensional semantic vector through a series of predefined calculation steps. This vector captures the core features of the user's query and converts it into a form that can be matched and retrieved in the hierarchical navigation graph. For example, if a user uploads a picture of sneakers, the feature extraction model will analyze the picture and extract features related to sneakers, such as the pattern of the sole, the material of the upper, etc., and encode these features into a high-dimensional vector.

[0134] In addition, the feature extraction model can also combine multimodal data, that is, process image and text data simultaneously to obtain a more comprehensive query vector. This multimodal feature extraction can improve the accuracy of search because it combines information from different data sources and provides richer query context.

[0135] It is not difficult to understand that, generally speaking, the feature extraction model for extracting high-dimensional semantic vectors for the second content in this embodiment can be the same feature extraction model as the feature extraction model for generating high-dimensional semantic vectors in node data based on information in product data in the previous text.

[0136] This embodiment achieves a deep understanding of and efficient response to the user's query intent by accurately parsing the user's product search events and extracting query categories and query vectors from them. Compared with other embodiments, the unique technical advantage of this embodiment is that it can convert the user's query content, whether it is a text-based description or image-based information, into precise query categories and high-dimensional semantic vectors, thereby performing fast and accurate matching in the multi-level navigation graph. This conversion not only improves the efficiency of the search because it reduces unnecessary calculations and comparisons, but also improves the relevance of the search results because it ensures that the search results are highly consistent with the user's query intent. In addition, by using advanced feature extraction models, this embodiment is able to process complex multimodal data, further enhancing the power and flexibility of the search system, and providing users with a richer and more personalized search experience.

[0137] Based on any embodiment of the method of the present application, retrieving a corresponding node of a commodity that semantically matches the query vector from the target navigation graph includes:

[0138] Step S3410: In the topmost sub-navigation layer of the target navigation graph, based on the data distance between the query vector and the high-dimensional semantic vector of each node, identify the set of nearest nodes;

[0139] The process of retrieving the corresponding nodes of the products that match the semantics of the query vector from the target navigation graph begins with the operation of the top-level sub-navigation layer. In the top-level sub-navigation layer, the query vector is compared with the high-dimensional semantic vectors of all nodes to determine which nodes are closest to the query vector. This comparison process can form a set of nearest neighbor nodes by calculating the data distance between the query vector and each node vector and selecting those nodes with the smallest distance using a preset threshold. This node set represents the nodes in the top-level sub-navigation layer that are most similar to the query vector.

[0140] For example, if the query vector represents a user's search request for a specific product, identify the nodes closest to the query vector in the topmost sub-navigation layer, which represent the products most relevant to the user's query. This process is part of a hierarchical search strategy that allows the search scope to be gradually narrowed down at different levels, ultimately locating the product node that best matches the user's query.

[0141] Step S3420: Select a node from the nearest node set as an entry point, adapt the navigation mechanism of the hierarchical navigable small-world graph algorithm, and recursively search the lower sub-navigation layer along the path with the shortest distance between the query vector and each node data;

[0142] The nodes selected from the nearest node set will be used as the starting point of the search, and the direction of the next search will be determined based on the data distance between these nodes and the query vector. In the hierarchical navigable small-world graph algorithm, each node is usually connected to multiple neighboring nodes, forming a complex network structure. Along the path with the shortest distance to the query vector data, the lower sub-navigation layer is recursively retrieved, gradually narrowing the search scope until the target level is reached.

[0143] For example, if a node in the topmost sub-navigation layer is closest to the query vector, this node can be selected as the entry point and moved along the edge connected to it to the corresponding node in the next sub-navigation layer. In the next sub-navigation layer, this process is repeated to continue looking for nodes that are closer to the query vector. This recursive process continues until the bottommost sub-navigation layer is reached or a product node that matches the query vector semantics is found.

[0144] When implementing the hierarchical navigable small-world graph algorithm, a variety of specific embodiments can be used. For example, different search strategies, such as breadth-first search or depth-first search, can be set to adapt to different search requirements. In addition, the parameters of the algorithm, such as the number of neighbor nodes to be selected or the depth of the recursive search, can also be adjusted to optimize the search efficiency and the accuracy of the results. These different embodiments demonstrate the flexibility and adaptability of the hierarchical navigable small-world graph algorithm, enabling it to be customized and optimized according to different application scenarios.

[0145] Step S3430: after recursively going to the bottom sub-navigation layer of the target navigation graph layer by layer, select the corresponding nodes of the products that match the semantics of the query vector;

[0146] After recursively returning to the bottom sub-navigation layer of the target navigation graph layer by layer, a filtering operation can be performed to identify the corresponding nodes of the products that match the semantics of the query vector, thereby ensuring that product information that is highly relevant to the user's query intent is accurately extracted from the candidate node set.

[0147] In the bottom sub-navigation layer, the data distance between the query vector and the high-dimensional semantic vector of each node is compared to determine which nodes are semantically closest to the query vector. The data distance can be implemented using metrics such as cosine similarity or Euclidean distance. Based on these calculation results, nodes whose distance to the query vector is within a preset threshold are selected and considered to match the query vector semantics.

[0148] For example, if a user searches for a specific style of sneakers, the high-dimensional semantic vectors of the nodes in the underlying sub-navigation layer are similar to the query vector to filter out the product nodes that represent the same or similar styles of sneakers. This screening process takes into account the multi-dimensional features of the product, including but not limited to the visual features, text descriptions, and categories of the product, ensuring the relevance and accuracy of the search results.

[0149] In addition, the selected node set includes not only nodes that directly match the query vector, but also nodes that are not exactly the same as the query vector but are similar in some key features. This flexible matching mechanism makes the search results richer and can cover more potential related products.

[0150] When implementing the screening operation, a variety of specific embodiments can be used. For example, different similarity thresholds can be set to meet the needs of different users for the accuracy of search results. In addition, the screening strategy can be adjusted in combination with the user's search history and preferences to make the search results more personalized. These embodiments demonstrate the flexibility and adaptability of the screening process, enabling it to be optimized according to different application scenarios and user needs, thereby providing users with a fast and accurate product search service experience.

[0151] Step S3440: construct a corresponding product recommendation list based on the filtered nodes and return it as a search result to complete the response to the product search event.

[0152] The product recommendation list constructed based on the filtered nodes can be returned to the user as a search result, thereby completing the response to the product search event. The product recommendation list contains product information that is most relevant to the user's query.

[0153] In order to generate a list of recommended products, the filtered nodes are first sorted. The sorting can be based on the similarity score between the node and the query vector, or combined with other factors, such as the sales volume, user reviews, or price of the product. The purpose of sorting is to ensure that the products in the recommended product list are arranged according to the degree of relevance to the user's query, so that the most relevant products appear at the top of the list.

[0154] When generating the recommendation list, the detailed information of the product corresponding to each node is integrated, such as the product image, description, price, store information, product link, etc. This information provides users with comprehensive shopping decision support. In addition, the product recommendation list can be personalized according to the user's personal preferences and historical purchasing behavior to improve the accuracy and attractiveness of the recommendation.

[0155] A variety of specific embodiments can be used to construct the list. For example, different sorting algorithms can be set, such as sorting based on vector similarity, sorting based on user behavior, or comprehensive sorting combining multiple factors. In addition, the recommendation strategy can be dynamically adjusted based on user feedback to optimize future search results.

[0156] The unique technical advantage of this embodiment is that it can accurately and efficiently retrieve commodity nodes that match the user query semantics from complex commodity data. By identifying the node set closest to the query vector in the topmost sub-navigation layer of the target navigation graph, and recursively retrieving to the bottom sub-navigation layer along the path with the shortest data distance, this embodiment implements a hierarchical and refined search strategy. This method not only optimizes the search path and reduces unnecessary calculations, but also ensures the high relevance and accuracy of the retrieval results by performing precise screening operations in the bottom sub-navigation layer. In addition, by constructing a commodity recommendation list containing rich commodity information and sorting it according to various factors such as sales volume, user evaluation, etc., this embodiment further enhances the user experience and makes the search results more personalized and practical. This method of multi-level search using high-dimensional semantic vectors and attribute scalars provides faster search speed and higher search quality compared to other embodiments, and is particularly suitable for processing complex query requirements in large-scale e-commerce platforms.

[0157] Based on any embodiment of the method of the present application, after retrieving the corresponding node of the commodity that matches the semantics of the query vector from the target navigation graph, the method includes:

[0158] Step S7100: monitor the change events of the product data of each online store in the e-commerce platform, and respond to the change events;

[0159] In the daily operation of e-commerce platforms, changes in product data are common, including updates to product information, the listing of new products, and the removal of old products. In order to maintain the timeliness and accuracy of navigation maps at all levels, monitoring these change events and responding to them can ensure the effectiveness of product search services. Monitoring product data change events of online stores on e-commerce platforms requires real-time tracking of product data changes, including but not limited to modifications to product descriptions, price adjustments, inventory changes, or updates to product images.

[0160] Monitoring of change events can be achieved in a variety of ways. For example, polling can be used to regularly check changes in product data, or an event-driven model can be used, in which the backend service of the e-commerce platform actively sends notifications when data changes. In addition, real-time monitoring can be achieved by using database triggers or message queue mechanisms, which can immediately notify and process data changes. Once a change event is detected, it can include waking up a background task, calling an API interface, or starting a workflow to handle the change.

[0161] Step S7200: When the change data belongs to the commodity data of an existing commodity, the node data of the corresponding commodity in the first layered navigation graph and the associated second layered navigation graph are updated in real time using the corresponding change data;

[0162] When the change data corresponding to the monitored change event belongs to the commodity data of the existing commodity in the first layered navigation graph, especially when the metadata of the high-dimensional semantic vector and the attribute title is not affected, the node data of the corresponding commodity in the first layered navigation graph and the associated second layered navigation graph are updated in real time using these change data. For example, if the price of a commodity changes, the price data in the corresponding node data in the first layered navigation graph and the second layered navigation graph containing the nodes of the corresponding commodity can be updated accordingly to reflect these changes. This process ensures that the navigation graph can accurately reflect the latest status of the commodity, thereby providing the latest search results.

[0163] Step S7300: When the change data is data content corresponding to the high-dimensional semantic vector, the high-dimensional semantic vector in the node of the updated product is regenerated based on the change data, and the node corresponding to the updated product is relocated in the first layered navigation graph and the second layered navigation graph;

[0164] When the monitored product data changes contain key feature data such as data content for generating high-dimensional semantic vectors, such as updates to product images or description texts, in addition to updating the corresponding metadata in the node data, it is also necessary to regenerate the high-dimensional semantic vectors of the affected products and update the positions of the nodes in each navigation graph accordingly.

[0165] Specifically, when the key feature data of a product changes, such as updating the product image or modifying the product description, these changes will directly affect the generation of high-dimensional semantic vectors. Therefore, it is necessary to re-extract features based on the latest product data and generate new high-dimensional semantic vectors.

[0166] After generating a new high-dimensional semantic vector, it is necessary to update the node data of the affected products in the first hierarchical navigation graph and the associated second hierarchical navigation graph, and redefine the sub-navigation layer affiliation and edge connection relationship of these nodes to ensure that the position of the product node in the navigation graph can accurately reflect the latest high-dimensional semantic vector, thereby ensuring that the search results can accurately reflect the latest features of the product.

[0167] For example, if a product image is updated, resulting in a significant change in the visual features of the product, the high-dimensional semantic vector regenerated based on the updated image will be different. This new vector must be used to reposition the product node in the navigation graph to ensure its accuracy in search results and to ensure that the structure of the navigation graph reflects the latest status of product data.

[0168] Step S7400: When the changed data belongs to the product data of a newly added product, generate a high-dimensional semantic vector corresponding to the newly added product, construct a node corresponding to the newly added product, insert the node into the first layered navigation graph, and insert the node into the corresponding second layered navigation graph according to the belonging category indicated by the attribute scalar of the newly added product.

[0169] The listing of new products is an ongoing activity of an online store. In order to ensure that the navigation graph can reflect the latest product information, data changes of newly added products need to be specially handled. When the product data of newly added products is monitored, a series of operations are performed to include these new products in the search structure of the navigation graph.

[0170] First, a high-dimensional semantic vector is generated for the newly added product. To this end, feature extraction is performed on the product image and / or description text, and the product data of the new product is processed using the same feature extraction model as before to generate a high-dimensional semantic vector representing the product features, ensuring that the new product can be accurately represented and identified in the navigation graph. After the high-dimensional semantic vector is generated, the node of the new product is constructed and inserted into the first hierarchical navigation graph, including determining the position of the new product node in the navigation graph to ensure that the new product can be included in the existing search structure. The insertion of the node is based on the relationship between its high-dimensional semantic vector and the existing product node, as well as possible attribute scalars, such as store logo, product category or brand information. Subsequently, according to the belonging category indicated by the attribute scalar of the newly added product, the new product node is inserted into the corresponding second hierarchical navigation graph, ensuring that the new product is not only visible in the first hierarchical navigation graph at the full platform level, but also can be found in the second hierarchical navigation graph corresponding to the specific category view, enhancing the accuracy and efficiency of the search.

[0171] The process of inserting nodes can be implemented in a variety of ways, including but not limited to direct insertion, optimizing the insertion position through a sorting algorithm, or reserving space for new products in a specific area of ​​the navigation graph. Each implementation method aims to ensure that new products can be seamlessly integrated into the existing navigation graph while maintaining the performance and accuracy of the navigation graph.

[0172] The unique technical advantage of this embodiment is that it can adapt to the dynamic changes of product information on the e-commerce platform, ensuring the real-time update and high accuracy of the navigation map. By monitoring the change events of product data and responding quickly, this embodiment can not only update the metadata of existing products in real time, but also regenerate high-dimensional semantic vectors when the key feature data of the product changes, and reposition the product nodes in the navigation map, ensuring the accuracy and timeliness of the search results. In addition, for new products, this embodiment can quickly generate high-dimensional semantic vectors and integrate their nodes into the corresponding navigation map, ensuring that new products can be quickly included in the search range, enhancing the comprehensiveness of the search results. This dynamic update mechanism enables the e-commerce platform to continuously provide the latest and most accurate product search services, improve the user experience, and optimize the response speed and processing capabilities of the e-commerce platform to product changes, which is difficult to match with other embodiments.

[0173] See also Figure 5According to one aspect of the present application, a multi-level hybrid search device is provided, comprising a base graph calling module 3100, a sub-graph construction module 3200, an event response module 3300, and a response execution module 3400, wherein the base graph calling module 3100 is configured to obtain a first hierarchical navigation graph for representing the proximity relationship between corresponding nodes of commodities in multiple online stores of an e-commerce platform, wherein the nodes include a high-dimensional semantic vector for describing the corresponding commodities and an attribute scalar for indicating the belonging category of the corresponding commodities in the hierarchical classification system of the e-commerce platform; the sub-graph construction module 3200 is configured to construct a second hierarchical navigation graph corresponding to each belonging category indicating the proximity relationship between corresponding nodes of the commodities corresponding to the belonging category for each belonging category indicated by the corresponding attribute scalar; the event response module 3300 is configured to respond to a commodity search event and determine the query category and query vector corresponding to the query information carried by the event; the response execution module 3400 is configured to call the second hierarchical navigation graph having the same belonging category as the query category as the target navigation graph, and retrieve the corresponding nodes of the commodities that match the semantics of the query vector from the target navigation graph.

[0174] On the basis of any embodiment of the device of the present application, the device of the present application further includes: a data acquisition module, configured to acquire product data of each online store in the e-commerce platform, and perform feature extraction on the product image and / or product description text of each product in the product data, generate a corresponding high-dimensional semantic vector and add it to the corresponding product data, the product data also includes an attribute scalar for indicating the category to which the corresponding product belongs within the hierarchical classification system of the e-commerce platform; a composition execution module, configured to adapt to the layered navigable small-world graph algorithm, perform a structured transformation on each product data to obtain the node data of the corresponding node, and determine the sub-navigation layer to which each node belongs and the edge connection relationship between each node according to the data distance between the high-dimensional semantic vectors of each node data; a mapping merging module, configured to merge multiple sub-navigation layers into a first layered navigation map covering all products on the e-commerce platform.

[0175] Based on any embodiment of the device of the present application, the sub-graph construction module 3200 includes: a filtering and screening module, which is configured to determine the target attribution category that needs to be constructed for the second layered navigation graph, and filters the corresponding attribute scalars of the first layered navigation graph according to the target attribution category to obtain a target node set containing the target attribution category in the node data; a sub-graph construction module, which is configured to adapt to the layered navigable small-world graph algorithm, and determines the sub-navigation layer to which each node in the target node set belongs and the edge connection relationship between each node according to the data distance between the high-dimensional semantic vectors of each node data in the target node set; and a multi-graph merging module, which is configured to merge multiple sub-navigation layers into a second layered navigation graph covering the target attribution category products.

[0176] On the basis of any embodiment of the device of the present application, the device of the present application also includes: a node merging module, configured to de-duplicate the node data of the nodes of the second layered navigation graph and the node data of the nodes of the first layered navigation graph and merge them into the same node data; a connection optimization module, configured to represent the edge connection relationships of the corresponding nodes in the first layered navigation graph and the second layered navigation graph respectively through different edge connection relationship attributes in the merged node data, so as to realize data compression of the first layered navigation graph and the second layered navigation graph to obtain a fusion graph, and the fusion graph is used to provide the first layered navigation graph and the second layered navigation graph.

[0177] Based on any embodiment of the device of the present application, the event response module 3300 includes: an event parsing module, which is configured to respond to a product search event and parse the event to obtain query information carried therein, wherein the data content of the query information includes a first content corresponding to a category indicating the product, and a second content corresponding to the data content for generating a high-dimensional semantic vector of the product; a hierarchical processing module, which is configured to format the first content to convert the first content into a query category; and a vector processing module, which is configured to call a feature extraction model for generating a high-dimensional semantic vector of the product, extract deep semantic information from the second content, and obtain a corresponding query vector.

[0178] Based on any embodiment of the device of the present application, the response execution module 3400 includes: a starting identification module, which is set in the topmost sub-navigation layer of the target navigation graph, and identifies the nearest node set based on the data distance between the query vector and the high-dimensional semantic vector of each node; a recursive retrieval module, which is set to select a node from the nearest node set as an entry point, adapt to the navigation mechanism of the hierarchical navigable small-world graph algorithm, and recursively search the lower sub-navigation layer along the path with the shortest distance between the query vector and the data of each node; a bottom-level screening module, which is set to screen out the corresponding nodes of the products that match the query vector semantics after recursing layer by layer to the bottom-level sub-navigation layer of the target navigation graph; a result output module, which is set to construct a corresponding product recommendation list based on the screened nodes and return it as a retrieval result to complete the response to the product search event.

[0179] On the basis of any embodiment of the device of the present application, the device of the present application also includes: a change response module, which is configured to monitor change events of product data of each online store in the e-commerce platform and respond to the change events; a data update module, which is configured to use the corresponding change data to perform real-time updates on the node data of the corresponding product in the first layered navigation graph and the associated second layered navigation graph when the change data belongs to the product data of an existing product; a node positioning module, which is configured to regenerate the high-dimensional semantic vector in the node of the updated product based on the change data when the change data is the data content corresponding to the high-dimensional semantic vector generated, and relocate the corresponding node of the updated product in the first layered navigation graph and the second layered navigation graph; a new insertion module, which is configured to generate a high-dimensional semantic vector corresponding to the new product when the changed data belongs to the product data of the new product, construct a node corresponding to the new product, insert the node into the first layered navigation graph, and insert the node into the corresponding second layered navigation graph according to the belonging category indicated by the attribute scalar of the new product.

[0180] Another embodiment of the present application also provides a multi-level hybrid search device. Figure 6 As shown, a schematic diagram of the internal structure of a multi-level hybrid search device. The multi-level hybrid search device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. Among them, the computer-readable non-volatile readable storage medium of the multi-level hybrid search device stores an operating system, a database, and computer-readable instructions. The database may store an information sequence. When the computer-readable instructions are executed by the processor, the processor may implement a multi-level hybrid search method.

[0181] The processor of the multi-level hybrid search device is used to provide computing and control capabilities to support the operation of the entire multi-level hybrid search device. The memory of the multi-level hybrid search device may store computer-readable instructions, which, when executed by the processor, may enable the processor to execute the multi-level hybrid search method of the present application. The network interface of the multi-level hybrid search device is used to connect and communicate with a terminal.

[0182] Those skilled in the art will understand that Figure 6 The structure shown in the figure is merely a block diagram of a portion of the structure related to the scheme of the present application, and does not constitute a limitation on the multi-level hybrid search device to which the scheme of the present application is applied. The specific multi-level hybrid search device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0183] In this embodiment, the processor is used to execute Figure 5The memory stores the program code and various data required to execute the above modules or submodules. The network interface is used to realize data transmission between user terminals or servers. The non-volatile readable storage medium in this embodiment stores the program code and data required to execute all modules in the multi-level hybrid search device of this application, and the server can call the program code and data of the server to execute the functions of all modules.

[0184] The present application also provides a non-volatile readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the multi-level hybrid search method of any embodiment of the present application.

[0185] The present application also provides a computer program product, including a computer program / instruction, which implements the steps of the method described in any embodiment of the present application when executed by one or more processors.

[0186] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments of the present application can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a computer-readable storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0187] In summary, this application provides an efficient product search solution for e-commerce platforms. Through a multi-level hybrid search strategy, it realizes fast and accurate search of massive product data, while taking into account the processing of vector and scalar data, significantly improving the speed and accuracy of retrieval, optimizing the user experience, and providing strong technical support for complex queries on e-commerce platforms.

Claims

1. A multi-level hybrid search method, characterized in that: include: Obtaining a first hierarchical navigation graph for representing proximity relationships between corresponding nodes of products in a plurality of online stores on an e-commerce platform, wherein the nodes include high-dimensional semantic vectors for describing corresponding products and attribute scalars for indicating the categories to which the corresponding products belong in a hierarchical classification system of the e-commerce platform; Corresponding to each attribution category indicated by the attribute scalar, respectively constructing a second hierarchical navigation graph corresponding to each attribution category, which represents the neighbor relationship between the corresponding nodes of the commodities corresponding to the attribution category; In response to a product search event, determine the query category and query vector corresponding to the query information carried by the event; A second hierarchical navigation graph with the same belonging category as the query category is called as a target navigation graph, and a corresponding node of the commodity that matches the semantics of the query vector is retrieved from the target navigation graph.

2. The multi-level hybrid search method according to claim 1, characterized in that: Before obtaining the first hierarchical navigation graph for representing the proximity relationship between corresponding nodes of products in multiple online stores of the e-commerce platform, the following steps are included: Obtain product data of each online store in the e-commerce platform, and perform feature extraction on the product image and / or product description text of each product in the product data, generate a corresponding high-dimensional semantic vector and add it to the corresponding product data, wherein the product data also includes an attribute scalar indicating the category to which the corresponding product belongs in the hierarchical classification system of the e-commerce platform; Adapting to the hierarchical navigable small-world graph algorithm, each commodity data is structured and converted to obtain the node data of the corresponding node. According to the data distance between the high-dimensional semantic vectors of each node data, the sub-navigation layer to which each node belongs and the edge connection relationship between each node are determined; Merge multiple sub-navigation layers into a first-level navigation map that covers all products on the e-commerce platform.

3. The multi-level hybrid search method according to claim 2, characterized in that: For each attribution category indicated by the corresponding attribute scalar, a second hierarchical navigation graph corresponding to each attribution category and indicating the neighbor relationship between the corresponding nodes of the commodity corresponding to the attribution category is constructed, including: Determine the target attribution category for which the second hierarchical navigation graph needs to be constructed, filter the corresponding attribute scalar of the first hierarchical navigation graph according to the target attribution category, and obtain a target node set containing the target attribution category in the node data; Adapting to the hierarchical navigable small-world graph algorithm, according to the data distance between the high-dimensional semantic vectors of the data of each node in the target node set, the sub-navigation layer to which each node in the target node set belongs and the edge connection relationship between each node are determined; Merge multiple sub-navigation layers into a second layered navigation map covering the target category products.

4. The multi-level hybrid search method according to any one of claims 1 to 3, characterized in that: After constructing, for each attribution category indicated by the attribute scalar, a second hierarchical navigation graph corresponding to each attribution category and indicating the neighbor relationship between the corresponding nodes of the commodity corresponding to the attribution category, the following steps are included: De-duplication and merging of node data of nodes of the second hierarchical navigation graph and node data of nodes of the first hierarchical navigation graph into the same node data; In the merged node data, different edge connection relationship attributes are used to respectively represent the edge connection relationships of the corresponding nodes in the first layered navigation graph and the second layered navigation graph, so as to achieve data compression of the first layered navigation graph and the second layered navigation graph to obtain a fusion graph, and the fusion graph is used to provide the first layered navigation graph and the second layered navigation graph.

5. The multi-level hybrid search method according to any one of claims 1 to 3, characterized in that: In response to a product search event, the query category and query vector corresponding to the query information carried by the event are determined, including: In response to a product search event, the event is parsed to obtain query information carried therein, wherein the data content of the query information includes first content corresponding to a category to which the product belongs and second content corresponding to data content for generating a high-dimensional semantic vector of the product; performing formatting processing on the first content to convert the first content into a query category; A feature extraction model for generating a high-dimensional semantic vector of a product is called to extract deep semantic information from the second content to obtain a corresponding query vector.

6. The multi-level hybrid search method according to any one of claims 1 to 3, characterized in that: Retrieving a corresponding node of a product that matches the semantics of the query vector from the target navigation graph includes: In the topmost sub-navigation layer of the target navigation graph, the nearest node set is identified based on the data distance between the query vector and the high-dimensional semantic vector of each node; Select a node from the nearest node set as the entry point, adapt to the navigation mechanism of the hierarchical navigable small-world graph algorithm, and recursively search the lower sub-navigation layer along the path with the shortest distance between the query vector and each node data; After recursively traversing the bottom sub-navigation layer of the target navigation graph layer by layer, the corresponding nodes of the products that match the semantics of the query vector are selected; According to the filtered nodes, a corresponding product recommendation list is constructed and returned as the search result to complete the response to the product search event.

7. The multi-level hybrid search method according to any one of claims 1 to 3, characterized in that: After retrieving the corresponding nodes of the products that match the semantics of the query vector from the target navigation graph, the method includes: Monitor the change events of product data of each online store on the e-commerce platform and respond to the change events; When the change data belongs to the commodity data of the existing commodity, the node data of the corresponding commodity in the first layered navigation graph and the associated second layered navigation graph are updated in real time using the corresponding change data; When the change data is data content corresponding to the high-dimensional semantic vector, the high-dimensional semantic vector in the node of the updated product is regenerated based on the change data, and the node corresponding to the updated product is relocated in the first layered navigation graph and the second layered navigation graph; When the changed data belongs to the product data of a newly added product, a high-dimensional semantic vector corresponding to the newly added product is generated, a node corresponding to the newly added product is constructed, the node is inserted into the first layered navigation graph, and the node is inserted into the corresponding second layered navigation graph according to the belonging category indicated by the attribute scalar of the newly added product.

8. A multi-level hybrid search device, characterized in that: include: a base graph calling module, configured to obtain a first hierarchical navigation graph for representing proximity relationships between corresponding nodes of products in a plurality of online stores on the e-commerce platform, wherein the nodes include high-dimensional semantic vectors for describing the corresponding products and attribute scalars for indicating the categories to which the corresponding products belong in the hierarchical classification system of the e-commerce platform; A sub-graph construction module is set to each attribution category indicated by the corresponding attribute scalar, and constructs a second hierarchical navigation graph corresponding to each attribution category, which represents the neighbor relationship between the corresponding nodes of the commodities corresponding to the attribution category; An event response module, configured to respond to a product search event and determine a query category and a query vector corresponding to the query information carried by the event; The response execution module is configured to call a second hierarchical navigation graph with the same belonging category as the query category as a target navigation graph, and retrieve corresponding nodes of commodities that match the semantics of the query vector from the target navigation graph.

9. A multi-level hybrid search device, comprising a central processing unit and a memory, characterized in that: The central processing unit is used to call and run the computer program stored in the memory to execute the steps of the method according to any one of claims 1 to 7.

10. A non-volatile readable storage medium, characterized in that: It stores a computer program implemented according to the method described in any one of claims 1 to 7 in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the corresponding method are executed.

Images