Vertical e-commerce platform commodity recommendation method and device, equipment and storage medium

Abstract

The application provides a vertical e-commerce platform commodity recommendation method, device and equipment and a storage medium, the method comprising: collecting data streams including user comments, trend topics and commodity information from multiple data sources, and preprocessing the data streams to obtain standardized text data; performing entity recognition, relationship extraction and sentiment analysis on the standardized text data to obtain a knowledge triple set containing entities, entity relationship and sentiment attributes; fusing the knowledge triple set with an existing knowledge graph, resolving conflict information in the fusion process, and obtaining an updated domain knowledge graph; determining candidate commodities and associated paths with the user according to user information and entity association relationships in the domain knowledge graph, and obtaining a commodity recommendation result. The application continuously fuses external multi-source dynamic information and performs incremental graph updating, so that the recommendation system can timely perceive market changes and improve the timeliness and accuracy of recommendations.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a product recommendation method, apparatus, device, and storage medium for a vertical e-commerce platform. Background Technology

[0002] Vertical e-commerce platforms derive their core competitiveness from their professionalism and precision. However, existing product recommendation systems primarily rely on users' historical transaction data and collaborative filtering algorithms, making it difficult to effectively utilize dynamic information from external platforms such as social media and industry analytics websites. This external information often includes key data such as market trend changes, real-time user preferences, and competitor activities, which are of significant value for accurate recommendations in vertical e-commerce.

[0003] In existing technologies, knowledge graph-based recommendation methods typically use static knowledge graphs as auxiliary information sources. For example, graph embedding-based recommendation methods map entities and relationships in the knowledge graph to low-dimensional vectors and combine them with user-item interaction matrices for recommendations. Other methods attempt to combine knowledge graphs with collaborative filtering, leveraging entity relationship paths within the graph to enhance the interpretability of recommendations. However, the knowledge graphs relied upon by these methods are usually built offline and not updated, failing to reflect dynamic market changes. Their construction still relies on manual annotation or semi-automated batch extraction processes, resulting in high update costs and poor timeliness, with update cycles lasting days or even weeks. When emerging trends appear on social media or technological innovations occur in the industry, recommendation systems miss marketing opportunities due to knowledge lag, leading to a disconnect between recommended content and market reality, impacting user experience and platform conversion rates. Furthermore, existing knowledge graphs lack effective conflict resolution mechanisms when integrating multi-source heterogeneous data. When different data sources provide contradictory information about the same entity attribute, it is difficult to automatically determine the credibility of the information, resulting in inconsistent knowledge within the graph and consequently affecting recommendation quality. Summary of the Invention

[0004] The main objective of this invention is to solve the technical problems of existing product recommendation systems being unable to integrate external dynamic information in a timely manner, resulting in delayed knowledge graph updates, and lacking an effective conflict resolution mechanism when integrating multi-source data, leading to inconsistent knowledge. This invention provides a product recommendation method for a vertical e-commerce platform, characterized in that the product recommendation method for the vertical e-commerce platform includes: Data streams are collected from multiple data sources and preprocessed to obtain standardized text data, wherein the data streams include user comments, trending topics, and product information from vertical e-commerce platforms; Entity recognition, relation extraction, and sentiment analysis are performed on the standardized text data to obtain a set of knowledge triples containing entities, relationships between entities, and sentiment attributes. The knowledge triple set is fused with the existing knowledge graph, and conflicting information during the fusion process is resolved to obtain an updated domain knowledge graph. Obtain user information of the target user, including browsing history, purchase history, and favorited items; Based on user information and entity relationships in the domain knowledge graph, candidate products and their association paths with users are determined, resulting in product recommendation results.

[0005] The present invention also provides a product recommendation device for a vertical e-commerce platform, characterized in that the product recommendation device for the vertical e-commerce platform includes: The data acquisition module is used to collect data streams from multiple data sources, preprocess the data streams to obtain standardized text data, wherein the data streams include user comments, trending topics and product information from vertical e-commerce platforms; The knowledge extraction module is used to perform entity recognition, relation extraction and sentiment analysis on the standardized text data to obtain a set of knowledge triples containing entities, relationships between entities and sentiment attributes. The knowledge graph fusion module is used to fuse the knowledge triple set with the existing knowledge graph and resolve conflicting information during the fusion process to obtain an updated domain knowledge graph. The information acquisition module is used to acquire user information of the target user, including browsing history, purchase history and favorited items. The recommendation generation module is used to determine candidate products and their association paths with users based on user information and entity relationships in the domain knowledge graph, and to obtain product recommendation results.

[0006] The present invention also provides a product recommendation device for a vertical e-commerce platform, comprising: a memory and at least one processor, wherein the memory stores instructions, and the memory and the at least one processor are interconnected via a circuit; the at least one processor invokes the instructions in the memory to cause the product recommendation device for the vertical e-commerce platform to perform the steps of the product recommendation method for the vertical e-commerce platform described above.

[0007] The present invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the steps of the above-described product recommendation method for a vertical e-commerce platform.

[0008] The aforementioned product recommendation method, apparatus, equipment, and storage medium for vertical e-commerce platforms collect data streams from multiple data sources, including user reviews, trending topics, and product information. These data streams are preprocessed to obtain standardized text data. Entity recognition, relation extraction, and sentiment analysis are performed on the standardized text data to obtain a set of knowledge triples containing entities, inter-entity relationships, and sentiment attributes. This set of knowledge triples is then fused with an existing knowledge graph, and conflicting information during the fusion process is resolved to obtain an updated domain knowledge graph. Based on user information and entity relationships in the domain knowledge graph, candidate products and their association paths with users are determined, resulting in product recommendations. This invention, by continuously fusing external multi-source dynamic information and incrementally updating the knowledge graph, enables the recommendation system to promptly perceive market changes, improving the timeliness and accuracy of recommendations.

[0009] Beneficial Effects: By continuously collecting and integrating user comments, trending topics, and product information from multiple data sources, the domain knowledge graph can dynamically reflect market changes. Compared to existing technologies that use offline batch updates, this invention incrementally integrates and resolves conflicts in the knowledge triple set, avoiding service interruptions and update delays caused by full reconstruction, and reducing knowledge update delays from days in traditional batch methods to minutes. Since the domain knowledge graph includes inter-entity relationships and sentiment attributes, recommendations can be based on entity association paths, enhancing the interpretability of the recommendation results. When new trends emerge on social media or competitors release new products, the knowledge graph can promptly supplement relevant knowledge triples, enabling the recommendation system to quickly respond to market hotspots and reduce recommendation bias caused by knowledge lag. Furthermore, through a conflict resolution mechanism based on authority scoring and time decay, the accuracy of conflict resolution in the knowledge graph is significantly improved compared to a simple timestamp-first strategy, further ensuring the accuracy of recommendation results.

[0010] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0011] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the first embodiment of the product recommendation method for a vertical e-commerce platform in this invention; Figure 2 This is a schematic diagram of a second embodiment of the product recommendation method for a vertical e-commerce platform in this invention. Figure 3 This is a schematic diagram of one embodiment of the product recommendation device for a vertical e-commerce platform in this invention. Figure 4 This is a schematic diagram of one embodiment of a product recommendation device for a vertical e-commerce platform according to an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0015] To facilitate understanding of this embodiment, a product recommendation method for a vertical e-commerce platform disclosed in this invention will first be described in detail. For example... Figure 1 As shown, this method includes the following steps: 101. Collect data streams from multiple data sources, preprocess the data streams to obtain standardized text data, wherein the data streams include user comments, trending topics, and product information from vertical e-commerce platforms; In this embodiment, the step of collecting data streams from multiple data sources and preprocessing the data streams to obtain standardized text data includes: obtaining raw data streams from the multiple data sources, including social media platforms, vertical e-commerce platforms, and industry analysis websites, through API interfaces or web crawlers; pushing the raw data streams to a stream processing framework to perform deduplication and format conversion on the raw data streams to obtain a pre-processed data stream; performing language type identification on the text content in the pre-processed data streams, inputting the identified non-target language text into a machine translation model for translation to obtain target language text; and performing text cleaning and noise filtering on the target language text to obtain standardized text data.

[0016] Specifically, the recommendation systems of vertical e-commerce platforms need to acquire dynamic information from outside the platform to supplement the lack of internal transaction data. This external information mainly comes from three types of data sources: social media platforms, vertical e-commerce platforms, and industry analysis websites. Social media platforms can reflect users' real-time preferences and emerging trends, such as a certain type of product being discussed or shared extensively in a short period of time; vertical e-commerce platforms contain structured and unstructured data such as product descriptions and user reviews from their own platform and competitors' platforms; and professional reports and technology updates provided by industry analysis websites help to understand the technological evolution and market changes in the vertical industry.

[0017] The system can periodically retrieve raw data streams from the aforementioned data sources using API interfaces or web crawling techniques. API interfaces are suitable for platforms that provide standardized interfaces, allowing data requests to be made at a preset frequency. For data sources that do not provide APIs, web crawling techniques are used to extract the required content by parsing the webpage structure. The raw data stream is typically in a mixed format, containing multiple formats such as JSON, XML, and HTML, and may contain duplicate data.

[0018] To improve the efficiency of subsequent processing, the raw data stream needs to be pushed to a stream processing framework for initial processing. The stream processing framework's role is to receive data in real time and perform initial deduplication and format conversion. Deduplication is based on unique identifiers of the data, such as the ID of a social media post or the timestamp of an e-commerce review combined with the user ID, using a hash algorithm to quickly determine if the data already exists. Format conversion unifies heterogeneous data from different sources into a preset intermediate format, facilitating unified processing by subsequent modules.

[0019] The initial data stream may contain text in multiple languages. For vertical e-commerce platforms targeting international markets, user reviews and social media discussions often involve multiple languages ​​such as English, Japanese, and Korean. The language type identification module determines the language type of the text by analyzing the character distribution and common vocabulary. When non-target language text is identified, the system inputs it into a machine translation model for translation. The selection of the machine translation model needs to balance translation quality and response speed, and pre-trained neural machine translation models are typically used. The translated text is then standardized to the target language, such as Chinese or English, thereby eliminating language barriers.

[0020] The target language text still contains noisy content, including HTML tags, special symbols, emojis, and meaningless repetitive characters. Text cleaning removes this noise using regular expressions and natural language processing tools. Specifically, HTML tags can be directly stripped using a parsing library; special symbols are filtered according to a pre-defined retention list, retaining symbols related to product attributes, such as " / " used to indicate size specifications; emojis, although common in social media, contribute little to knowledge extraction and are usually removed. The noise-filtered text is standardized text data, with a uniform format, consistent language, and clean content, suitable as input for the subsequent knowledge extraction module.

[0021] It's worth noting that the introduction of stream processing frameworks enables near real-time data processing. Compared to traditional batch ETL processes, stream processing frameworks can complete preprocessing within seconds or minutes of data arrival, which is crucial for capturing emerging trending topics on social media. For example, after a new product generates buzz on social media, the system can quickly collect, clean, and translate the data, providing timely input for subsequent knowledge extraction and graph updates.

[0022] 102. Perform entity recognition, relation extraction, and sentiment analysis on the standardized text data to obtain a set of knowledge triples containing entities, relationships between entities, and sentiment attributes; In this embodiment, standardized text data needs to undergo deep semantic parsing to be transformed into structured knowledge triples. This process relies on the semantic understanding capabilities of a pre-trained language model, and knowledge extraction is completed collaboratively through three sub-tasks: named entity recognition, relation extraction, and sentiment analysis.

[0023] After receiving standardized text data, the pre-trained language model first converts the text into vector representations through a word embedding layer, and then captures contextual semantics through a multi-layer Transformer structure. The named entity recognition module then labels entities in the text based on this. Specifically, this module uses sequence labeling to assign an entity type label to each word in the text. In vertical e-commerce scenarios, entity types include not only general product and brand entities, but also domain-specific attribute entities and user entities. Attribute entities refer to professional attributes such as the functional characteristics, ingredients, or design style of a product, such as the "niacinamide" ingredient in skincare products or the "refresh rate" parameter in mobile phones. User entities include commenters, opinion leader accounts, etc.

[0024] After entity annotation is completed, the relation extraction module is responsible for determining the semantic relationships between entity pairs. The difficulty of relation extraction lies in understanding the implicit connections between entities, rather than relying solely on obvious syntactic structures. For example, in the text "This serum contains a high concentration of hyaluronic acid, with obvious moisturizing effects," it is necessary to identify the inclusion relationship between "serum" and "hyaluronic acid," and the efficacy relationship between "hyaluronic acid" and "moisturizing effect." The relation classification module determines the relation type by analyzing the semantic vectors of entity pairs and their context. Common relation types include recommendation relations, inclusion relations, competition relations, and evaluation relations. Recommendation relations typically appear in the content of opinion leaders, indicating that a user or account recommends a product; inclusion relations reflect the product's ingredient composition or configuration information; competition relations reflect the comparison of different products within the same category; this invention also provides evaluation relations, which connect user entities and product entities, expressing user experience.

[0025] The introduction of the sentiment analysis module aims to capture users' fine-grained emotional inclinations towards product attributes. Traditional sentiment analysis typically classifies entire texts as positive, negative, or neutral. However, in vertical e-commerce scenarios, user reviews are often multifaceted. A product might receive positive reviews on some attributes while being criticized on others. Therefore, sentiment analysis needs to be conducted at the attribute level. The system first locates text segments containing attribute entities, and then identifies modifiers and evaluative words describing the attribute characteristics within those segments. For example, in the text "The screen display is amazing, but the battery life is average," the modifier "amazing" corresponding to "screen" and the evaluative word "average" corresponding to "battery life" need to be extracted separately. The sentiment classification module then determines the sentiment polarity of these modifiers, labeling "amazing" as positive and "average" as neutral or negative.

[0026] When the same attribute entity is mentioned multiple times in the text and corresponds to different sentiments, the system uses a weighted average strategy to determine the overall sentiment polarity. The position and intensity of modifiers in the text affect their weights; for example, modifiers appearing at the beginning of a sentence or with emphatic adverbs usually have higher weights.

[0027] After completing entity recognition, relation extraction, and sentiment analysis, the system integrates the extraction results into knowledge triples. Each knowledge triple contains four elements: head entity, relation type, tail entity, and sentiment attribute. The head and tail entities can be products, brands, attributes, or users; the relation type describes the semantic connection between them; and the sentiment attribute is attached to triples involving evaluation relations, recording the user's sentiment towards a specific attribute. For example, from the text above, triples (screen, attribute, positive) and (battery life, attribute, negative) can be constructed, and these triples together constitute the knowledge triple set.

[0028] It should be noted that pre-trained language models, trained on general corpora, may have misunderstandings regarding specialized terminology and expressions in vertical domains. To improve extraction accuracy, the model can be fine-tuned on labeled data from vertical e-commerce domains to adapt it to domain-specific entity types and relationship patterns. During fine-tuning, the loss function simultaneously considers the prediction errors of entity recognition, relationship classification, and sentiment classification, thereby improving overall performance through multi-task learning.

[0029] 103. The knowledge triple set is fused with the existing knowledge graph, and conflicting information in the fusion process is resolved to obtain an updated domain knowledge graph; In this embodiment, the process of fusing the knowledge triple set with the existing knowledge graph and resolving conflicting information during the fusion process to obtain an updated domain knowledge graph includes: calculating the similarity between entities in the knowledge triple set and entities in the existing knowledge graph; determining that entities are the same entity when the similarity exceeds a preset threshold, thus obtaining an entity alignment mapping relationship; adding new triples from the knowledge triple set to the existing knowledge graph or updating existing triples using atomic operations based on the entity alignment mapping relationship; detecting inconsistent information provided by multiple data sources on the same attribute of the same entity, thus obtaining a conflict information set; calculating the comprehensive confidence level for each conflict item in the conflict information set based on the authority score and corresponding timestamp of each data source, and selecting the knowledge triple with the highest comprehensive confidence level to replace the corresponding conflict item, thus obtaining the updated domain knowledge graph.

[0030] Specifically, existing knowledge graphs are structured knowledge bases built on historical data, storing entities such as products, brands, attributes, and users, along with their relationships. Newly extracted sets of knowledge triples need to be integrated with existing knowledge graphs to achieve continuous knowledge updates. The core challenges of this integration process lie in entity alignment and conflict resolution, as these two aspects directly determine the accuracy and consistency of the knowledge graph.

[0031] Entity alignment is a prerequisite for fusion. Since different data sources may represent the same entity differently—for example, the standard model name and the abbreviation of a mobile phone may refer to the same product, and "hyaluronic acid" and "hyaluronic acid" are different names for the same ingredient—the system needs to determine whether these representations refer to the same entity. Similarity calculation is accomplished by comparing the literal form of the entity, semantic vectors, and contextual features. Literal form comparison is based on edit distance or character overlap, which can capture spelling differences and abbreviations; semantic vectors map entities to a high-dimensional vector space through a pre-trained language model, where similar entities are closer together; contextual features consider the co-occurring words of the entity in the text. For example, mobile phone products often co-occur with words like "screen" and "processor," and this consistency in patterns can serve as an alignment criterion.

[0032] When the overall similarity between entities exceeds a preset threshold, the system determines that they are the same entity and establishes an entity alignment mapping relationship. This mapping relationship records the correspondence between newly extracted entities and entities in the existing graph, and is used for subsequent triple addition and update operations. It is important to note that the similarity threshold setting needs to balance recall and precision. A threshold that is too low will lead to misalignment, mistaking different entities for the same one; a threshold that is too high may miss synonymous entities, resulting in redundant nodes in the graph.

[0033] After determining the entity alignment relationships, the system updates the knowledge graph using atomic operations. Atomicity means that each update is executed as an independent transaction, either succeeding completely or failing completely, without any intermediate partial update states. This mechanism ensures the consistency and reliability of graph updates. Specifically, when the entity involved in a triple in the knowledge triple set does not exist in the existing graph, the system creates a new entity node and relation edge, adding the triple to the graph; when the entity involved in the triple already exists but the relation or attribute value has changed, the system updates the content of the existing triple.

[0034] Graph databases' transaction mechanisms ensure the implementation of atomic operations. Each add or update operation is encapsulated as a database transaction. If an exception occurs during transaction execution, the system automatically rolls back to the state before the operation, preventing inconsistencies in the graph. This incremental update method has significant advantages over full reconstruction. Full reconstruction requires taking the graph service offline, exporting all data, merging new knowledge, rebuilding the index, and loading it again—a process that can take several hours. In contrast, atomic incremental updates allow the graph to continuously provide query services during the update process, with each update operation typically taking milliseconds, thus supporting near real-time knowledge update requirements.

[0035] Conflict detection occurs after entity alignment. When multiple data sources provide different information about the same attribute of the same entity, the system needs to determine which information is more credible. The conflict detection module traverses the knowledge triples to be added, checking for cases where the head entity and attribute are the same but the tail entity or sentiment polarity differs. For example, social media data shows a skincare product is priced at 300 yuan, while competitor platform data shows the same product is priced at 280 yuan, constituting a conflict in the price attribute. Detected conflicts are collected into a conflict information set, awaiting further resolution.

[0036] Conflict resolution is based on two dimensions: the authority rating of the data source and the timestamp. The authority rating reflects the credibility of the data source; for example, information published by an official brand account is more authoritative than comments from ordinary users, and industry analysis reports are more authoritative than fragmented discussions on social media. The authority rating can be determined comprehensively based on factors such as the historical accuracy of the data source, the certification level of the publishing entity, and the professionalism of the content. The timestamp reflects the freshness of the information; for dynamic attributes such as product prices and inventory status, newer information is often more accurate.

[0037] The calculation of overall confidence combines authority scores and timestamps. Timestamps are converted into time decay coefficients using a time decay function; the further back in time, the smaller the decay coefficient. The time decay function can employ an exponential decay form, causing the timeliness of information to decrease rapidly over time. The overall confidence of the triple is obtained by multiplying the authority score of the data source by the time decay coefficient of the corresponding knowledge triple. The system selects the triple with the highest overall confidence as retained information, replacing the corresponding conflicting item in the conflicting information set.

[0038] Furthermore, the step of calculating the comprehensive confidence level for each conflict item in the conflict information set based on the authority rating and corresponding timestamp of each data source, and selecting the knowledge triple with the highest comprehensive confidence level to replace the corresponding conflict item includes: obtaining multiple candidate knowledge triples corresponding to each conflict item in the conflict information set, as well as the authority rating and timestamp of the data source to which each candidate knowledge triple belongs; calculating a time decay coefficient based on the time difference between the current time and the timestamp of each candidate knowledge triple, wherein the larger the time difference, the smaller the time decay coefficient; multiplying the authority rating of the data source to which each candidate knowledge triple belongs by the corresponding time decay coefficient to obtain the comprehensive confidence level of each candidate knowledge triple; and selecting the candidate knowledge triple with the highest comprehensive confidence level as retained information to replace the corresponding conflict item in the conflict information set.

[0039] Specifically, the conflict resolution process requires selecting the most credible triple from multiple conflicting candidate knowledge triples. Each conflicting item actually corresponds to multiple candidate knowledge triples from different data sources. These triples are identical in head entity and relation type, but differ in tail entity or sentiment polarity. The system first obtains these candidate knowledge triples, along with the authority ratings and timestamp information of their respective data sources.

[0040] The acquisition of authority ratings relies on the pre-assessment of data sources. Different types of data sources vary significantly in terms of information reliability. Product information published by official brand accounts is typically internally reviewed and therefore highly accurate; product detail pages on e-commerce platforms are maintained by merchants and are generally reliable in terms of structured information such as specifications; however, user discussions on social media may contain subjective speculation or outdated information. The system assigns a fixed authority rating to each data source based on factors such as the type of data source, historical accuracy, and the authentication status of the publishing entity. This rating is determined during the data collection phase and stored along with the data source information.

[0041] Timestamps record the publication or collection time of the original data corresponding to the knowledge triple. In terms of dynamic attributes, the timeliness of information is crucial. For example, product prices change with promotional activities, inventory status updates with sales, and user sentiment towards products may shift with version iterations. A comment published six months ago might target an older version of the product, the issues of which have been fixed in the newer version. Therefore, newer information should be given higher weight in conflict resolution.

[0042] The calculation of the time decay coefficient reflects the impact of information freshness on credibility. The system first calculates the time difference between the current time and the timestamp of the candidate knowledge triplet; a larger time difference indicates that the information is more outdated. The time decay function uses an exponential decay form, specifically calculated as: d(Δt) = e^(-λ·Δt), where Δt is the time difference between the current time and the timestamp of the knowledge triplet, and λ is the decay rate parameter used to control the decay speed; a larger λ value indicates faster decay of information timeliness. When Δt is zero, d(Δt) equals 1, indicating that newly published information has the highest time credibility; as Δt increases, d(Δt) tends towards zero. Linear decay assumes that the credibility of information decreases uniformly over time, while exponential decay assumes that information maintains high credibility initially and then decays rapidly. In vertical e-commerce scenarios, due to the rapid pace of market changes, exponential decay is more consistent with reality. The time decay coefficient is designed to be a value between zero and one; the decay coefficient for newly published information is close to one, gradually approaching zero over time.

[0043] The overall confidence score is obtained by multiplying the authority score by the time decay coefficient. This multiplication operation integrates two dimensions: the authority score reflects the inherent credibility of the data source, and the time decay coefficient reflects the timeliness of the information. The product of the two takes into account both "who said it" and "when it was said it." For example, information from an authoritative data source but published a long time ago will have a lower overall confidence score due to time decay; while information from an ordinary data source but very recent, although having a low authority score, may have a competitive overall confidence score due to a time decay coefficient close to one.

[0044] After calculating the overall confidence of all candidate knowledge triples, the system selects the triple with the highest confidence as the retained information. This triple is considered the most reliable knowledge at present and is used to replace the corresponding conflicting item in the conflicting information set. The replacement operation is performed in the graph database; the system deletes or marks other candidate triples, retaining only the selected triple.

[0045] It's worth noting that conflict resolution strategies may need adjustment for different attribute types. For relatively stable attributes, such as product composition and design style, the impact of time decay is relatively small, and authority ratings dominate; however, for dynamic attributes, such as price, inventory, and user sentiment, the weight of time decay should be increased. The system can dynamically adjust the weight ratio of authority ratings and time decay coefficients in the overall confidence calculation based on attribute type, making conflict resolution more accurate.

[0046] 104. Obtain user information of the target user, including browsing history, purchase history, and favorited items; 105. Based on user information and entity relationships in the domain knowledge graph, determine candidate products and their association paths with users to obtain product recommendation results.

[0047] In this embodiment, determining candidate products and their associated paths with users based on user information and entity relationships in the domain knowledge graph to obtain product recommendation results includes: identifying trend nodes that are newly added or have significantly increased weight within a preset time window from the domain knowledge graph to obtain a set of hot trend nodes; locating the corresponding user nodes in the domain knowledge graph based on the user information, calculating the path length and path weight from the user nodes to each product node using a graph traversal algorithm to obtain a set of candidate products and their corresponding graph distance scores; for products in the candidate product set that are associated with the set of hot trend nodes, adjusting and sorting the graph distance scores according to the association strength, and selecting a preset number of products with the highest scores as recommended products; extracting the associated paths from the user nodes to each recommended product, converting the associated paths into natural language descriptions, and obtaining product recommendation results containing recommended products and recommendation reasons.

[0048] Specifically, the recommendation generation process is based on an updated domain knowledge graph. The knowledge graph stores entities and relationships in a graph structure, with users, products, brands, attributes, trending topics, etc., all existing as nodes, and their semantic relationships connected by weighted edges. The task of the recommendation system is to find product nodes relevant to a specific user within this graph structure and explain the basis for the recommendations.

[0049] Identifying trending topics is the first step in the recommendation process. As the domain knowledge graph is continuously updated, certain nodes may undergo significant changes due to a concentrated influx of external information. The system identifies these trending nodes by monitoring their dynamic changes. Specifically, the system sets a time window and compares the number of associated edges or the sum of their weights within that window with the historical average outside the window. When the increase exceeds a preset threshold, the node is marked as a trending topic node. These nodes constitute a set of trending topics, representing the current focus of market attention.

[0050] User node location relies on user information. User information includes behavioral data such as the user's browsing history, purchase history, favorited items, and followed accounts. The system locates or creates corresponding user nodes in the domain knowledge graph based on this information. If the user already has historical interaction records, their node already exists in the graph and is connected to related entity nodes through relational edges; if the user is a new user, the system creates a user node based on the interest tags or initial browsing behavior provided during user registration and establishes initial associations.

[0051] Graph traversal algorithms are used to calculate the path length and path weight from user nodes to each product node. Path length refers to the number of edges traversed from a user node to a product node, reflecting their distance in the graph structure. Path weight is the cumulative weight of all edges on the path, reflecting the semantic strength of the path. The system employs a weighted shortest path algorithm, prioritizing edges with larger weights during traversal to find the semantically strongest path from the user node to each product node.

[0052] Path length and path weight together determine the graph distance score. A shorter path length means a more direct association between the user and the product, while a higher path weight means a stronger semantic association. The graph distance score is typically calculated using the formula: S = W(p) / (1 + L(p)), where W(p) is the sum of the weights of all edges on path p, L(p) is the number of edges traversed by path p, and adding 1 to the denominator avoids division by zero and smooths the path length. The shorter the path length and the higher the path weight, the higher the score. All reachable product nodes and their graph distance scores constitute the candidate product set.

[0053] The introduction of trending topics enhances the timeliness of recommendations. Some products in the candidate product set may be associated with trending topic sets. The system detects the association between candidate product nodes and trending topic nodes and adjusts the graph distance scores of candidate products based on the strength of the association. The strength of the association is measured by the edge weights between product nodes and trending topic nodes; if a product is connected to multiple trending topic nodes, the weights of these edges are accumulated. After the weighted adjustment, products closely associated with trending topics will have significantly higher scores, thus occupying a higher position in the ranking.

[0054] The system sorts candidate products according to their adjusted scores and selects a preset number of products with the highest scores as recommended products. The preset number can be dynamically adjusted according to the recommendation scenario.

[0055] The generation of recommendation reasons is crucial to the interpretability of recommendation results. The system extracts the association paths from user nodes to each recommended product node. This path records the intermediate nodes and relational edges traversed from the user to the product. The conversion of paths to natural language relies on predefined templates and relation type mappings. The system designs corresponding language expressions for each relation type, traverses the nodes and edges in the path, generates sentence fragments according to mapping rules, and then concatenates the fragments into complete sentences.

[0056] In this embodiment, data streams including user reviews, trending topics, and product information are collected from multiple data sources. The data streams are preprocessed to obtain standardized text data. Entity recognition, relation extraction, and sentiment analysis are performed on the standardized text data to obtain a set of knowledge triples containing entities, inter-entity relationships, and sentiment attributes. This set of knowledge triples is then fused with an existing knowledge graph, and conflicting information during the fusion process is resolved to obtain an updated domain knowledge graph. Based on user information and entity relationships in the domain knowledge graph, candidate products and their association paths with users are determined, resulting in product recommendations. This invention, by continuously fusing external multi-source dynamic information and performing incremental graph updates, enables the recommendation system to promptly perceive market changes, improving the timeliness and accuracy of recommendations.

[0057] Please see Figure 2 Another embodiment of the product recommendation method for vertical e-commerce platforms in this application includes: 201. Collect data streams from multiple data sources, preprocess the data streams to obtain standardized text data, wherein the data streams include user comments, trending topics and product information from vertical e-commerce platforms; In this embodiment, step 201 is similar to step 101 in the first embodiment, and will not be described again here.

[0058] 202. Input the standardized text data into the pre-trained language model, and use the named entity recognition module to annotate the entities in the text to obtain an entity annotation sequence, wherein the entity annotation sequence includes product entities, brand entities, attribute entities and user entities; In this embodiment, the pre-trained language model undertakes the task of text semantic understanding. Standardized text data is converted into vector representations through a word embedding layer and then input into a multi-layer Transformer encoder for context modeling. Based on the encoder output, the named entity recognition module annotates the entity type of each word in the text.

[0059] Named entity recognition uses sequence labeling to classify words in text into tags such as the beginning, middle, end, or non-entity of an entity. In vertical e-commerce scenarios, the definition of entity types needs to be combined with domain characteristics. Product entities refer to specific product names or models, brand entities refer to product manufacturers or brand identifiers, attribute entities cover professional attributes such as product functions, composition, and design style, and user entities include commenters, opinion leader accounts, etc.

[0060] Attribute entity identification is crucial in this embodiment. In vertical e-commerce, user reviews often revolve around specific product attributes rather than general evaluations of the product as a whole. Attribute entities may appear explicitly or be implicit in descriptive phrases. The named entity recognition module analyzes the semantic roles and syntactic structures of words to label these attribute-related words as attribute entities. After entity labeling is complete, the system outputs an entity labeling sequence. This sequence records the type, start position, and end position of each entity in the text.

[0061] 203. Determine the semantic relationships between entity pairs in the entity annotation sequence to obtain the relationship annotation results; 204. Based on the text fragments containing attribute entities in the standardized text data, identify the user's sentiment towards each attribute and assign a sentiment polarity label to each attribute entity. In this embodiment, the step of identifying the user's sentiment towards each attribute based on text fragments containing attribute entities in the standardized text data and assigning a sentiment polarity label to each attribute entity includes: extracting text fragments containing attribute entities from the standardized text data; identifying modifiers and evaluation words describing attribute features in the text fragments to obtain an attribute-modifier pair set; inputting the attribute-modifier pair set into the sentiment classification module to classify and judge the sentiment tendency of each modifier to obtain positive, negative, or neutral sentiment polarity labels; associating the sentiment polarity labels with the corresponding attribute entities; when multiple sentiment polarity labels exist for the same attribute entity, performing a weighted average according to the weight of the modifiers in the text to obtain the comprehensive sentiment polarity label of the attribute entity.

[0062] Specifically, attribute-level sentiment analysis needs to capture users' attitudes towards different attributes at a fine-grained level. The system first extracts text fragments containing attribute entities from standardized text data. The boundaries of the text fragments are usually determined at the sentence level, meaning that complete sentences containing attribute entities are extracted as the analysis object.

[0063] In the extracted text fragments, the system identifies modifiers and evaluative words that describe attribute features. Modifiers are typically adjectives or adverbs that directly modify attribute entities, while evaluative words express the user's subjective judgment of the attribute. The identification process relies on dependency parsing, which analyzes the grammatical dependencies between words to determine which modifiers or evaluative words are semantically associated with attribute entities. The system pairs attribute entities with their corresponding modifiers, forming a set of attribute-modifier pairs.

[0064] The attribute-modifier pair set is then input into the sentiment classification module. Based on a sentiment dictionary or a trained classifier, the sentiment classification module determines the sentiment polarity for each modifier. Sentiment polarity is categorized into three types: positive, negative, and neutral. Positive polarity indicates user satisfaction or approval of the attribute, negative polarity indicates dissatisfaction or criticism, and neutral polarity indicates an objective description or a vague attitude. The classification module outputs the sentiment polarity label for each modifier.

[0065] Sentiment polarity labels need to be associated with corresponding attribute entities. In simple cases, an attribute entity corresponds to only one modifier in the text, and the sentiment polarity label of that modifier can be directly assigned to the attribute entity. However, in complex texts, the same attribute entity may be described by multiple modifiers, and the sentiment polarities of these modifiers may be inconsistent. In this case, the system uses a weighted average strategy to determine the overall sentiment polarity of the attribute entity.

[0066] The weight of modifiers in the text is determined by several factors. First, the distance between the modifier and the attribute entity in the syntactic tree affects the weight; the closer the distance, the higher the weight. Second, the emotional intensity of the modifier itself also affects the weight; strong emotional words such as "very" and "extremely" will increase the weight of their modified objects. The system assigns a weight value to each modifier based on these factors and calculates a weighted average to obtain the comprehensive emotional polarity label of the attribute entity. If the weighted sum of positive emotions is greater than that of negative emotions, the attribute entity is labeled as positive; otherwise, it is labeled as negative; if the two are similar, it is labeled as neutral.

[0067] It's worth noting that attribute-level sentiment analysis provides more granular user preference information compared to document-level or sentence-level sentiment analysis. Users may like certain attributes of a product while being dissatisfied with others; this multi-dimensional sentiment information is of significant value to recommendation systems in vertical e-commerce.

[0068] 205. Based on the entity annotation sequence, the relation annotation result, and the sentiment polarity label, construct a knowledge triplet containing a head entity, relation type, tail entity, and sentiment attribute to obtain a knowledge triplet set; In this embodiment, before performing entity recognition, relation extraction, and sentiment analysis on the standardized text data, the method further includes: source labeling of the standardized text data, classifying the text as social media context, e-commerce platform context, or professional forum context based on the platform type of the data source, to obtain context-labeled data; extracting professional terms from the context-labeled data, statistically analyzing the co-occurring words and modifiers of each professional term in different contexts, calculating the semantic vector representation of each professional term in different contexts, to obtain a set of polysemous terms; for homographs in the set of polygraphs whose semantic vector distance exceeds a preset threshold, creating multiple semantic nodes in the domain knowledge graph, and associating each semantic node with a corresponding context label, wherein the context label is used to select the corresponding semantic version of the term node for matching during entity recognition and relation extraction, and to establish a mapping relationship between different semantic nodes in the knowledge graph when cross-contextual term references are detected, the mapping relationship is used to record the semantic differences of homographs in different contexts, and to identify users' cross-contextual product preferences during recommendation generation, preventing recommendation errors caused by term polysemy.

[0069] Specifically, the construction of knowledge triples requires integrating the outputs of three modules: entity recognition, relation extraction, and sentiment analysis. Knowledge triples are the basic units of structured knowledge representation, with each triple containing four elements: head entity, relation type, tail entity, and sentiment attribute. The head and tail entities are derived from the entity annotation sequence, the relation type from the relation annotation results, and the sentiment attribute from the sentiment polarity label.

[0070] In vertical e-commerce scenarios, the ambiguity of technical terms poses a challenge to knowledge extraction. The same term may carry different meanings in different platform or community contexts. Before performing entity recognition and relation extraction, the system needs to perform contextual annotation on standardized text data. Contextual annotation categorizes texts into social media contexts, e-commerce platform contexts, or professional forum contexts based on the platform type from which the data originates. Expressions in social media contexts are usually more colloquial and emotionally charged, e-commerce platform contexts focus on product function descriptions and user experience, while professional forum contexts contain more technical details and industry terminology. The system identifies the metadata of the data source and attaches contextual tags to each text segment to obtain context-annotated data.

[0071] Based on context-annotated data, the system extracts domain-specific terms and analyzes their semantic differences. The extraction of these terms relies on a term extraction algorithm that combines word frequency statistics and domain dictionary matching to identify specialized vocabulary in the text. For each term, the system analyzes its co-occurring words and modifiers in different contexts. Co-occurring words are those that frequently appear in the same sentence or paragraph as the term, while modifiers are adjectives or adverbs that directly describe the characteristics of the term. By analyzing the distribution differences of these co-occurring words and modifiers, the system can capture the semantic changes of terms in different contexts.

[0072] Semantic vector representation is a key technology for quantifying the semantics of terms. The system constructs a semantic vector for each term in various contexts. This vector integrates the term's word embeddings, the average vector of co-occurring words, and the sentiment features of modifiers. Semantic vectors from different contexts are compared using cosine similarity or Euclidean distance. When the semantic vector distance of the same term in different contexts exceeds a preset threshold, the system determines that the term has multiple meanings and includes it in a set of term polysemy.

[0073] For homographs within a set of polysemous terms, the system creates multiple semantic nodes in the domain knowledge graph. Each semantic node corresponds to the semantic version of the term in a specific context and is associated with a corresponding contextual tag. This design enables the knowledge graph to distinguish the different meanings of the same literal form in different contexts, avoiding semantic confusion.

[0074] When performing entity recognition and relation extraction, the system selects term nodes with corresponding semantic versions for matching based on the contextual tags of the current text. When the text is labeled as belonging to a social media context, the system prioritizes matching term nodes associated with social media context tags; when the text comes from a professional forum, it matches term nodes from the professional forum context. This context-aware matching mechanism improves the accuracy of entity recognition and relation extraction.

[0075] In some cases, semantic relationships exist between term nodes in different contexts. The system detects cross-contextual term references, and when it finds that a term refers to related but different concepts in multiple contexts, it establishes a mapping relationship between different semantic nodes in the knowledge graph. This mapping relationship records the cross-contextual changes in term semantics, which helps to understand how users express the same product or attribute differently on different platforms.

[0076] After completing context-aware entity recognition and relation extraction, the system integrates the outputs of each module into knowledge triples. For each identified entity pair, the system searches for the corresponding relation type and checks for the existence of associated sentiment polarity tags. If the relation involves evaluation or recommendation, the sentiment polarity tag is attached as the sentiment attribute of the triple; if the relation is a neutral relation such as inclusion or competition, the sentiment attribute can be left blank or marked as neutral. All constructed knowledge triples are aggregated into a knowledge triple set, which serves as input for subsequent knowledge graph fusion.

[0077] 206. The knowledge triple set is fused with the existing knowledge graph, and conflicting information in the fusion process is resolved to obtain an updated domain knowledge graph. In this embodiment, when the number of associated triples of a newly added product node in the domain knowledge graph is lower than a preset threshold, the method further includes: identifying highly authoritative text content in the domain knowledge graph originating from opinion leader accounts or brand official accounts, and using the highly authoritative text content as expert knowledge samples; extracting attribute descriptions and efficacy descriptions related to the newly added product from the expert knowledge samples, and constructing an initial knowledge triple set for the newly added product through named entity recognition and relation extraction; retrieving mature product nodes under the same category in the domain knowledge graph based on the category tag of the newly added product, calculating the similarity of the newly added product with each mature product in terms of attribute features, and obtaining a set of similar products; using the associated triples of each product in the set of similar products as transfer candidate knowledge, filtering and replacing attribute values ​​of the transfer candidate knowledge based on attribute matching degree, and generating transfer knowledge triples adapted to the newly added product; merging the initial knowledge triple set with the transfer knowledge triple set to construct a cold start knowledge subgraph for the newly added product.

[0078] Specifically, the system encounters a cold start problem with newly added product nodes. When the number of associated triples in the domain knowledge graph for a newly added product node falls below a preset threshold, it means that the product lacks sufficient knowledge support. This situation is common on vertical e-commerce platforms, where new products typically lack user reviews and social media discussions, making it difficult for the system to extract enough knowledge triples from regular data sources. Without special handling for these products, the recommendation system will be unable to generate effective recommendations when calculating the path from user nodes to newly added product nodes due to missing associated edges, causing the new product to be excluded from the recommendation list.

[0079] To address the cold start problem, the system employs a dual-path knowledge construction strategy. The first path extracts expert knowledge from highly authoritative text content, while the second path transfers relevant knowledge from established products. These two paths complement each other, jointly constructing the initial knowledge base for new products.

[0080] The identification of highly authoritative text content is based on the identity verification information of the data source. Opinion leader accounts are usually verified by the platform, have a large number of followers and high interaction rates, and their reviews are highly professional and credible. Information published by official brand accounts comes directly from the product manufacturers or sellers, and is authoritative in terms of product attributes and efficacy descriptions. The system marks the source information of these accounts in the domain knowledge graph and filters their published text content as expert knowledge samples.

[0081] Although the number of expert knowledge samples is limited, their information density is high. The system extracts attribute descriptions and efficacy descriptions related to the new product from these samples. Attribute descriptions include objective information such as the product's composition, specifications, and design features, while efficacy descriptions cover the product's applicable scenarios and expected effects. By performing named entity recognition and relation extraction on the expert knowledge samples, the system can construct an initial set of knowledge triples for the new product. Although this set is small in size, its accuracy is high.

[0082] The core of the knowledge transfer path lies in leveraging the existing knowledge of mature products. Mature products typically have accumulated rich association triples in the knowledge graph, including user reviews, attribute characteristics, target audience, and other information. Based on the category tags of newly added products, the system retrieves mature product nodes within the same category from the domain knowledge graph. Category tags provide product classification information; using category tags can narrow the search scope and improve search efficiency.

[0083] Among the retrieved mature products, the system calculates the similarity between the new product and each mature product in terms of attribute features. Attribute features include the product's price range, functional positioning, and target user group. Similarity calculation is performed by comparing the attribute vectors of the two products, which are encoded by the attribute values ​​of each product. Mature products with high similarity are included in the similar product set. These products are similar to the new product in terms of functional positioning or usage scenarios, and their knowledge has transfer value.

[0084] The triples associated with each product in a set of similar products are considered candidate knowledge for transfer. However, this knowledge cannot be directly copied to new products because different products differ in specific attribute values. The system filters candidate knowledge based on attribute matching. Attribute matching measures whether the attributes in the candidate knowledge are applicable to the new product. If the attribute involved in a triple also exists in the new product, then the triple has transfer value; if the involved attribute does not exist in the new product or has been replaced, then the triple is filtered out.

[0085] The selected candidate knowledge still needs to undergo attribute value replacement. Attribute value replacement involves adjusting the specific values ​​or descriptions in the triples to match the corresponding information for the new product. The system identifies the content in the triples that is bound to specific attributes of the product and replaces it with the attribute values ​​of the new product, generating transfer knowledge triples adapted to the new product. This process ensures that the transferred knowledge is semantically consistent with the new product.

[0086] The initial set of knowledge triples and the transferred knowledge triples come from expert content and mature products, respectively, complementing each other in terms of knowledge source and coverage. The system merges these two parts of knowledge to construct a cold start knowledge subgraph for new products. The cold start knowledge subgraph is a local structure in the domain knowledge graph surrounding the nodes of new products, containing the product's attributes, effects, applicable scenarios, and relationships with other entities.

[0087] Once the cold-start knowledge subgraph is constructed, the recommendation system can access this subgraph when generating recommendations. When a user browses or searches for categories related to a new product, the graph traversal algorithm calculates the path length and path weight from the user node to the new product node. Because the cold-start knowledge subgraph already contains the product's attribute features, efficacy descriptions, and similarity relationships with mature products in the same category, the recommendation system can generate recommendation results based on these associated paths. Simultaneously, the system can use the attribute descriptions and efficacy information in the knowledge subgraph to construct recommendation reasons, explaining to the user why the new product is recommended. This mechanism allows new products to be effectively recommended even in the absence of user reviews, preventing them from being excluded from the recommendation list due to data sparsity.

[0088] 207. Obtain user information of the target user, including browsing history, purchase history, and favorited items; 208. Based on user information and entity relationships in the domain knowledge graph, determine candidate products and their association paths with users to obtain product recommendation results.

[0089] In this embodiment, data streams including user reviews, trending topics, and product information are collected from multiple data sources. The data streams are preprocessed to obtain standardized text data. Entity recognition, relation extraction, and sentiment analysis are performed on the standardized text data to obtain a set of knowledge triples containing entities, inter-entity relationships, and sentiment attributes. This set of knowledge triples is then fused with an existing knowledge graph, and conflicting information during the fusion process is resolved to obtain an updated domain knowledge graph. Based on user information and entity relationships in the domain knowledge graph, candidate products and their association paths with users are determined, resulting in product recommendations. This invention, by continuously fusing external multi-source dynamic information and performing incremental graph updates, enables the recommendation system to promptly perceive market changes, improving the timeliness and accuracy of recommendations.

[0090] The product recommendation method for a vertical e-commerce platform in this embodiment of the invention has been described above. The product recommendation device for a vertical e-commerce platform in this embodiment of the invention will be described below. Please refer to [link to relevant documentation] for details on this product recommendation device. Figure 3 One embodiment of the product recommendation device for a vertical e-commerce platform in this invention includes: The data acquisition module 301 is used to acquire data streams from multiple data sources, preprocess the data streams to obtain standardized text data, wherein the data streams include user comments, trending topics and product information from vertical e-commerce platforms; The knowledge extraction module 302 is used to perform entity recognition, relation extraction and sentiment analysis on the standardized text data to obtain a set of knowledge triples containing entities, relationships between entities and sentiment attributes. The graph fusion module 303 is used to fuse the knowledge triple set with the existing knowledge graph and resolve conflict information in the fusion process to obtain an updated domain knowledge graph. Information acquisition module 304 is used to acquire user information of the target user, including browsing history, purchase history and favorited items; The recommendation generation module 305 is used to determine candidate products and their association paths with users based on user information and entity relationships in the domain knowledge graph, and to obtain product recommendation results.

[0091] In this embodiment of the invention, the product recommendation device of the vertical e-commerce platform operates the aforementioned product recommendation method for the vertical e-commerce platform. The device collects data streams from multiple data sources, including user reviews, trending topics, and product information. The data streams are preprocessed to obtain standardized text data. Entity recognition, relation extraction, and sentiment analysis are performed on the standardized text data to obtain a set of knowledge triples containing entities, inter-entity relationships, and sentiment attributes. This set of knowledge triples is then fused with an existing knowledge graph, and conflicting information during the fusion process is resolved to obtain an updated domain knowledge graph. Based on user information and entity relationships in the domain knowledge graph, candidate products and their association paths with users are determined, resulting in product recommendation results. This invention, by continuously fusing external multi-source dynamic information and performing incremental graph updates, enables the recommendation system to promptly perceive market changes, improving the timeliness and accuracy of recommendations.

[0092] above Figure 3 The product recommendation device of the vertical e-commerce platform in the embodiments of the present invention will be described in detail from the perspective of unitized functional entities. The product recommendation device of the vertical e-commerce platform in the embodiments of the present invention will be described in detail from the perspective of hardware processing.

[0093] Figure 4This is a schematic diagram of the structure of a product recommendation device for a vertical e-commerce platform according to an embodiment of the present invention. The product recommendation device 400 for this vertical e-commerce platform can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 410 (e.g., one or more processors) and a memory 420, and one or more storage media 430 (e.g., one or more mass storage devices) for storing application programs 433 or data 432. The memory 420 and storage media 430 can be temporary or persistent storage. The program stored in the storage media 430 may include one or more units (not shown in the diagram), each unit may include a series of instruction operations on the product recommendation device 400 of the vertical e-commerce platform. Furthermore, the processor 410 may be configured to communicate with the storage media 430 and execute the series of instruction operations in the storage media 430 on the product recommendation device 400 of the vertical e-commerce platform to implement the steps of the product recommendation method of the vertical e-commerce platform described above.

[0094] The product recommendation device 400 of a vertical e-commerce platform may also include one or more power supplies 440, one or more wired or wireless network interfaces 450, one or more input / output interfaces 460, and / or one or more operating systems 431, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 4 The illustrated structure of the product recommendation device for a vertical e-commerce platform does not constitute a limitation on the product recommendation device for a vertical e-commerce platform provided by this invention. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0095] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the product recommendation method of the vertical e-commerce platform.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0097] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0098] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A product recommendation method for a vertical e-commerce platform, characterized in that, The product recommendation methods of the vertical e-commerce platform include: Data streams are collected from multiple data sources and preprocessed to obtain standardized text data, wherein the data streams include user comments, trending topics, and product information from vertical e-commerce platforms; Entity recognition, relation extraction, and sentiment analysis are performed on the standardized text data to obtain a set of knowledge triples containing entities, relationships between entities, and sentiment attributes. The knowledge triple set is fused with the existing knowledge graph, and conflicting information during the fusion process is resolved to obtain an updated domain knowledge graph. Obtain user information of the target user, including browsing history, purchase history, and favorited items; Based on user information and entity relationships in the domain knowledge graph, candidate products and their association paths with users are determined, resulting in product recommendation results.

2. The product recommendation method for a vertical e-commerce platform according to claim 1, characterized in that, The process of collecting data streams from multiple data sources and preprocessing the data streams to obtain standardized text data includes: Raw data streams are obtained from the multiple data sources, including social media platforms, vertical e-commerce platforms, and industry analysis websites, through API interfaces or web crawlers. The original data stream is pushed to the stream processing framework, where it is deduplicated and converted in format to obtain a pre-processed data stream. The text content in the pre-processed data stream is subjected to language type identification. The identified non-target language text is then input into a machine translation model for translation to obtain the target language text. The target language text is cleaned and noise filtered to obtain standardized text data.

3. The product recommendation method for a vertical e-commerce platform according to claim 1, characterized in that, The process of performing entity recognition, relation extraction, and sentiment analysis on the standardized text data to obtain a set of knowledge triples containing entities, inter-entity relationships, and sentiment attributes includes: The standardized text data is input into a pre-trained language model, and the entities in the text are labeled by the named entity recognition module to obtain an entity labeling sequence, wherein the entity labeling sequence includes product entities, brand entities, attribute entities and user entities; The semantic relationships between entity pairs in the entity annotation sequence are determined to obtain the relationship annotation results; Based on the text fragments containing attribute entities in the standardized text data, identify the user's sentiment towards each attribute and assign a sentiment polarity label to each attribute entity. Based on the entity annotation sequence, the relation annotation result, and the sentiment polarity label, a knowledge triplet containing a head entity, relation type, tail entity, and sentiment attribute is constructed to obtain a knowledge triplet set.

4. The product recommendation method for a vertical e-commerce platform according to claim 3, characterized in that, The step of identifying the user's sentiment towards each attribute based on text fragments containing attribute entities in the standardized text data, and assigning a sentiment polarity label to each attribute entity, includes: Extract text fragments containing attribute entities from the standardized text data, identify modifiers and evaluation words describing attribute features in the text fragments, and obtain a set of attribute-modifier pairs; The attribute-modifier pair set is input into the sentiment classification module, and the sentiment tendency of each modifier is classified and judged to obtain positive, negative or neutral sentiment polarity labels; The sentiment polarity tags are associated with the corresponding attribute entities. When there are multiple sentiment polarity tags for the same attribute entity, a weighted average is performed according to the weight of the modifiers in the text to obtain the comprehensive sentiment polarity tag of the attribute entity.

5. The product recommendation method for a vertical e-commerce platform according to claim 3, characterized in that, Before performing entity recognition, relation extraction, and sentiment analysis on the standardized text data, the following steps are also included: The standardized text data is source-labeled, and the text is labeled as social media context, e-commerce platform context, or professional forum context according to the platform type of the data source, to obtain context-labeled data; Professional terms are extracted from the context-annotated data. The co-occurring words and modifiers of each professional term in different contexts are statistically analyzed. The semantic vector representation of each professional term in different contexts is calculated to obtain a set of polysemous terms. For homographs in the term polysemy set whose semantic vector distance exceeds a preset threshold, multiple semantic nodes are created in the domain knowledge graph, and each semantic node is associated with a corresponding context label. The context label is used to select the term node with the corresponding semantic version for matching when performing entity recognition and relation extraction. When cross-contextual term references are detected, a mapping relationship between different semantic nodes is established in the knowledge graph. The mapping relationship is used to record the semantic differences of homographs in different contexts and to identify users' cross-contextual product preferences when generating recommendations, thereby preventing recommendation errors caused by term polysemy.

6. The product recommendation method for a vertical e-commerce platform according to claim 1, characterized in that, The process of fusing the knowledge triple set with the existing knowledge graph and resolving conflicting information during the fusion process to obtain the updated domain knowledge graph includes: The similarity between entities in the knowledge triple set and entities in the existing knowledge graph is calculated. When the similarity exceeds a preset threshold, they are determined to be the same entity, and the entity alignment mapping relationship is obtained. Based on the entity alignment mapping relationship, new triples in the knowledge triple set are added to the existing knowledge graph in an atomic operation manner, or existing triples are updated. Detect inconsistent information provided by multiple data sources on the same attribute of the same entity to obtain a set of conflict information; For each conflict item in the conflict information set, the comprehensive confidence level is calculated based on the authority score of each data source and the corresponding timestamp. The knowledge triple with the highest comprehensive confidence level is selected to replace the corresponding conflict item, resulting in an updated domain knowledge graph.

7. The product recommendation method for a vertical e-commerce platform according to claim 6, characterized in that, Following the obtained updated domain knowledge graph, the following is also included: When the number of associated triples of newly added product nodes in the domain knowledge graph is lower than a preset threshold, highly authoritative text content from opinion leader accounts or brand official accounts in the domain knowledge graph is identified and used as expert knowledge samples. Extract attribute descriptions and efficacy descriptions related to the newly added product node from the expert knowledge sample, and construct an initial knowledge triplet set for the newly added product node through named entity recognition and relation extraction; Based on the category tag of the newly added product node, mature product nodes of the same category are retrieved in the domain knowledge graph, and the similarity of the newly added product node with each mature product node in terms of attribute features is calculated to obtain a set of similar products. The association triples of each product in the similar product set are used as candidate knowledge for migration. The candidate knowledge is filtered and the attribute values ​​are replaced according to the attribute matching degree to generate a migration knowledge triple that is adapted to the newly added product node. The initial set of knowledge triples is combined with the transferred knowledge triples to construct the cold start knowledge subgraph of the newly added product node.

8. The product recommendation method for a vertical e-commerce platform according to claim 6, characterized in that, The step of calculating the comprehensive confidence level for each conflict item in the conflict information set based on the authority rating of each data source and the corresponding timestamp, and selecting the knowledge triple with the highest comprehensive confidence level to replace the corresponding conflict item includes: Obtain multiple candidate knowledge triples corresponding to each conflict item in the conflict information set, as well as the authority score and timestamp of the data source to which each candidate knowledge triple belongs; The time decay coefficient is calculated based on the time difference between the current time and the timestamps of each candidate knowledge triplet. The larger the time difference, the smaller the time decay coefficient. The overall confidence level of each candidate knowledge triple is obtained by multiplying the authority score of the data source to which each candidate knowledge triple belongs with the corresponding time decay coefficient. The candidate knowledge triple with the highest overall confidence is selected as the retained information and used to replace the corresponding conflict item in the conflict information set.

9. The product recommendation method for a vertical e-commerce platform according to claim 1, characterized in that, The step of determining candidate products and their association paths with users based on user information and entity relationships in the domain knowledge graph, and obtaining product recommendation results, includes: From the domain knowledge graph, identify trend nodes that are newly added or have significantly increased weight within a preset time window to obtain a set of hot trend nodes; Based on the user information, the corresponding user node in the domain knowledge graph is located. The path length and path weight from the user node to each product node are calculated by the graph traversal algorithm to obtain the candidate product set and the corresponding graph distance score. For products in the candidate product set that are associated with the hot trend set, the graph distance scores are weighted and sorted according to the association strength, and a preset number of products with the highest scores are selected as recommended products. Extract the association paths from the user nodes to each recommended product, convert the association paths into natural language descriptions, and obtain product recommendation results that include recommended products and reasons for recommendation.

10. A product recommendation device for a vertical e-commerce platform, characterized in that, The product recommendation device of the vertical e-commerce platform includes: The data acquisition module is used to collect data streams from multiple data sources, preprocess the data streams to obtain standardized text data, wherein the data streams include user comments, trending topics and product information from vertical e-commerce platforms; The knowledge extraction module is used to perform entity recognition, relation extraction and sentiment analysis on the standardized text data to obtain a set of knowledge triples containing entities, relationships between entities and sentiment attributes. The knowledge graph fusion module is used to fuse the knowledge triple set with the existing knowledge graph and resolve conflicting information during the fusion process to obtain an updated domain knowledge graph. The information acquisition module is used to acquire user information of the target user, including browsing history, purchase history and favorited items. The recommendation generation module is used to determine candidate products and their association paths with users based on user information and entity relationships in the domain knowledge graph, and to obtain product recommendation results.

11. A product recommendation device for a vertical e-commerce platform, characterized in that, The product recommendation device of the vertical e-commerce platform includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the product recommendation device of the vertical e-commerce platform to perform the steps of the product recommendation method of the vertical e-commerce platform as described in any one of claims 1-9.

12. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements the steps of the product recommendation method for a vertical e-commerce platform as described in any one of claims 1-9.

Images