An ai product selection management platform with dynamic knowledge base updating

Abstract

The application relates to the technical field of artificial intelligence recommendation, and discloses an AI product selection management platform with dynamic knowledge base updating, which comprises the following steps: collecting multi-source product data to construct a dynamic product knowledge graph, performing incremental updating based on influence domain analysis, and only re-embedding calculation is performed on the subgraph related to the changed entity to reduce the updating cost; analyzing user natural language requirements to generate a structured query vector; adopting a multi-path recall strategy of semantic matching, collaborative filtering and knowledge graph path to obtain candidate products; performing normalized comparison and analysis on the candidate products in multiple dimensions such as function, performance and price, and combining the user demand weight to perform weighted comprehensive score sorting to generate a personalized recommendation list and an interpretable recommendation reason; obtaining user interaction feedback and synchronously updating a user preference model and a knowledge graph to form a closed loop optimization. The application reduces the maintenance cost of the knowledge graph, realizes fine-grained product comparison, and significantly enhances the explainability and self-adaptive optimization capability of the recommendation result.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and information recommendation technology, specifically to an AI product selection management platform with dynamic knowledge base updates, which can be applied to scenarios such as e-commerce, intelligent shopping guides, and product selection that require intelligent comparison and personalized recommendations of multiple products. Background Technology

[0002] With the rapid development of e-commerce and online shopping platforms, the number of products is growing exponentially, posing a significant challenge to users in quickly and accurately finding suitable products that meet their needs from this vast array of options. Recommendation systems, as an effective means of alleviating information overload, have been widely adopted on various platforms.

[0003] However, existing product recommendation systems still have the following shortcomings: First, there's the issue of static knowledge bases. Most existing recommendation systems rely on pre-built static knowledge graphs or product databases, resulting in delayed product information updates and an inability to promptly reflect dynamic information such as changes in product parameters, price fluctuations, and user reviews. When new products are launched or existing product information changes, the system needs to expend significant resources to rebuild or fully update the knowledge graph, leading to inefficiency.

[0004] Second, there is a lack of granular product comparison capabilities. Existing recommendation systems typically rely solely on collaborative filtering based on users' historical behavior or content-based recommendations, making it difficult to support users' needs for refined comparisons of multiple candidate products across multiple dimensions. When purchasing products, users often need to comprehensively compare multiple dimensions such as functionality, performance, price, and reputation, and existing systems struggle to provide this structured and visualized comparison service.

[0005] Third, the interpretability of recommendation results is insufficient. Deep learning-based recommendation models typically operate in a "black box" manner, meaning users cannot understand the basis for generating the recommendation results, which reduces users' trust in the recommendation system and their willingness to adopt it.

[0006] Fourth, the understanding of user needs is not precise enough. Existing systems mostly rely on passive recommendations based on users' historical behavior data, lacking the ability to proactively understand and accurately match users' current specific needs, making it difficult to handle "cold start" scenarios where users are visiting for the first time or have no historical behavior data.

[0007] To address the aforementioned issues, some studies have attempted to introduce knowledge graphs into recommender systems; however, most solutions still suffer from high knowledge update costs, limited comparison dimensions, and a disconnect between recommendation and comparison. Therefore, there is an urgent need for an intelligent product comparison and recommendation system that can dynamically update the knowledge base, support multi-dimensional product comparison, and possess interpretability. Summary of the Invention

[0008] This invention aims to overcome the shortcomings of existing technologies and provide an AI product selection management platform with dynamic knowledge base updates. By constructing a dynamically updatable product knowledge graph, a multi-dimensional product comparison engine, and an intelligent recommendation module based on a large language model, it achieves an integrated solution for real-time product knowledge updates, multi-dimensional refined comparisons, and personalized intelligent recommendations.

[0009] To achieve the above-mentioned objectives, the present invention provides the following technical solution: This invention provides an AI product selection management platform with dynamic knowledge base updates, characterized by comprising the following steps: S1: Collect product data from multiple data sources, clean, standardize and structure the product data, and extract the core attribute information of the products; S2: Based on the product data processed in step S1, construct a dynamic product knowledge graph with product entities as core nodes, and perform vector embedding on the knowledge graph; when product information changes or new product data is detected, perform incremental update operation, the incremental update includes: identifying the product entities and related subgraphs involved in the changed data, recalculating and updating only the nodes and edges within the scope of the change, and keeping the knowledge graph embedding of the unaffected parts unchanged; S3: Obtain the product query requirements input by the user, perform intent parsing on the user input, extract the core requirement elements of the user, and transform the requirement elements into a structured query vector; S4: Based on the query vector obtained in step S3 and the dynamic knowledge graph constructed in step S2, a multi-path recall strategy is adopted to recall a set of candidate products that match the user's needs from the knowledge graph. S5: For each product in the candidate product set recalled in step S4, perform comparative analysis on multiple preset dimensions to generate structured product comparison results. S6: Based on the comparative analysis results of step S5, and combined with the weight of user needs, the candidate product set is comprehensively scored and ranked to generate a personalized recommendation list and explanation of the recommendation reasons; S7: Obtain user interaction feedback on the recommendation results, update the user preference model based on the interaction feedback, and feed the new knowledge contained in the interaction feedback into the knowledge graph update mechanism in step S2.

[0010] Furthermore, the construction of the dynamic product knowledge graph in step S2 of the present invention specifically includes: S2.1: The preprocessed product data is subjected to entity recognition and relationship extraction according to the preset graph pattern to generate an initial product knowledge graph. The knowledge graph includes product entity nodes, attribute nodes, category nodes, brand nodes and their relationships. S2.2: A graph convolutional neural network is used to embed vectors into the initial knowledge graph, generating a fixed-dimensional vector representation for each entity node; S2.3: When a change in product information or the addition of new product data is detected in the data source, the knowledge graph is expanded outward by a preset number of neighboring nodes, with the entity that has changed as the center. The nodes and edges within the expansion range are defined as the influence domain. Only the knowledge graph elements within the influence domain are re-embedded and recalculated, and the recalculated embedding vector is updated to the corresponding position in the entity vector matrix.

[0011] Furthermore, the intent parsing described in step S3 of the present invention specifically includes: Semantic parsing is performed on the user's natural language input to extract the user's core needs, which include: target product category, functional requirements, performance requirements, price range, and brand preference. The parsed user demand elements are transformed into structured query vectors according to preset encoding rules. The query vectors are located in the same vector space as the vector representations of product entities in the knowledge graph.

[0012] Furthermore, the multi-path recall strategy in step S4 of the present invention includes: First-line recall - semantic matching recall: calculate the semantic similarity between the user demand vector and the product entity vector in the knowledge graph, and select products with similarity higher than the first preset threshold as candidates; The second approach to recall is collaborative filtering recall: based on users' historical behavior data, products similar to users' historical preferences are recalled through collaborative filtering algorithms. The third approach to recall is knowledge graph path recall: searching for the associated paths connecting user demand entities and product entities in the knowledge graph, and recalling products based on the semantic relevance of the paths. The three recall results are deduplicated and merged to generate a candidate product set.

[0013] Furthermore, the semantic similarity calculation described in this invention employs a text embedding method based on a large language model, encoding the user requirement text and product description text into semantic vectors respectively, and determining the degree of semantic matching by calculating the cosine similarity of the vectors.

[0014] Furthermore, the preset comparison dimensions mentioned in step S5 of the present invention include: functional dimension, performance dimension, price dimension, quality dimension, reputation dimension, and service dimension; The comparative analysis specifically includes: S5.1: Extract the feature values ​​of each product in each comparison dimension from the knowledge graph; S5.2: Normalize the feature values ​​of different dimensions to eliminate dimensional differences; S5.3: Calculate the similarity scores between each pair of candidate products in each dimension, as well as the overall similarity score; S5.4: Based on the comparison results, identify the advantages and disadvantages of each product relative to other candidate products.

[0015] Furthermore, the comprehensive score in step S6 of the present invention adopts a weighted summation model:

[0016] in, Let be the overall score of the i-th product, and D be the total number of comparison dimensions. Let d be the weight of user demand in the d-th dimension. Let be the normalized score of the i-th product on the d-th dimension; sort the candidate products in descending order according to the comprehensive score, and select the top N products as the recommendation results.

[0017] Furthermore, the user demand weight described in this invention is dynamically adjusted based on the attention paid to each dimension during user interaction, and the attention is quantified by the user's dwell time, number of clicks, and details expansion behavior in each dimension.

[0018] Furthermore, the interactive feedback in step S7 of the present invention includes: user clicks, browsing, favorites, comparisons and purchases of the recommendation results, as well as user-submitted evaluations and preference adjustments; The method for updating the user preference model based on the interactive feedback is as follows: the product features corresponding to the positive feedback signal are used as positive samples, and the product features corresponding to the negative feedback signal are used as negative samples, and the user preference vector is incrementally learned and updated.

[0019] Furthermore, the method of the present invention also includes: presenting the structured product comparison results generated in step S5 to the user in a visualization manner such as a radar chart or a parallel coordinate graph, so that the user can intuitively compare the performance differences of each product in different dimensions.

[0020] Compared with the prior art, the present invention has the following beneficial technical effects: 1. Achieve low-cost dynamic updates of the knowledge base. This invention utilizes an incremental update mechanism based on subgraph influence domain analysis. When the knowledge graph changes, only the affected parts are recalculated, eliminating the need for a full reconstruction of the knowledge graph. This significantly reduces the computational cost and time overhead of knowledge updates. Actual testing shows that, for product knowledge graphs of the same size, the incremental update time of this invention is only 10% to 20% of that of traditional full-scale update methods.

[0021] 2. Provides fine-grained, multi-dimensional product comparison capabilities. This invention, by constructing a structured, multi-dimensional comparison and analysis framework, enables refined comparison of candidate products across multiple dimensions such as functionality, performance, price, quality, reputation, and service. This solves the problem that existing recommendation systems can only output recommendation lists without providing detailed comparison criteria, significantly improving the efficiency of product selection for users.

[0022] 3. Enhanced interpretability of recommendation results. This invention uses knowledge graphs as the basis for recommendation reasoning. Recommendation results can be traced back to specific entities and relationship paths within the knowledge graph, while generating structured explanations of recommendation reasons. This allows users to understand "why this product is recommended," improving the transparency of the recommendation system and user trust.

[0023] 4. Achieved accurate understanding and matching of user needs. This invention, through natural language intent parsing technology, can accurately understand users' product needs expressed in natural language and transform them into structured query vectors for precise matching, effectively solving the problem of traditional recommendation systems' inability to make effective recommendations in cold start scenarios.

[0024] 5. A data-driven closed-loop optimization system was constructed. This invention uses a user interaction feedback mechanism to feed back user behavior data such as clicks, browsing, and purchases to the knowledge graph update module and user preference model in real time, forming a closed loop of "collection-update-recommendation-feedback-re-update". This allows the system to continuously optimize itself as users use it, and the recommendation accuracy continues to improve over time. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of the AI ​​product selection management platform with dynamic knowledge base updates according to the present invention. Figure 2 This is a flowchart illustrating the dynamic product knowledge graph incremental update method of the present invention. Figure 3 This is a schematic diagram of the multi-path recall strategy for candidate products of the present invention; Figure 4 This is a schematic diagram of the multi-dimensional product comparison and analysis framework of the present invention; Figure 5 This is a schematic diagram of the process for generating and interacting with the feedback loop of the recommended results in this invention. Detailed Implementation

[0026] like Figures 1-5 As shown, this invention provides an AI product selection management platform with dynamic knowledge base updates, comprising the following steps: Step S1: Multi-source product data collection and preprocessing. Collect product data from multiple data sources, including but not limited to: e-commerce platform product pages, official product specifications, user reviews and ratings, price tracking data, and social media discussions. Clean, standardize, and structure the collected product data to extract core product attribute information. These core attributes include: product identification information, functional parameters, performance indicators, price information, brand information, and user ratings and review summaries.

[0027] Step S2: Dynamic Product Knowledge Graph Construction and Update. Based on the product data processed in Step S1, a product knowledge graph is constructed with product entities as core nodes. The knowledge graph includes product entity nodes, attribute nodes, category nodes, brand nodes, and the relationships between them.

[0028] The construction of knowledge graphs specifically includes: S2.1 Initial Knowledge Graph Construction: The preprocessed product data is used to perform entity recognition and relationship extraction according to the preset knowledge graph pattern to generate an initial product knowledge graph; S2.2 Incremental Update Mechanism: When changes in product information or new product data are detected in the data source, an incremental update operation is performed. The incremental update includes: identifying the product entities and related subgraphs involved in the changed data, recalculating and updating only the nodes and edges within the scope of the change's impact, and keeping the knowledge graph embedding of the unaffected parts unchanged. S2.3 Dynamic Embedding Update: The updated knowledge graph is updated by vector embedding. A graph neural network is used to encode the newly added or changed entities and their associated subgraphs to generate updated entity vector representations. These updated entity vector representations are then combined with the vector representations of the original unchanged parts to form the updated complete knowledge graph embedding.

[0029] Step S3: User Needs Understanding and Intent Parsing. Obtain the user's product query needs, including: the user's natural language description, the user's selected filtering conditions, and the user's browsing and purchase history. Perform intent parsing on the user input, specifically including: semantic parsing of the user's natural language input to extract the user's core needs elements, including: target product category, functional requirements, performance requirements, price range, and brand preference; transform the parsed user needs elements into a structured query vector, which serves as input for subsequent product recall and comparison.

[0030] Step S4: Intelligent Recall of Candidate Products. Based on the user demand vector obtained in Step S3 and the dynamic knowledge graph constructed in Step S2, a set of candidate products matching the user demand is recalled from the knowledge graph. The candidate product recall employs a multi-path recall strategy: First-line recall - semantic matching recall: calculate the semantic similarity between the user demand vector and the product entity vector in the knowledge graph, and select products with similarity higher than the first preset threshold as candidates; The second approach to recall is collaborative filtering recall: based on users' historical behavior data, products similar to users' historical preferences are recalled through collaborative filtering algorithms. The third approach to recall – knowledge graph path recall: Searching for the associated paths connecting user demand entities and product entities in the knowledge graph, and recalling products based on the semantic relevance of the paths.

[0031] The three recall results are deduplicated and merged to generate a candidate product set.

[0032] Step S5: Multi-dimensional product comparison analysis. For each product in the candidate product set recalled in Step S4, a comparison analysis is performed on multiple preset dimensions to generate structured product comparison results.

[0033] The comparison dimensions include, but are not limited to: functional dimension: comparison of the functional features of the products; performance dimension: comparison of the key performance indicators of the products; price dimension: comparison of product price and cost-effectiveness; quality dimension: comparison of product quality, durability, and reliability; reputation dimension: comparison of user ratings and evaluation sentiment; and service dimension: comparison of after-sales service and warranty policies.

[0034] The comparative analysis specifically includes: S5.1 Feature Extraction: Extract feature values ​​for each product across all comparison dimensions from the knowledge graph; S5.2 Normalization: Normalize the feature values ​​across different dimensions to eliminate dimensional differences; S5.3 Similarity Calculation: Calculate the pairwise similarity scores between candidate products across all dimensions, as well as the overall similarity score; S5.4 Advantages and Disadvantages Identification: Based on the comparison results, identify the advantages and disadvantages of each product relative to other candidate products.

[0035] Step S6: Intelligent recommendation ranking and generation. Based on the comparison and analysis results of step S5, and combined with the user demand weight, the candidate product set is comprehensively scored and ranked to generate a personalized recommendation list.

[0036] The overall score uses a weighted summation model:

[0037] in, Let be the overall score of the i-th product, and D be the total number of comparison dimensions. Let d be the weight of user demand in the d-th dimension. Let be the normalized score of the i-th product on the d-th dimension; sort the candidate products in descending order according to the comprehensive score, and select the top N products as the recommendation results.

[0038] Simultaneously, a recommendation reason explanation is generated, which includes: the product's performance in each comparison dimension, its matching points with user needs, and its advantages compared to other candidate products.

[0039] Step S7: Interactive feedback and knowledge base closed-loop update. The recommendation results generated in step S6 are displayed to the user, and the user's interactive feedback is obtained. The feedback includes the user's clicks, browsing, collection, comparison, and purchase behavior on the recommendation results, as well as the user's actively submitted evaluations and preference adjustments.

[0040] The user preference model is updated based on user feedback, and the new knowledge contained in the user feedback is fed back into the knowledge graph update mechanism in step S2, forming a closed-loop system of "data collection - graph update - recall comparison - recommendation generation - feedback learning".

[0041] Preferably, the incremental update mechanism in step S2.2 adopts a subgraph influence domain analysis method, specifically: when a product entity is updated, the knowledge graph expands outward by a preset number of neighboring nodes with the entity as the center, and the nodes and edges within the expansion range are defined as the influence domain, and only the knowledge graph elements within the influence domain are re-embedded and calculated.

[0042] Preferably, the semantic similarity calculation in step S4 adopts a text embedding method based on a large language model, which encodes the user requirement text and product description text into semantic vectors respectively, and determines the degree of semantic matching by calculating the cosine similarity of the vectors.

[0043] Preferably, the multi-dimensional comparison analysis in step S5 uses a radar chart or parallel coordinate graph to present the comparison results, making it easier for users to intuitively compare the performance differences of each product in different dimensions.

[0044] Preferably, the user requirements described in step S6 The weights are dynamically adjusted based on the user's attention to each dimension during the interaction process. The attention is quantified by the user's dwell time, number of clicks, and actions such as expanding details on each dimension.

[0045] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to illustrate the present invention and should not be regarded as limiting the scope of the present invention.

[0046] Example 1: System Overall Architecture like Figure 1As shown, the AI ​​product selection management platform with dynamic knowledge base updates provided by this invention mainly includes the following functional modules: data collection and preprocessing module, dynamic knowledge graph construction and updating module, user needs understanding and intent parsing module, candidate product intelligent recall module, multi-dimensional product comparison and analysis module, intelligent recommendation sorting and generation module, and interactive feedback and knowledge base closed-loop update module.

[0047] The data flow between modules is as follows: the data acquisition module collects raw product data from a multi-source data platform, and after preprocessing, it is input into the dynamic knowledge graph construction module to generate a product knowledge graph; when a user initiates a query request, the demand understanding module parses the user's intent and generates a query vector; the recall module recalls candidate products from the knowledge graph; the comparison module performs multi-dimensional analysis on the candidate products; the recommendation module generates a sorted recommendation list and outputs it to the user; user feedback is fed back to the knowledge graph and user preference model through the closed-loop update module.

[0048] Example 2: Incremental Update of Dynamic Knowledge Graph like Figure 2 As shown, this embodiment describes in detail the incremental update method for dynamic product knowledge graphs.

[0049] Step S2.1: Initial map construction Taking electronic product recommendation as an example, structured data of mobile phone products is collected from multiple e-commerce platforms, including: product model, brand, processor model, screen size, camera parameters, battery capacity, storage capacity, price, user ratings, etc. Entity recognition is performed according to a preset graph pattern to establish product entity nodes, attribute nodes, category nodes, brand nodes, and the relationships between them.

[0050] A graph convolutional neural network is used to embed vectors into the initial knowledge graph, generating a fixed-dimensional vector representation for each entity node and storing the entity vector matrix and relation vector matrix.

[0051] Step S2.2: Incremental Update Mechanism When the system detects that the price of "iPhone 15 Pro" in a certain data source has changed from $999 to $899, it performs the following incremental update operation: (1) Locate the changed entity: Locate the entity node “iPhone_15_Pro” in the knowledge graph; (2) Determine the domain of influence: Taking the “iPhone_15_Pro” node as the center, expand outwards to two layers of neighboring nodes to determine the domain of influence, which includes the entity node itself, its price attribute node “Price_999” and its related hierarchical nodes; (3) Incremental calculation: Only nodes and edges within the influence domain are re-embedded and calculated, while the vector embeddings of other unaffected parts of the knowledge graph remain unchanged; (4) Update storage: Update the recalculated embedding vector to the corresponding position in the entity vector matrix.

[0052] Through the above incremental update method, the update calculation for a single product information change only involves about 5% to 15% of the nodes in the knowledge graph, which greatly reduces the update cost -8.

[0053] Step S2.3: Dynamic Embedded Update For newly added product entities, the following embedded update process is executed: (1) Create new entity nodes and associated edges in the knowledge graph based on the attribute information of the newly added products; (2) Using the trained graph convolutional neural network parameters, with the subgraph of the new node and its neighboring nodes as input, calculate the initial embedding vector of the new node; (3) Add the embedding vector of the newly added node to the entity vector matrix to complete the dynamic expansion of the knowledge graph.

[0054] Example 3: Multi-channel recall of candidate products like Figure 3 As shown in the figure, this embodiment describes in detail the multi-path recall strategy for intelligent recall of candidate products.

[0055] For example, if a user inputs "I want to buy a phone with good camera, long battery life, and a price around 5000 yuan": First-line recall – semantic matching recall: (1) Input the user query text into the pre-trained large language model to generate the query semantic vector Q. (2) Calculate the cosine similarity between Q and the entity vectors of all mobile phone products in the knowledge graph; (3) Select the top 50 products with the highest similarity as the semantic matching recall results.

[0056] Second-stage recall – Collaborative filtering recall: (1) Obtain the user's browsing, favorites, and purchase history; (2) Based on the collaborative filtering algorithm for items, calculate other products similar to the products that the user has interacted with in the past; (3) Select the top 30 products with the highest similarity as the collaborative filtering recall results.

[0057] Third-party recall – Knowledge graph path recall: (1) Locate the entity nodes in the knowledge graph that correspond to the demand elements such as "taking photos", "battery life" and "price"; (2) Search for the association path between the demand entity and the product entity; (3) Based on the semantic relevance and length of the path, the path is scored, and the 30 products with the highest scores are selected as the path recall results.

[0058] The three recall results are merged and deduplicated to obtain the final set of candidate products.

[0059] Example 4: Multi-dimensional Product Comparison Analysis like Figure 4 As shown, this embodiment describes in detail the implementation method of multi-dimensional product comparison analysis.

[0060] Suppose the candidate product set contains three mobile phone products: Product A, Product B, and Product C. The system extracts feature values ​​for each product across six dimensions from the knowledge graph:

[0061] Step S5.1 Feature Extraction: Read the original feature values ​​of each product in each dimension from the knowledge graph.

[0062] Step S5.2 Normalization: The Min-Max normalization method is used to map the feature values ​​of each dimension to the [0,1] interval. Taking the price dimension as an example (lower prices are better, inverse normalization is used):

[0063] Step S5.3 Similarity Calculation: Calculate the overall similarity between each pair of products:

[0064] Step S5.4 Strengths and Weaknesses Identification: Based on the normalized scores of each dimension, identify the strengths (dimensions with scores higher than the mean of the candidate set) and weaknesses (dimensions with scores lower than the mean of the candidate set) of each product, and generate a structured strengths and weaknesses analysis report.

[0065] Example 5: Intelligent Recommendation Ranking and Interactive Feedback like Figure 5 As shown in the figure, this embodiment describes in detail the process of generating recommendation results and updating them in a closed loop.

[0066] Recommended sorting: Assume the weights of user needs expressed through historical behavior and current interaction are as follows: Function weight 0.25, Performance weight 0.20, Price weight 0.30, Quality weight 0.10, Reputation weight 0.10, and Service weight 0.05.

[0067] Calculate the score for each product based on the comprehensive scoring formula:

[0068] Similarly, calculate the scores for product B and product C, and generate a recommendation list by sorting them in descending order of scores.

[0069] Recommendation reason generation: For the top-ranked product, the recommendation reason is: "This product has an advantage in terms of price and performs excellently in terms of functionality, which highly matches your need for 'good photography'." Interactive feedback and closed-loop updates: After displaying the recommendation results to the user, the system collects the user's interaction behavior in real time: (1) If a user clicks to view the details of a product, the system records the product as a positive feedback signal; (2) If a user adds a product to the comparison list, the system records that the product as a high-interest signal; (3) If a user ultimately purchases a product, the system records that the product as a strong positive feedback signal; (4) If a user quickly skips a product in the recommendation list, the system records that product as a negative feedback signal.

[0070] The aforementioned feedback signal is used to: (a) update the weights of each dimension in the user preference vector. (b) Input new factual data into the knowledge graph update module to enrich the relationships between product entities; (c) Optimize the similarity threshold and ranking model parameters in the recall strategy.

[0071] Through the aforementioned closed-loop mechanism, the system can continuously optimize recommendation accuracy as the frequency of user use increases, achieving an adaptive recommendation effect that becomes more accurate with repeated use.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI product selection management platform with dynamic knowledge base updates, characterized in that, Includes the following steps: S1: Collect product data from multiple data sources, clean, standardize and structure the product data, and extract the core attribute information of the products; S2: Based on the product data processed in step S1, construct a dynamic product knowledge graph with product entities as core nodes, and perform vector embedding on the knowledge graph; when product information changes or new product data is detected, perform incremental update operation, the incremental update includes: identifying the product entities and related subgraphs involved in the changed data, recalculating and updating only the nodes and edges within the scope of the change, and keeping the knowledge graph embedding of the unaffected parts unchanged; S3: Obtain the product query requirements input by the user, perform intent parsing on the user input, extract the core requirement elements of the user, and transform the requirement elements into a structured query vector; S4: Based on the query vector obtained in step S3 and the dynamic knowledge graph constructed in step S2, a multi-path recall strategy is adopted to recall a set of candidate products that match the user's needs from the knowledge graph. S5: For each product in the candidate product set recalled in step S4, perform comparative analysis on multiple preset dimensions to generate structured product comparison results. S6: Based on the comparative analysis results of step S5, and combined with the weight of user needs, the candidate product set is comprehensively scored and ranked to generate a personalized recommendation list and explanation of the recommendation reasons; S7: Obtain user interaction feedback on the recommendation results, update the user preference model based on the interaction feedback, and feed the new knowledge contained in the interaction feedback into the knowledge graph update mechanism in step S2.

2. The AI ​​product selection management platform with dynamic knowledge base updates according to claim 1, characterized in that, The construction of the dynamic product knowledge graph in step S2 specifically includes: S2.1: The preprocessed product data is subjected to entity recognition and relationship extraction according to the preset graph pattern to generate an initial product knowledge graph. The knowledge graph includes product entity nodes, attribute nodes, category nodes, brand nodes and their relationships. S2.2: A graph convolutional neural network is used to embed vectors into the initial knowledge graph, generating a fixed-dimensional vector representation for each entity node; S2.3: When a change in product information or the addition of new product data is detected in the data source, the knowledge graph is expanded outward by a preset number of neighboring nodes, with the entity that has changed as the center. The nodes and edges within the expansion range are defined as the influence domain. Only the knowledge graph elements within the influence domain are re-embedded and recalculated, and the recalculated embedding vector is updated to the corresponding position in the entity vector matrix.

3. The AI ​​product selection management platform with dynamic knowledge base updates according to claim 1, characterized in that, The intent parsing described in step S3 specifically includes: Semantic parsing is performed on the user's natural language input to extract the user's core needs, which include: target product category, functional requirements, performance requirements, price range, and brand preference. The parsed user demand elements are transformed into structured query vectors according to preset encoding rules. The query vectors are located in the same vector space as the vector representations of product entities in the knowledge graph.

4. The AI ​​product selection management platform with dynamic knowledge base updates according to claim 1, characterized in that, The multi-path recall strategy mentioned in step S4 includes: First-line recall - semantic matching recall: calculate the semantic similarity between the user demand vector and the product entity vector in the knowledge graph, and select products with similarity higher than the first preset threshold as candidates; The second approach to recall is collaborative filtering recall: based on users' historical behavior data, products similar to users' historical preferences are recalled through collaborative filtering algorithms. The third approach to recall is knowledge graph path recall: searching for the associated paths connecting user demand entities and product entities in the knowledge graph, and recalling products based on the semantic relevance of the paths. The three recall results are deduplicated and merged to generate a candidate product set.

5. The AI ​​product selection management platform with dynamic knowledge base updates according to claim 4, characterized in that, The semantic similarity calculation adopts a text embedding method based on a large language model, which encodes the user requirement text and product description text into semantic vectors respectively, and determines the degree of semantic matching by calculating the cosine similarity of the vectors.

6. The AI ​​product selection management platform with dynamic knowledge base updates according to claim 1, characterized in that, The preset comparison dimensions mentioned in step S5 include: functional dimension, performance dimension, price dimension, quality dimension, reputation dimension, and service dimension; The comparative analysis specifically includes: S5.1: Extract the feature values ​​of each product in each comparison dimension from the knowledge graph; S5.2: Normalize the feature values ​​of different dimensions to eliminate dimensional differences; S5.3: Calculate the similarity scores between each pair of candidate products in each dimension, as well as the overall similarity score; S5.4: Based on the comparison results, identify the advantages and disadvantages of each product relative to other candidate products.

7. The AI ​​product selection management platform with dynamic knowledge base updates according to claim 1, characterized in that, The comprehensive score mentioned in step S6 adopts a weighted summation model: in, Let be the overall score of the i-th product, and D be the total number of comparison dimensions. Let d be the weight of user demand in the d-th dimension. Let be the normalized score of the i-th product on the d-th dimension; Candidate products are sorted in descending order based on their overall scores, and the top N products are selected as the recommended results.

8. The AI ​​product selection management platform with dynamic knowledge base updates according to claim 7, characterized in that, The user demand weights are dynamically adjusted based on the user's attention to each dimension during the user interaction process. The attention is quantified by the user's dwell time, number of clicks, and actions of expanding details on each dimension.

9. The AI ​​product selection management platform with dynamic knowledge base updates according to claim 1, characterized in that, The interactive feedback mentioned in step S7 includes: user clicks, browsing, favorites, comparisons and purchases of the recommendation results, as well as user-submitted evaluations and preference adjustments; The method for updating the user preference model based on the interactive feedback is as follows: the product features corresponding to the positive feedback signal are used as positive samples, and the product features corresponding to the negative feedback signal are used as negative samples, and the user preference vector is incrementally learned and updated.

10. The AI ​​product selection management platform with dynamic knowledge base updates according to claim 1, characterized in that, The method further includes: presenting the structured product comparison results generated in step S5 to the user in a visual manner using a radar chart or parallel coordinate graph, so that the user can intuitively compare the performance differences of each product in different dimensions.

Images