Chemical and plastic product recommendation method and system based on user requirement

Abstract

The present invention relates to the technical field of chemical and plastic product recommendation. Provided are a chemical and plastic product recommendation method and system based on a user requirement. The method comprises: S1, constructing a knowledge graph of the chemical and plastic industry; S2, acquiring historical user data and the current user requirement, wherein the historical user data comprises historical search data, historical browsing data, historical interaction data and historical transaction data; S3, preprocessing the historical user data, and analyzing the preprocessed historical user data, so as to extract a historical requirement feature; S4, using a semantic understanding model to perform keyword extraction on the current user requirement, and in combination with the knowledge graph of the chemical and plastic industry, generating an optimized query condition; S5, using the optimized query condition to search a chemical and plastic product database, so as to obtain a first list; and S6, using the knowledge graph of the chemical and plastic industry to perform personalized ranking of the first list, taking the top k items to generate a recommendation list, and presenting the recommendation list to a user. The present invention can improve the relevance and accuracy of recommendations, thereby improving the user experience.

Description

Chemical and plastic product recommendation method and system based on user demand TECHNICAL FIELD

[0001] The present application relates to the technical field of chemical and plastic product recommendation, and particularly relates to a chemical and plastic product recommendation method and system based on user demand. BACKGROUND

[0002] The chemical and plastic industry is an important part of modern manufacturing, and its products are widely used in many fields, including automobiles, construction, packaging, electronics, and medical care. This industry has a wide range of product types, complex specifications, and each product often has multiple application scenarios and attribute characteristics. Therefore, providing personalized chemical and plastic product recommendation services based on user demand not only improves procurement efficiency, but also better meets the business needs of customers.

[0003] Existing chemical and plastic product recommendation techniques mainly include the following aspects: collaborative filtering-based recommendation method: find similar users or similar products by analyzing user historical behavior data for recommendation. Content-based recommendation method: match product attributes and user preferences. Knowledge graph assisted recommendation: use domain knowledge to enhance the interpretability and accuracy of the recommendation system.

[0004] Although the above research has made some progress in the field of chemical and plastic product recommendation, the existing technology still has the following defects: lack of deep integration of chemical and plastic industry professional knowledge, difficult to accurately understand and match complex user demand; the multi-dimensional relationship between products and the time sequence characteristics of user behavior are not fully considered in the recommendation process; the processing capacity for new users and cold start problem is limited; the individuality and interpretability of the recommendation results need to be improved.

[0005] In summary, the existing chemical and plastic product recommendation technology still lacks a comprehensive solution that can deeply integrate domain knowledge, accurately understand user demand, and dynamically adapt to market changes when dealing with personalized recommendation problems in complex professional fields. SUMMARY

[0006] Therefore, the present application provides a chemical and plastic product recommendation method and system based on user demand, which improves the relevance and accuracy of the recommendation by constructing a multi-dimensional chemical and plastic industry knowledge graph, deeply understanding user demand, optimizing query condition generation, and personalized sorting, thereby improving user experience, optimizing procurement decisions, and providing strong support for the digital transformation of the chemical and plastic industry.

[0007] The technical solution of the present application is as follows:

[0008] On the one hand, the present application provides a chemical and plastic product recommendation method based on user demand, comprising:

[0009] S1: Construct a chemical and plastic industry knowledge graph;

[0010] S2 obtains user historical data and current user demand, including historical search data, historical browsing data, historical interaction data and historical transaction data;

[0011] S3 pre-processes the user historical data, and analyzes the pre-processed user historical data to extract historical demand features;

[0012] S4 extracts keywords from the current user demand using a semantic understanding model, and generates optimized query conditions in combination with the plastic industry knowledge graph;

[0013] S5 searches in the plastic product database using the optimized query conditions to obtain a first list;

[0014] S6 uses the plastic industry knowledge graph to perform personalized sorting on the first list, takes the top k to generate a recommendation list, and presents the recommendation list to the user.

[0015] Preferably, the construction process of the plastic industry knowledge graph comprises:

[0016] Relevant information is extracted from the plastic product database, industry standard documents, expert knowledge base and user interaction records;

[0017] The named entity recognition method is used to identify plastic products, product categories, attributes, attribute values, professional terms and application scenarios, and the relationship extraction algorithm is used to extract semantic relationships between entities;

[0018] A multi-dimensional knowledge graph is constructed, including a first-dimensional sub-graph, a second-dimensional sub-graph, a third-dimensional sub-graph and a fourth-dimensional sub-graph, wherein the first-dimensional sub-graph is constructed in a triple format according to product-product category-attribute, the second-dimensional sub-graph is constructed in a triple format according to product-attribute-attribute value, the third-dimensional sub-graph is constructed in a triple format according to product-professional term-synonym / antonym, and the fourth-dimensional sub-graph is constructed in a triple format according to product-attribute-application scenario;

[0019] TransE algorithm is used to perform representation learning on the multi-dimensional knowledge graph to generate low-dimensional vector representation of entities and relationships;

[0020] Entity alignment and relationship mapping are performed on the sub-graphs of different dimensions, and a graph fusion algorithm is used to integrate the sub-graphs of different dimensions into a unified plastic industry knowledge graph.

[0021] Preferably, entity alignment and relationship mapping are performed on the sub-graphs of different dimensions, and a graph fusion algorithm is used to integrate the sub-graphs of different dimensions into a unified plastic industry knowledge graph, comprising:

[0022] Step 1: For each sub-map G i Entity e in i :

[0023] Calculate e i Other sub-maps G j All entities e in j similarity sim(e i ,e j );

[0024] If sim(e) i ,e j If e > threshold T, then e i and e j Marked as the same entity;

[0025] Step 2: Construct the relation mapping matrix R, where R [i][j] Sub-graph G i China-Right Relations i To G j China-Right Relations j The mapping, for each pair of mapping relations (r) i ,r j ):

[0026] Choose a consistent expression format and update R. [i][j] and R [j][i] ;

[0027] Step 3: Use the first-dimensional sub-graph G1 as the initial fused graph G, and for the remaining dimensions of the sub-graph G... i :

[0028] G i The entities and relations in the first step are added to G, and the results of steps one and two are used to handle duplicates and conflicts of entities or relations.

[0029] Step 4: For each entity e in G:

[0030] If there is a conflict in the attribute value of 'e', ​​the attribute value with the latest timestamp will be retained.

[0031] For each relation r in G:

[0032] If r has a conflict, retain the relation with the highest confidence.

[0033] Step 5: For any two triples in G: (entity e1, relation r1, entity e2) and (entity e2, relation r2, entity e3):

[0034] If entities e1 and e3 satisfy the predefined inference rules, then add a new triple (entity e1, relation r) new Entity e3) into G;

[0035] Step six, delete all dangling entities and relationships in G, for each entity e in G:

[0036] If the connectivity of e is less than the threshold K, mark e as pending review;

[0037] Step seven, calculate the importance PR of all entities in G, for entities whose PR is below the threshold P and their corresponding relationships:

[0038] If it does not affect the overall structure of G, remove it from G;

[0039] Step eight, for the entity e marked as pending review in step six:

[0040] Check the importance of e in its original subgraph, if the importance exceeds the threshold P, keep e and increase its connectivity, if its importance is below the threshold P, remove it;

[0041] Step nine, repeat steps one to eight until the fusion is complete, store the fusion graph G in the graph database D, and establish a multi-dimensional index structure on D, that is, obtain the plastic industry knowledge graph.

[0042] Preferably, step S3 comprises:

[0043] S31 cleans the user historical data, removes invalid or abnormal data, and unifies the format to obtain preprocessed user historical data;

[0044] S32 extracts high-frequency words from historical search data and historical browsing data using the TF-IDF method, converts high-frequency words into vector representations using the word2vec model, obtains high-frequency word vectors, and clusters high-frequency word vectors based on clustering algorithms. Each clustering cluster represents a user theme, and a user theme set is obtained;

[0045] S33 extracts target product attributes from historical browsing data and historical interaction data, calculates the attention frequency and dwell time of each product attribute, and generates an attribute preference vector;

[0046] S34 analyzes the price distribution in historical transaction data, calculates the average transaction price and price volatility of the user, and generates a price sensitivity index;

[0047] S35 analyzes the time series of historical transaction data, identifies the procurement cycle and seasonal patterns, calculates the average procurement quantity and procurement frequency, and generates a procurement behavior matrix;

[0048] S36 uses a Naive Bayes classifier to infer application scenarios based on historical search data and historical transaction data, and calculates the probability score of each application scenario;

[0049] S37 calculates the similarity of the user to each user topic, converts the user topic set into a topic similarity vector, takes the topic similarity vector, attribute preference vector, price sensitivity index, purchase behavior matrix, and probability score of application scenario as features, standardizes the features, and performs weighted fusion to form a historical demand feature.

[0050] Preferably, step S37 further comprises: performing time weight calculation on each feature, and performing standardization processing after applying the time weight to each feature, wherein the calculation formula of the time weight is:

[0051] In the formula, w is the time weight, λ is the time decay coefficient, t1 is the current time, t0 is the time when the feature is generated, λ0 is the base decay coefficient, and ts is the time window.

[0052] Preferably, step S4 comprises:

[0053] S41 inputs the current user demand into a pre-trained semantic understanding model to extract a keyword, a semantic feature vector, and prediction information;

[0054] S42 performs entity linking and expansion on the extracted keyword using a knowledge graph of the plastic industry, including synonyms, near synonyms, and related concepts;

[0055] S43 adjusts the weight of the expanded keyword in combination with the historical demand feature and the entity importance in the knowledge graph;

[0056] S44 calculates the similarity of the semantic feature vector and the historical demand feature, and obtains a query preference in combination with the entity relationship in the knowledge graph;

[0057] S45 generates a preliminary query condition based on the weight-adjusted keyword and the query preference, in combination with the attribute preference, price sensitivity, and application scenario information in the historical demand feature, and the product attribute hierarchical structure in the knowledge graph;

[0058] S46 expands the preliminary query condition using the semantic feature vector, prediction information, and association rules in the knowledge graph;

[0059] S47 performs similarity calculation on the expanded query condition, and removes redundant or contradictory conditions using the constraint relationship in the knowledge graph;

[0060] S48 reorders the remaining conditions based on the entity importance in the knowledge graph to obtain the final optimized query condition.

[0061] Preferably, the semantic understanding model comprises an input layer, an embedding layer, a multi-layer transformer encoder, a context enhancement layer, a multi-scale feature extraction layer, an adaptive pooling layer, and a multi-task output layer, wherein:

[0062] The multi-layer transformer encoder comprises L transformer layers, each of which comprises a multi-head self-attention sublayer and a feedforward neural network sublayer, the multi-head self-attention sublayer comprises h attention heads, each of which has a dimension of d / h, the feedforward neural network sublayer adopts a two-layer fully connected network, and a GELU activation function is used in the middle, and LN and a residual connection are used after each sublayer of the multi-layer transformer encoder;

[0063] The context enhancement layer adopts a bidirectional long short-term memory network.

[0064] The multi-scale feature extraction layer comprises a plurality of one-dimensional convolutions with different kernel sizes, each of which is followed by a LeakyReLU activation function and a max-pooling operation.

[0065] The adaptive pooling layer adopts a self-attention mechanism.

[0066] The multi-task output layer comprises three branches, the first branch adopts a conditional random field to extract keywords, the second branch uses a nonlinear transformation to generate a semantic feature vector, and the third branch is a multi-label classifier for generating predicted information, including predicted product categories, predicted attributes, and predicted application scenarios.

[0067] Preferably, the processing process of the semantic understanding model for the current user demand is as follows:

[0068] The input layer receives the current user demand, converts the text into a token sequence, and adds [CLS] and [SEP] markers, wherein [CLS] is used to capture the semantic information of the entire input, and [SEP] is used to distinguish different input segments.

[0069] The embedding layer generates word embeddings, position embeddings, and segment embeddings for each token, and adds the three kinds of embeddings to obtain an initial representation.

[0070] The initial representation is processed by the multi-layer transformer encoder, and each layer comprises:

[0071] The multi-head self-attention calculates the attention weight of each token relative to other tokens to extract sentence-level semantic features.

[0072] The feedforward neural network processes the attention output through a nonlinear transformation.

[0073] The output of the multi-layer transformer encoder is processed using a bidirectional long short-term memory network to add context information.

[0074] The multi-scale feature extraction layer applies multiple convolution kernels to extract features, and splices the features output by the different convolution kernels to obtain multi-scale features;

[0075] The adaptive pooling layer converts the multi-scale features into fixed-length semantic representations through a self-attention mechanism;

[0076] In the multi-task output layer, the first branch uses a conditional random field to perform sequence labeling on each token, outputs a keyword list, the second branch converts the pooled semantic representations into a semantic feature vector through a nonlinear layer, and the third branch uses a multi-label classifier to perform structured prediction on product categories, key attributes, and application scenarios.

[0077] Preferably, step S6 comprises:

[0078] S61 calculates the semantic relevance of each plastics product in the first list to the optimization query condition based on the plastics industry knowledge graph;

[0079] S62 calculates the personalized score of each plastics product in combination with user historical data;

[0080] S63 calculates the similarity between plastics products using multi-dimensional information of the plastics industry knowledge graph;

[0081] S64 combines the semantic relevance, the personalized score, and the similarity by weighting to obtain a comprehensive score of each plastics product, rearranges the first list according to the comprehensive score to obtain a second list;

[0082] S65 selects the top k options from the second list to generate a recommendation list, and presents the recommendation list to the user.

[0083] On the other hand, the present application also provides a plastics product recommendation system based on user demand, which is used to execute the method of any one of the above aspects, and comprises:

[0084] A knowledge graph construction module configured to extract information from a plastics product database, industry standard documents, and an expert knowledge base, form a multi-dimensional knowledge graph using entity recognition and relationship extraction, and perform entity alignment, relationship mapping, and graph fusion operations to generate a unified plastics industry knowledge graph;

[0085] A data acquisition module configured to obtain historical search data, browsing data, interaction data, and transaction data from user interaction activities;

[0086] A preprocessing module connected to the data acquisition module and configured to perform data cleaning, format standardization, and outlier removal on the obtained user historical data to generate a preprocessed user data set;

[0087] a user modeling module configured to extract user historical behavior features by analyzing the preprocessed user data set, and to generate a user interest topic, an attribute preference vector, a price sensitivity index, a purchase behavior matrix, and a probability score of an application scenario;

[0088] a semantic understanding module built-in semantic understanding model, including an input layer, an embedding layer, a multi-layer transformer encoder, a context enhancement layer, a multi-scale feature extraction layer, an adaptive pooling layer, and a multi-task output layer, for performing semantic analysis on the current user demand, extracting keywords, semantic feature vectors, and prediction information;

[0089] a query optimization module connected between the semantic understanding module and the knowledge graph construction module, and configured to perform semantic expansion, weight adjustment, and condition optimization on the user query in combination with the user historical demand features and the plastic industry knowledge graph, to generate an optimized query condition;

[0090] a product search module configured to perform a search operation in the plastic product database according to the optimized query condition, to generate a first list;

[0091] a personalized sorting module connected to the product search module, and configured to comprehensively calculate a comprehensive score of each plastic product according to semantic relevance, personalized score, and product similarity, to sort and adjust the first list;

[0092] a recommendation generation module configured to select a preset number of products from the first list after sorting and adjustment, to generate a final recommendation list and output to the user;

[0093] a feedback collection module for collecting user feedback on the recommendation result, to optimize the plastic industry knowledge graph.

[0094] The present application has the following beneficial effects over the prior art:

[0095] (1) By fusing multi-dimensional knowledge graph, deep semantic understanding, query optimization and personalized sorting, the present application realizes accurate understanding and matching of user demand for plastic products, significantly improves the accuracy and relevance of the recommendation, and effectively solves the product recommendation problem in the complex semantic environment of the plastic industry;

[0096] (2) By constructing a multi-dimensional plastic industry knowledge graph and performing graph fusion, the present application systematically integrates product categories, attributes, professional terms and application scenarios, and other information, provides a rich knowledge base for subsequent semantic understanding and query optimization, and enhances the ability of the recommendation system to grasp the characteristics of the industry;

[0097] (3) Utilizing a semantic understanding model based on a multi-layer Transformer encoder, combining context enhancement and multi-scale feature extraction methods, the user demand is deeply analyzed, and the key words, semantic feature vectors and prediction information are extracted, greatly improving the understanding ability of the system to complex user demand;

[0098] (4) Adopting a multi-stage query optimization strategy, including entity linking, semantic expansion, weight adjustment and condition generation, etc., effectively converting user demand into optimized query conditions, improving the retrieval efficiency and relevance of the results;

[0099] (5) Through the integrated ranking mechanism of semantic relevance, personalized score and product similarity, the personalized rearrangement of search results is realized, ensuring the high matching of the recommendation list with user demand, while considering the product diversity;

[0100] (6) The recommendation system realizes the intelligent processing of the whole process from user demand input to final recommendation list output through modular design and integration of multiple innovative technologies, providing efficient, accurate and personalized product recommendation services for the users in the plastic industry, significantly improving the user experience and procurement decision efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0101] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0102] Fig. 1 is a flow chart of the method of the present application;

[0103] Fig. 2 is the knowledge graph construction process of the present application;

[0104] Fig. 3 is a graph fusion algorithm flow chart of the present application;

[0105] Fig. 4 is a framework diagram of the semantic understanding model of the present application;

[0106] Fig. 5 is a multi-layer transformer encoder structure diagram of the present application;

[0107] Fig. 6 is a multi-scale feature extraction layer structure diagram of the present application;

[0108] Fig. 7 is a multi-task output layer structure diagram of the present application;

[0109] Fig. 8 is a system framework diagram of the present application. DETAILED DESCRIPTION

[0110] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0111] As shown in FIG. 1, first, the present application provides a plastic product recommendation method based on user demand, comprising:

[0112] S1, constructing a plastic industry knowledge graph;

[0113] S2, obtaining user historical data and current user demand, including historical search data, historical browsing data, historical interaction data and historical transaction data;

[0114] S3, preprocessing the user historical data, and analyzing the preprocessed user historical data to extract historical demand features;

[0115] S4, extracting keywords from the current user demand by using a semantic understanding model, and combining the plastic industry knowledge graph to generate an optimized query condition;

[0116] S5, searching in the plastic product database by using the optimized query condition to obtain a first list;

[0117] S6, performing personalized sorting on the first list by using the plastic industry knowledge graph, taking the top k to generate a recommendation list, and presenting the recommendation list to the user.

[0118] Specifically, as shown in FIG. 2, in an embodiment of the present application, the construction process of the plastic industry knowledge graph comprises:

[0119] Extracting relevant information from the plastic product database, industry standard documents, expert knowledge base and user interaction records;

[0120] The purpose of this step is to obtain comprehensive industry data, including product information, technical specifications, expert experience and actual application scenarios. The specific implementation method includes: using SQL query to extract product information from structured database, including product name, specification, performance parameter, etc.; using web crawler technology to obtain standard documents from industry websites and public document libraries, such as GB / T, ISO, etc.; accessing expert knowledge management system through API interface to extract expert contributed knowledge points, including expert description of product application, technical points; extracting user query and feedback information from user interaction system log data.

[0121] The named entity recognition method is used to identify plastic products, product categories, attributes, attribute values, professional terms and application scenarios, and the relationship extraction algorithm is used to extract semantic relationships between entities.

[0122] Specifically, the named entity recognition is performed based on a BERT-CRF deep learning model, which is fine-tuned based on the plastic industry corpus and can accurately identify industry-specific entities; the relationship extraction is performed using a neural network model based on remote supervision, and the attention mechanism is used to improve the identification accuracy of long-distance dependent relationships.

[0123] A multi-dimensional knowledge graph is constructed, including a first-dimensional sub-graph, a second-dimensional sub-graph, a third-dimensional sub-graph and a fourth-dimensional sub-graph. The first-dimensional sub-graph is constructed in a triple format according to product-product category-attribute, the second-dimensional sub-graph is constructed in a triple format according to product-attribute-attribute value, the third-dimensional sub-graph is constructed in a triple format according to product-professional term-synonym / antonym, and the fourth-dimensional sub-graph is constructed in a triple format according to product-attribute-application scenario.

[0124] Specifically, the embodiment constructs four-dimensional sub-graphs, each of which focuses on a different aspect of plastic industry knowledge to comprehensively capture the structure of industry knowledge. The first dimension reflects the classification system of products, the second dimension describes product characteristics, the third dimension constructs a term system, and the fourth dimension associates products with actual applications. The purpose of constructing these four-dimensional sub-graphs is to comprehensively represent plastic industry knowledge from different angles, facilitating subsequent complex queries and reasoning.

[0125] TransE algorithm is used to perform representation learning on the multi-dimensional knowledge graph to generate low-dimensional vector representations of entities and relationships.

[0126] The reason for selecting the TransE algorithm in this embodiment is that it is simple and effective, suitable for processing large-scale knowledge graphs, and the generated vector representations facilitate subsequent similarity calculation and reasoning tasks.

[0127] Entity alignment and relationship mapping are performed on different dimensional sub-graphs, and a graph fusion algorithm is used to integrate the sub-graphs of each dimension into a unified plastic industry knowledge graph. As shown in FIG. 3, this step includes:

[0128] Step one, for each sub-graph G i entity e i :

[0129] Calculate the similarity sim(e i , e j ) between e j and all entities e i in other sub-graphs G j .

[0130] If sim(e i ,e j )> threshold T, mark e i and e j as the same entity.

[0131] This step is the entity alignment step, aiming to identify entities representing the same concept in different sub-graphs.

[0132] Step two, construct the relation mapping matrix R, where R [i][j] represents the mapping of relation r i in sub-graph G i to relation r j in G j , for each pair of mapping relations (r i , r j ):

[0133] Select a unified form of expression, update R [i][j] and R [j][i] .

[0134] This step is the relation mapping step, which ensures that relations with the same semantics but different expressions in different sub-graphs can be uniformly represented.

[0135] Step three, take the first-dimensional sub-graph G1 as the initialized fusion graph G, and for the remaining dimensional sub-graphs G i :

[0136] Add entities and relations in G i to G, and use the results of step one and step two to handle the repetition and conflict of entities or relations.

[0137] Step four, for each entity e in G:

[0138] If there is a property value conflict, keep the property value with the latest timestamp;

[0139] For each relation r in G:

[0140] If there is a conflict, keep the relation with the highest confidence.

[0141] Step five, for any two triples (entity e1, relation r1, entity e2), (entity e2, relation r2, entity e3) in G:

[0142] If entity e1 and entity e3 satisfy the pre-defined inference rule, add a new triple (entity e1, relation r new , entity e3) to G.

[0143] Step six, delete all dangling entities and relationships in G, for each entity e in G:

[0144] If the connectivity of e is less than threshold K, mark e as to be reviewed.

[0145] Specifically, the connectivity = 2*actual number of connections between neighbors / (number of neighbors*(number of neighbors-1)).

[0146] Step seven, calculate the importance PR of all entities in G, for entities whose PR is below the threshold P and their corresponding relationships:

[0147] If it does not affect the overall structure of G, remove it from G.

[0148] The importance PR of the entity is calculated as follows:

[0149] PR(e) = (1-d) + d*∑(PR(i) / L(i));

[0150] Where d is the damping coefficient (value 0.85), i refers to the entity pointing to e, and L(i) is the out-degree of i.

[0151] Step eight, for the entity e marked as to be reviewed in step six:

[0152] Check the importance of e in its original subgraph, if the importance exceeds the threshold P, keep e and increase its connectivity, if its importance is below the threshold P, remove it.

[0153] Specifically, the way to increase the connectivity is as follows:

[0154] Use existing relationship patterns to infer and add new reasonable relationships. For example: if A is a subclass of B, B has attribute C, then it can be inferred that A may also have attribute C.

[0155] Based on the attributes of similar entities, supplement the possible missing attributes of entity e.

[0156] Add synonymous or near-synonymous entities related to entity e and establish corresponding relationships.

[0157] Step nine, repeat steps one to eight until the fusion is completed, store the fusion graph G into the graph database D, and establish a multi-dimensional index structure on D, that is, obtain the plastic industry knowledge graph.

[0158] Specifically, establishing a multi-dimensional index structure includes:

[0159] According to the user's needs and the working mode of the query optimization module, analyze the most common query types. Determine the query dimensions that need to be optimized, such as product attributes, application scenarios, professional terms, etc.

[0160] Product category index and attribute index are established based on the first dimension sub-graph (product-product category-attribute). Attribute value index is established based on the second dimension sub-graph (product-attribute-attribute value). Term index is established based on the third dimension sub-graph (product-specialized term-synonym / paronym). Application scenario index is established based on the fourth dimension sub-graph (product-attribute-application scenario).

[0161] For text attributes such as product name and chemical composition, full-text index such as inverted index is used. For numerical attributes such as physical properties, B-tree or R-tree index is used. For category attributes, hash index is used. For time-related attributes such as update time, time index is used.

[0162] Product category index: tree structure is used to represent product category hierarchy. Attribute index: B-tree index is created for common attributes. Attribute value index: range index is created for numerical attribute values. Term index: inverted index of specialized terms to products is established. Application scenario index: mapping index of application scenarios to products is created.

[0163] Based on common query patterns, composite index of multi-attribute combination is created. For example, combination index of product category-attribute-attribute value.

[0164] Adjacency list index is created to optimize graph traversal operation. Path index is implemented to accelerate multi-hop query.

[0165] Vector index is created for entity and relation vector representation generated by TransE algorithm. Approximate nearest neighbor search algorithm such as HNSW is used to realize efficient similarity query.

[0166] Data is partitioned according to product category or other key attributes. Local index structure is established on each partition.

[0167] Through the above steps, a comprehensive multi-dimensional index structure can be established on the graph database D to support efficient query and recommendation functions of the plastic industry knowledge graph.

[0168] Through the construction of multi-dimensional sub-graphs, the knowledge structure of the plastic industry is comprehensively captured, including product classification, attribute characteristics, specialized terms, and application scenarios. Deep learning models are used for entity recognition and relation extraction to improve the accuracy of knowledge extraction. Knowledge graph representation learning is used to provide a good semantic foundation for subsequent intelligent applications. Through graph fusion algorithms, multi-dimensional knowledge is organically integrated to construct a unified plastic industry knowledge graph.

[0169] Specifically, in an embodiment of the present application, step S2 comprises:

[0170] Acquire historical search data: Record all search queries of users on the platform. Including search keywords, search time, search frequency.

[0171] Acquire historical browsing data: Record product pages browsed by users. Including browsed product ID, browsing time, browsing duration. Record the scrolling behavior of users on product pages to determine attention.

[0172] Acquire historical interaction data: Record various interaction behaviors of users with the platform. Including operations such as collecting products, adding shopping cart, product comparison, etc.

[0173] Acquire historical transaction data: Record all purchase behaviors of users. Including product ID, purchase time, purchase quantity, transaction amount. Record order status changes such as ordering, payment, delivery, receipt, etc.

[0174] Current user demand acquisition: Capture the current search query or browsing behavior of the user. Analyze the behavior sequence of the user in the current session.

[0175] Specifically, in an embodiment of the present application, step S3 includes:

[0176] S31 cleans the user historical data, removes invalid or abnormal data, and unifies the format to obtain preprocessed user historical data;

[0177] S32 extracts high-frequency words from historical search data and historical browsing data using TF-IDF method, converts high-frequency words into vector representation using word2vec model, obtains high-frequency word vector, clusters high-frequency word vector based on clustering algorithm, each cluster represents a user theme, and obtains user theme set;

[0178] S33 extracts target product attributes from historical browsing data and historical interaction data, calculates the attention frequency and dwell time of each product attribute, and generates attribute preference vector;

[0179] S34 analyzes the price distribution in historical transaction data, calculates the average transaction price and price volatility of the user, and generates the price sensitivity index;

[0180] S35 analyzes the time series of historical transaction data, identifies the procurement cycle and seasonal pattern, calculates the average procurement quantity and procurement frequency, and generates the procurement behavior matrix;

[0181] S36 infers application scenarios based on historical search data and historical transaction data using Naive Bayes classifier, and calculates the probability score of each application scenario;

[0182] S37 Calculate the similarity of the user to each user topic, convert the user topic set to a topic similarity vector, and use the topic similarity vector, attribute preference vector, price sensitivity index, purchase behavior matrix, and probability score of application scenario as features. Standardize the features and perform weighted fusion to form the historical demand characteristics.

[0183] As an example:

[0184] User A's original historical data is as follows:

[0185] Search records: ["polypropylene", "PP material", "transparent plastic", "high-strength plastic bag"]; browsing records: [product ID1, product ID2, product ID3]; transaction records: [{product ID: 1, price: 5000, quantity: 1000, date: "2023-01-15"}, {product ID: 2, price: 6000, quantity: 800, date: "2023-03-20"}].

[0186] Preprocess the data: remove duplicate search records; unify product ID format (e.g., convert all to strings); check and remove abnormal prices and quantities (e.g., values that are significantly deviating from the normal range); unify date format to YYYY-MM-DD.

[0187] The cleaned data is:

[0188] Search records: ["polypropylene", "PP material", "transparent plastic", "high-strength plastic bag"]; browsing records: ["001", "002", "003"]; transaction records: [{product ID: "001", price: 5000, quantity: 1000, date: "2023-01-15"}, {product ID: "002", price: 6000, quantity: 800, date: "2023-03-20"}].

[0189] Extract high-frequency words using TF-IDF: TF-IDF(t, d) = TF(t, d) * IDF(t), where TF(t, d) is the frequency of word t in document d, IDF(t) = log(N / DF(t)), N is the total number of documents, and DF(t) is the number of documents containing word t. The extracted high-frequency words are: ["polypropylene", "PP material", "transparent", "high-strength"].

[0190] High-frequency words are converted into 100-dimensional vectors using word2vec. The word vectors are clustered using the K-means clustering algorithm, where k = 3, resulting in 3 clusters: Cluster 1: ["polypropylene", "PP material"]; Cluster 2: ["transparent"]; Cluster 3: ["high strength"]. These three clusters represent three user topics: raw material type, appearance characteristics, and strength characteristics.

[0191] The target attributes are: material type, transparency, strength, and price. For each attribute, calculate the attention frequency and average dwell time, then construct an attribute preference vector as [(material type attention frequency, material type average dwell time), (transparency attention frequency, transparency average dwell time), (strength attention frequency, strength average dwell time), (price attention frequency, price average dwell time)].

[0192] Calculate the average transaction price, then calculate the price volatility: price volatility = standard deviation / average price. Then generate the price sensitivity index: price sensitivity index = 1 / (1 + price volatility).

[0193] Identify the procurement cycle, calculate the average procurement quantity and frequency, and construct the procurement behavior matrix [procurement cycle, average procurement quantity, procurement frequency].

[0194] Set the target application scenarios as packaging, construction, and automotive parts. Use the Naive Bayes classifier to calculate the probability of each scenario based on historical search and transaction data: P(scene|data) = P(data|scene) * P(scene) / P(data). The probability score is: [packaging: 0.6, construction: 0.3, automotive parts: 0.1].

[0195] Use cosine similarity to calculate the similarity between the user and each topic, resulting in a topic similarity vector [raw material type: 0.8, appearance characteristics: 0.5, strength characteristics: 0.7]. Apply time weight to each feature, the formula for calculating time weight is:

[0196] where w is the time weight, λ is the time decay coefficient, t1 is the current time, t0 is the time when the feature was generated, λ0 is the base decay coefficient, t s is the time window.

[0197] At this time, the time window can be set to 1 day, i.e. take one day as the window, the longer the distance from the current time, the smaller the corresponding weight. Then perform min-max standardization on all features, and then perform weighted fusion to obtain the historical demand features.

[0198] The final historical demand features are a multi-dimensional vector containing information such as user interest topics, attribute preferences, price sensitivity, procurement behavior, and application scenario preferences.

[0199] Specifically, in an embodiment of the present application, step S4 comprises:

[0200] S41 inputting the current user demand into the pre-trained semantic understanding model to extract keywords, semantic feature vectors and prediction information;

[0201] Specifically, as shown in FIG. 4, the semantic understanding model comprises an input layer, an embedding layer, a multi-layer transformer encoder, a context enhancement layer, a multi-scale feature extraction layer, an adaptive pooling layer and a multi-task output layer, wherein:

[0202] As shown in FIG. 5, the multi-layer transformer encoder comprises L transformer layers, each layer comprising a multi-head self-attention sub-layer and a feed-forward neural network sub-layer, the multi-head self-attention sub-layer comprising h attention heads, each attention head having a dimension of d / h, the feed-forward neural network sub-layer adopting a two-layer fully connected network, using a GELU activation function in the middle, and LN and a residual connection being used after each sub-layer of the multi-layer transformer encoder;

[0203] The context enhancement layer adopts a bidirectional long short-term memory network;

[0204] As shown in FIG. 6, the multi-scale feature extraction layer comprises a plurality of one-dimensional convolutions with different kernel sizes, each convolution being followed by a LeakyReLU activation function and a max-pooling operation;

[0205] The adaptive pooling layer adopts a self-attention mechanism;

[0206] As shown in FIG. 7, the multi-task output layer comprises three branches, the first branch adopts a conditional random field to extract keywords, the second branch uses a nonlinear transformation to generate semantic feature vectors, and the third branch is a multi-label classifier for generating prediction information, including predicted product categories, predicted attributes and predicted application scenarios.

[0207] The processing process of the semantic understanding model for the current user demand is as follows:

[0208] The input layer receives the current user demand, converts the text into a token sequence, and adds [CLS] and [SEP] marks, wherein [CLS] is used to capture the semantic information of the entire input, and [SEP] is used to distinguish different input segments; specifically, [CLS] is at the beginning of the sequence, and [SEP] is at the end of the sequence, for example, "I need high-strength PP material"→[CLS], I, need, high, strength, of, PP, material, [SEP].

[0209] The embedding layer generates word embeddings, position embeddings and segment embeddings for each token, and adds the three kinds of embeddings to obtain an initial representation; wherein the word embeddings are semantic representations of each token, the position embeddings capture the position information of the token in the sequence, and the segment embeddings distinguish different input segments.

[0210] The initial representation is processed by a multi-layer transformer encoder, each layer including:

[0211] The multi-head self-attention calculates the attention weight of each token relative to other tokens, and extracts sentence-level semantic features.

[0212] The feedforward neural network processes the attention output through a nonlinear transformation.

[0213] Specifically, in this embodiment, the multi-layer transformer encoder includes 12 transformer layers, each layer including: h attention heads, each head having a dimension of d / h, and calculating the attention weight: The feedforward neural network includes two fully connected networks using a GELU activation function.

[0214] Each sub-layer is followed by an LN and a residual connection.

[0215] The output of the multi-layer transformer encoder is processed using a bidirectional long short-term memory network to add context information.

[0216] The multi-scale feature extraction layer applies multiple convolution kernels to extract features, and splices the features output by different convolution kernels to obtain multi-scale features. Specifically, the multi-scale feature extraction layer uses multiple parallel one-dimensional convolutions, each followed by a LeakyReLU activation and a max pooling.

[0217] The adaptive pooling layer converts the multi-scale features into a fixed-length semantic representation through a self-attention mechanism.

[0218] In the multi-task output layer, the first branch uses a conditional random field to perform sequence labeling on each token, outputs a keyword list, the second branch converts the pooled semantic representation into a semantic feature vector through a nonlinear layer, and the third branch uses a multi-label classifier to perform structured prediction on product categories, key attributes and application scenarios.

[0219] S42 uses the plastic industry knowledge graph to perform entity linking and expansion on the extracted keywords, including synonyms, near-synonyms and related concepts.

[0220] Specifically, the extracted keywords are matched with entities in the knowledge graph. A combination of string matching and semantic similarity calculation is used. For example: the keyword "PP material" is linked to the "polypropylene" entity in the knowledge graph.

[0221] Entity expansion includes synonym expansion, near-synonym expansion and related concept expansion.

[0222] Synonym expansion: find other entities that are synonymous with the linked entity. For example: "polypropylene" → {"PP", "polypropylene", "polypropylene polymer"}.

[0223] Near-synonym expansion: find entities that are similar to the linked entity. For example: "high-strength plastic" → {"high-rigidity plastic", "high-toughness plastic"}.

[0224] Related concept expansion: find other entities that have a close relationship with the linked entity. For example: "plastic bag" → {"packaging material", "film", "shopping bag"}.

[0225] S43 combines historical demand features and entity importance in the knowledge graph to adjust the weights of the expanded keywords.

[0226] Specifically, the similarity between each expanded keyword and the user's historical interest theme is calculated. The higher the similarity, the higher the weight. The importance of each keyword in the knowledge graph is calculated. The higher the importance, the higher the weight.

[0227] S44 calculates the similarity between the semantic feature vector and the historical demand features, and combines the entity relationships in the knowledge graph to obtain the query preference.

[0228] Specifically, the cosine similarity is used to calculate the similarity between the semantic feature vector and the historical demand features. The main components in the semantic feature vector are analyzed and the entity relationships in the knowledge graph are analyzed. For example, if the semantic vector emphasizes "strength", the weight of the properties related to strength is increased. Based on the similarity and entity relationship analysis, a preference vector is generated. The preference vector contains the weights of different attributes and characteristics.

[0229] S45 generates preliminary query conditions based on the weight-adjusted keywords and query preferences, combined with the attribute preferences, price sensitivity and application scenario information in the historical demand features, and the product attribute hierarchical structure in the knowledge graph.

[0230] Specifically, the weight-adjusted keywords and query preferences are combined. The attribute preference vector, price sensitivity index and application scenario information in the historical demand features are used. The user preferences are mapped to specific product attributes and attribute values in the knowledge graph. Then the preliminary query conditions are generated.

[0231] S46 uses the semantic feature vector, prediction information and association rules in the knowledge graph to expand the preliminary query conditions.

[0232] Specifically, the main semantic components in the semantic feature vector are analyzed, and relevant query conditions are added. According to the predicted product category, attribute and application scenario, the corresponding query condition is added. The association rules in the knowledge graph are used: the frequent item set mining algorithm is used to find the association rules between attributes. The query condition is extended according to the association rules.

[0233] S47, the extended query condition is calculated for similarity, and the constraint relationship in the knowledge graph is used to remove redundant or contradictory conditions.

[0234] Specifically, the semantic similarity between the extended query conditions is calculated using GloVe. Check if the condition violates the constraints defined in the knowledge graph. If the similarity of the two conditions exceeds the threshold, keep the more specific or more important one. If contradictory conditions are found, keep one according to the weight of the condition, or convert the contradictory conditions into a range query.

[0235] S48, based on the entity importance in the knowledge graph, the remaining conditions are reordered to obtain the final optimized query condition.

[0236] Specifically, the remaining query conditions are sorted using the entity importance in the knowledge graph. The sorted conditions are converted into a structured query format, such as SQL query.

[0237] Specifically, in an embodiment of the present application, step S5 comprises:

[0238] Access the plastic product database, execute the query in the database, and return the results from the database query as a preliminary product list, which contains all product records that meet the query conditions. The query results are preliminarily filtered and formatted, and the filtering method is the default relevance filtering. The filtered and formatted product records are output as the first list.

[0239] Specifically, in an embodiment of the present application, step S6 comprises:

[0240] S61, based on the plastic industry knowledge graph, calculates the semantic relevance of the optimized query condition for each plastic product in the first list.

[0241] Specifically, the optimized query condition is converted into a vector representation, and the cosine similarity is used to calculate the similarity between the query vector and the product vector. The specific attribute requirements in the query condition are compared with the product attributes, and the matching degree is calculated. The fuzzy matching algorithm is used to handle the approximate matching of numerical ranges and text descriptions. The graph structure of the knowledge graph is used to calculate the structural similarity of the query entity and the product entity in the graph. The vector similarity, attribute matching degree and graph structure similarity are considered comprehensively to obtain the final semantic relevance score. In this embodiment, the weighted average is used to obtain the semantic relevance.

[0242] S62 combines user historical data to calculate the individualization score of each plastic product.

[0243] Specifically, a user interest model is constructed using matrix decomposition techniques: a user-product interaction matrix R is constructed, where R[u,i] represents the interaction intensity of user u to product i, SVD (singular value decomposition) is applied for matrix decomposition: R ≈ U*Σ*V T , the user latent factor matrix U and the product latent factor matrix V are extracted, and the interest vector of user u: I u = U[u].

[0244] For each product attribute a, the preference degree of user u is calculated: P(u,a) = Σ(w(u,i)*v(i,a)) / Σw(u,i), where: w(u,i) is the interaction weight of user u to product i, and v(i,a) is the value of product i in attribute a.

[0245] A context vector C is constructed, containing information such as season, geographical location, etc., and a context-aware neural network model is used: f(u,i,C) = σ(W2*ReLU(W1*[I u ;V i ;C]+b1)+b2), where V i is the latent factor vector of product i.

[0246] User-based collaborative filtering is used: sim(u,v) = cos(I u ,I v ); CF(u,i) = Σ(sim(u,v)*R[v,i]) / Σsim(u,v).

[0247] A gradient boosting tree (GBT) model is used to fuse various factors to obtain the individualization score.

[0248] S63 uses the multi-dimensional information of the plastic industry knowledge graph to calculate the similarity between plastic products.

[0249] Specifically, multi-dimensional information includes attributes, functions, structures, etc., and the similarity of each dimension is fused to obtain a similarity score.

[0250] S64 combines the semantic relevance, individualization score, and similarity with weighting to obtain a comprehensive score for each plastic product, rearranges the first list according to the comprehensive score, and obtains a second list.

[0251] S65 selects the top k options from the second list to generate a recommendation list, and presents the recommendation list to the user.

[0252] Specifically, the user interaction behavior is tracked, and feedback information such as user clicks on the recommended results, browsing time, etc. is recorded. Subsequently, the knowledge graph is optimized.

[0253] In addition, referring to FIG. 8, the present application also provides a plastic product recommendation system based on user demand, which is used to execute any of the above-mentioned methods, and the system comprises:

[0254] A knowledge graph construction module configured to extract information from a plastic product database, industry standard documents and expert knowledge base, form a multi-dimensional knowledge graph using entity recognition and relationship extraction, and perform entity alignment, relationship mapping and graph fusion operations to generate a unified knowledge graph of the plastic industry;

[0255] A data acquisition module configured to obtain historical search data, browsing data, interaction data and transaction data from user interaction activities;

[0256] A preprocessing module connected to the data acquisition module and configured to perform data cleaning, format standardization and outlier removal on the obtained user historical data to generate a preprocessed user data set;

[0257] A user modeling module configured to extract user historical behavior features by analyzing the preprocessed user data set, generate user interest topics, attribute preference vectors, price sensitivity indexes, purchasing behavior matrices and probability scores of application scenarios;

[0258] A semantic understanding module with a built-in semantic understanding model including an input layer, an embedding layer, a multi-layer transformer encoder, a context enhancement layer, a multi-scale feature extraction layer, an adaptive pooling layer and a multi-task output layer, which performs semantic analysis on the current user demand, extracts keywords, semantic feature vectors and prediction information;

[0259] A query optimization module connected between the semantic understanding module and the knowledge graph construction module and configured to combine user historical demand features and the knowledge graph of the plastic industry to perform semantic expansion, weight adjustment and condition optimization on user queries to generate optimized query conditions;

[0260] A product search module configured to perform search operations in the plastic product database according to the optimized query conditions to generate a first list;

[0261] A personalized sorting module connected to the product search module and configured to comprehensively calculate the comprehensive score of each plastic product according to semantic relevance, personalized score and product similarity to sort and adjust the first list;

[0262] A recommendation generation module configured to select a preset number of products from the first list after sorting and adjustment to generate a final recommendation list and output to the user;

[0263] a feedback collection module configured to collect feedback of the user on the recommendation result, so as to optimize the plastic industry knowledge graph.

[0264] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for recommending a plastic product based on user needs, characterized by, Comprise: S1 build the knowledge graph of plastic industry; S2 obtain user historical data and current user demand, including historical search data, historical browsing data, historical interaction data and historical transaction data; S3 preprocess the user historical data and analyze the preprocessed user historical data to extract historical demand features; S4 use semantic understanding model to extract keywords from current user demand, and generate optimized query conditions combined with the knowledge graph of plastic industry; S5 search in the plastic product database using the optimized query conditions to get the first list; S6 use the knowledge graph of plastic industry to sort the first list, take the top k to generate the recommendation list, and present the recommendation list to the user.

2. The user demand-based plastic product recommendation method of claim 1, wherein, The construction process of the knowledge graph of plastic industry includes: Extract relevant information from the plastic product database, industry standard documents, expert knowledge base and user interaction records; Use named entity recognition method to identify plastic products, product categories, attributes, attribute values, professional terms and application scenarios, and use relationship extraction algorithm to extract semantic relationships between entities; Build multi-dimensional knowledge graph, including first dimension sub-graph, second dimension sub-graph, third dimension sub-graph and fourth dimension sub-graph, wherein the first dimension sub-graph is constructed according to the product-product category-attribute in the form of triple, the second dimension sub-graph is constructed according to the product-attribute-attribute value in the form of triple, the third dimension sub-graph is constructed according to the product-professional term-synonym / antonym in the form of triple, and the fourth dimension sub-graph is constructed according to the product-attribute-application scenario in the form of triple; Use TransE algorithm to perform representation learning on the multi-dimensional knowledge graph to generate low-dimensional vector representation of entities and relationships; Align entities and map relationships of different dimensional sub-graphs, and use graph fusion algorithm to integrate each dimensional sub-graph into a unified plastic industry knowledge graph.

3. The method of claim 2, wherein the product recommendation is based on a user demand. Align entities and map relationships of different dimensional sub-graphs, and use graph fusion algorithm to integrate each dimensional sub-graph into a unified plastic industry knowledge graph, including: Step one, for each sub-graph G i of entities e i : Calculate e i Other sub-maps G j All entities e in j similarity sim(e i ,e j ); If sim(e i , e j ) > threshold T, then e i and e j are labeled as the same entity. Step two, construct a relation mapping matrix R, where R [i][j] represents the mapping of relations r i in sub-graph G i to relations r j in G j , for each pair of mapping relations (r i , r j ): Select a uniform form of expression, update R [i][j] and R [j][i] ; Step three, take the first dimension sub-spectrum G1 as the initialized fusion spectrum G, and for the sub-spectrum G of the remaining dimensions i : Add entities and relations in G i to G, handle the duplicates and conflicts of entities or relations using the results of Step One and Step Two; Step four, for each entity e in G: If there is a conflict in the attribute value of e, keep the attribute value with the latest timestamp; For each relationship r in G: If there is a conflict, keep the relationship with the highest confidence; Step five, for any two triples (entity e1, relationship r1, entity e2), (entity e2, relationship r2, entity e3) in G: If the entities e1 and e3 satisfy a predefined inference rule, then a new triple (entity e1, relation r new , entity e3) is added to G. Step six, delete all dangling entities and relationships in G, and for each entity e in G: If the connectivity of e is less than the threshold K, mark e as pending; Step seven, calculate the importance PR of all entities in G, and for entities with PR lower than the threshold P and their corresponding relationships: If it does not affect the overall structure of G, remove it from G; Step eight, for the entity e marked as pending in step six: Check the importance of e in its original sub-graph, if the importance exceeds the threshold P, keep e and increase its connectivity, if the importance is lower than the threshold P, remove it; Step nine, repeat steps one to eight until the fusion is completed, store the fusion graph G into the graph database D, and establish a multi-dimensional index structure on D, that is, obtain the plastic industry knowledge graph.

4. The user demand-based plastic product recommendation method of claim 2, wherein, Step S3 comprises: S31 clean the user historical data, remove invalid or abnormal data, and unify the format to obtain preprocessed user historical data; S32 extract high-frequency words from historical search data and historical browsing data using the TF-IDF method, convert the high-frequency words into vector representations using the word2vec model, obtain high-frequency word vectors, and cluster the high-frequency word vectors based on a clustering algorithm, each cluster representing a user theme, and obtain a user theme set; S33 extract target product attributes from historical browsing data and historical interaction data, calculate the attention frequency and dwell time of each product attribute, and generate an attribute preference vector; S34 analyze the price distribution in the historical transaction data, calculate the average transaction price and price volatility of the user, and generate a price sensitivity index; S35 analyze the time series of historical transaction data, identify the procurement cycle and seasonal patterns, calculate the average procurement quantity and procurement frequency, and generate a procurement behavior matrix; S36 infer application scenarios based on historical search data and historical transaction data using a Naive Bayes classifier, and calculate the probability score of each application scenario; S37 calculate the similarity between the user and each user theme, convert the user theme set into a theme similarity vector, and use the theme similarity vector, the attribute preference vector, the price sensitivity index, the procurement behavior matrix, and the probability score of the application scenario as features, standardize the features, and perform weighted fusion to form historical demand features.

5. The user demand-based plastic product recommendation method of claim 4, wherein, The step S37 further comprises: performing time weight calculation on each feature, and performing standardization processing after applying the time weight to each feature, wherein a calculation formula of the time weight is: In the formula, w is the time weight, λ is the time decay coefficient, t1 is the current time, t0 is the time when the feature is generated, λ0 is the base decay coefficient, and ts is the time window.

6. The user demand-based plastic product recommendation method of claim 4, wherein, Step S4 comprises: S41 input the current user demand into the pre-trained semantic understanding model to extract keywords, semantic feature vectors, and prediction information; S42 perform entity linking and expansion on the extracted keywords using the plastic industry knowledge graph, including synonyms, near-synonyms, and related concepts; S43 adjust the weights of the expanded keywords in combination with the historical demand features and the importance of entities in the knowledge graph; S44 calculate the similarity between the semantic feature vectors and the historical demand features, and combine the entity relationships in the knowledge graph to obtain query preferences; S45 generate preliminary query conditions based on the weight-adjusted keywords and query preferences, in combination with the attribute preferences, price sensitivity, and application scenario information in the historical demand features, as well as the product attribute hierarchical structure in the knowledge graph; S46 expand the preliminary query conditions using the semantic feature vectors, prediction information, and association rules in the knowledge graph; S47 perform similarity calculation on the expanded query conditions, and use the constraint relationships in the knowledge graph to remove redundant or contradictory conditions; S48 reorder the remaining conditions based on the importance of entities in the knowledge graph to obtain the final optimized query conditions.

7. The user demand-based plastic product recommendation method of claim 6, wherein, The semantic understanding model comprises an input layer, an embedding layer, a multi-layer transformer encoder, a context enhancement layer, a multi-scale feature extraction layer, an adaptive pooling layer and a multi-task output layer, wherein: The multi-layer transformer encoder comprises L transformer layers, each of which comprises a multi-head self-attention sublayer and a feedforward neural network sublayer, the multi-head self-attention sublayer comprises h attention heads, each of which has a dimension of d / h, the feedforward neural network sublayer adopts a two-layer fully connected network, a GELU activation function is used in the middle, and LN and a residual connection are used after each sublayer of the multi-layer transformer encoder; The context enhancement layer adopts a bidirectional long short-term memory network; The multi-scale feature extraction layer comprises a plurality of one-dimensional convolutions with different kernel sizes, each of which is followed by a LeakyReLU activation function and a max-pooling operation; The adaptive pooling layer adopts a self-attention mechanism; The multi-task output layer comprises three branches, the first branch adopts a conditional random field to extract keywords, the second branch uses a nonlinear transformation to generate a semantic feature vector, and the third branch is a multi-label classifier for generating prediction information, including predicted product categories, predicted attributes and predicted application scenarios.

8. The user demand-based plastic product recommendation method of claim 7, wherein, The processing process of the semantic understanding model for the current user demand is as follows: The input layer receives the current user demand, converts the text into a token sequence, and adds [CLS] and [SEP] markers, wherein [CLS] is used to capture the semantic information of the entire input, and [SEP] is used to distinguish different input segments; The embedding layer generates word embeddings, position embeddings and segment embeddings for each token, and adds the three kinds of embeddings to obtain an initial representation; The initial representation is processed by the multi-layer transformer encoder, each layer of which comprises: The multi-head self-attention calculates the attention weight of each token relative to other tokens to extract sentence-level semantic features. The feedforward neural network processes the attention output through nonlinear transformation. The output of the multi-layer transformer encoder is processed using a bidirectional long short-term memory network to add context information. The multi-scale feature extraction layer extracts features using multiple convolution kernels, and splices the features output by different convolution kernels to obtain multi-scale features. The adaptive pooling layer converts the multi-scale features into fixed-length semantic representations through a self-attention mechanism. In the multi-task output layer, the first branch uses a conditional random field to perform sequence labeling on each token to output a keyword list, the second branch converts the pooled semantic representation into a semantic feature vector through a nonlinear layer, and the third branch uses a multi-label classifier to perform structured prediction on product categories, key attributes and application scenarios.

9. The user demand-based plastic product recommendation method of claim 2, wherein, Step S6 comprises: S61 Based on the plastic industry knowledge graph, calculate the semantic relevance of each plastic product in the first list to the optimization query condition; S62 Combine user historical data to calculate the personalized score of each plastic product; S63 Use the multi-dimensional information of the plastic industry knowledge graph to calculate the similarity between plastic products; S64 combines the semantic relevance, the personalized score, and the similarity by weighting, obtains a comprehensive score of each plastic product, rearranges the first list according to the comprehensive score, and obtains a second list; S65 selects the top k options from the second list, generates a recommendation list, and presents the recommendation list to the user.

10. A plastic product recommendation system based on user needs, characterized by, The system is used to perform the method of any one of claims 1-9, and the system comprises: a knowledge graph construction module configured to extract information from a plastic product database, industry standard documents, and an expert knowledge base, form a multi-dimensional knowledge graph using entity recognition and relationship extraction, and perform entity alignment, relationship mapping, and graph fusion operations to generate a unified plastic industry knowledge graph; a data collection module configured to obtain historical search data, browsing data, interaction data, and transaction data from user interaction activities; a preprocessing module connected to the data collection module and configured to perform data cleaning, format standardization, and outlier removal on the obtained user historical data to generate a preprocessed user data set; a user modeling module configured to extract user historical behavior features by analyzing the preprocessed user data set, generate user interest topics, attribute preference vectors, price sensitivity indices, procurement behavior matrices, and probability scores of application scenarios; a semantic understanding module with a built-in semantic understanding model including an input layer, an embedding layer, a multi-layer transformer encoder, a context enhancement layer, a multi-scale feature extraction layer, an adaptive pooling layer, and a multi-task output layer, which performs semantic analysis on the current user demand, extracts keywords, semantic feature vectors, and prediction information; a query optimization module connected between the semantic understanding module and the knowledge graph construction module and configured to perform semantic expansion, weight adjustment, and condition optimization on the user query in combination with the user historical demand features and the plastic industry knowledge graph to generate an optimized query condition; a product search module configured to perform a search operation in the plastic product database according to the optimized query condition to generate a first list; a personalized sorting module connected to the product search module and configured to calculate a comprehensive score of each plastic product according to the semantic relevance, the personalized score, and the product similarity to sort and adjust the first list; a recommendation generation module configured to select a preset number of products from the first list after sorting and adjustment to generate a final recommendation list and output to the user; a feedback collection module for collecting user feedback on the recommendation results to optimize the plastic industry knowledge graph.

Images