Personalized recommendation system and method for chemical and plastic products combined with user portraits

Abstract

The present invention proposes a personalized recommendation system and method for chemical and plastic products in combination with user portraits, which relates to the field of personalized recommendation technology and includes: a data acquisition module for receiving search conditions input by users and acquiring historical data of users; a chemical and plastic product knowledge base module, including a chemical and plastic product database, for storing and managing chemical and plastic product information; a user portrait module for extracting multi-dimensional features of users, constructing user portraits, and dynamically updating user portraits according to a portrait update mechanism; a matching module for matching in the chemical and plastic product database according to search conditions to generate a preliminary product candidate set; a personalized recommendation module for calculating the recommendation score of products based on user portraits and preliminary product candidate sets, sorting products according to the recommendation score, and generating a final personalized recommendation result. The present invention can provide users with more accurate, timely and personalized chemical and plastic product recommendations.

Description

Technical Field

[0001] The present invention relates to the field of personalized recommendation technology, and in particular to a personalized recommendation system and method for plastic products combined with user portraits. Background Art

[0002] With the rapid development of the chemical and plastics industry, a large number of raw materials, recycled materials, modified materials, degradable materials, additives / pigments and fillers have appeared on the market. In order to help users quickly find the required materials among the numerous products, the e-commerce platform of the chemical and plastics industry began to introduce recommendation systems. The early recommendation systems were mainly based on simple keyword matching and product classification, providing users with basic search and screening functions.

[0003] With the advancement of technology, recommendation systems have gradually introduced more complex algorithms and data analysis technologies. Some advanced platforms have begun to use collaborative filtering, content-based recommendation and other methods to recommend relevant products to users based on their browsing history and purchase behavior. At the same time, considering the particularity of the chemical and plastics industry, some platforms have also begun to integrate market dynamics, price trends and other information to provide users with more comprehensive decision-making support.

[0004] However, the existing chemical and plastic product recommendation system still has shortcomings in dealing with the diversity of professional terminology, the dynamic changes in user needs, and the precise matching of product characteristics with user needs, making it difficult to provide users with truly personalized and efficient recommendation services. Summary of the invention

[0005] In view of this, the present invention proposes a personalized recommendation system and method for chemical and plastic products combined with user portraits. By constructing a multi-dimensional, dynamically updated user portrait, combined with a chemical and plastic product knowledge base and synonym expansion technology, accurate understanding and matching of user needs are achieved. The system adopts a multi-level labeling system, a dynamic weight adjustment mechanism, and a personalized recommendation algorithm, which can accurately capture changes in user interests and preferences, and consider factors such as product popularity to provide users with more accurate, timely and personalized chemical and plastic product recommendations. In addition, the modular design and scalability of the system enable it to adapt to the rapid development of the chemical and plastic industry and changes in user needs, thereby improving user experience and decision-making efficiency, and promoting accurate matching of chemical and plastic product transactions and improving market efficiency.

[0006] The technical solution of the present invention is achieved in this way:

[0007] On the one hand, the present invention provides a personalized recommendation system for chemical plastic products combined with user portraits, including:

[0008] The data acquisition module is used to receive the search conditions input by the user and obtain the user's historical data;

[0009] The chemical and plastic product knowledge base module includes a chemical and plastic product database for storing and managing chemical and plastic product information;

[0010] User portrait module, used to extract multi-dimensional features of users, build user portraits, and dynamically update user portraits according to the portrait update mechanism;

[0011] A matching module is used to match the search conditions in the chemical and plastic product database to generate a preliminary set of product candidates;

[0012] The personalized recommendation module is used to calculate the recommendation score of the product based on the user portrait and the preliminary product candidate set, sort the products according to the recommendation score, and generate the final personalized recommendation results.

[0013] Based on the above technical solution, preferably, the user portrait module includes:

[0014] The label system construction unit is used to establish the user and product label system of the chemical and plastic industry. The label system adopts a multi-level label system, which is divided into four levels, including basic labels, behavioral labels, interest preference labels and value labels;

[0015] A feature extraction unit is used to extract multi-dimensional features from the user's historical data, map them to the tag system, and generate a user tag set;

[0016] A weight distribution unit, used to generate a weight for each tag in the user tag set according to a weight generation method to obtain a weight set;

[0017] A portrait construction unit, used to construct a user portrait according to a user tag set and a weight set;

[0018] The portrait update mechanism is used to dynamically adjust the labels and weights in the user portrait based on the user's latest behavior and feedback, using a sliding window and attenuation factor to update the user portrait.

[0019] On the basis of the above technical solution, preferably, the user's historical data includes historical browsing records, historical search records, historical purchase records, historical collection records, historical evaluation records, historical consultation records and historical recommendation records; the user's multi-dimensional features include features of five dimensions, namely: basic attribute features, behavioral features, interest preference features, time series features and external data features; among which:

[0020] Basic attribute features include user ID, enterprise type, geographic location, enterprise size, and registration time;

[0021] Behavioral characteristics include search behavior, browsing behavior, transaction behavior, and social behavior;

[0022] Interest preference characteristics include material type preference, product attribute preference, market area preference, industry dynamics preference, and content interaction preference;

[0023] Time series features include purchasing cycles, active time periods, seasonal demand patterns, and behavioral change trends;

[0024] External data characteristics include industry macro-indicator sensitivity, social media behavior, and customer feedback and satisfaction.

[0025] Based on the above technical solution, preferably, the process of generating a user tag set includes:

[0026] Preprocess user historical data, including data cleaning, deduplication, and format unification, and divide the preprocessed data into text data and numerical data;

[0027] Use the WordVec model to segment text data and extract keywords, and normalize numerical data;

[0028] Perform feature extraction on the processed data to obtain a preliminary feature set;

[0029] According to the constructed multi-level label system, the preliminary feature set is mapped with the label system to generate the initial user label set;

[0030] Use association rule mining algorithms to discover the association between tags and add relevant tags;

[0031] Apply the tag fusion algorithm to merge tags with similar semantics and delete tags with low confidence to obtain the final user tag set.

[0032] Based on the above technical solution, preferably, the weight generation method includes:

[0033] Based on the frequency of occurrence of tags in the user's historical data , calculate the initial weight of each label :

[0034] ,

[0035] In the formula, n is the total number of user tags;

[0036] Calculate the time weight of the label :

[0037] ,

[0038] In the formula, is the time interval from the last appearance of tag i to the current one, is the time attenuation coefficient;

[0039] Calculate the inverse user frequency IUF of a tag:

[0040] ,

[0041] Where N is the total number of stored users, is the number of users with label i;

[0042] Generate the weight of the tag based on the initial weight, time weight and inverse user frequency:

[0043] ,

[0044] According to the above steps, a weight is generated for each tag in the user tag set to obtain a weight set.

[0045] Based on the above technical solution, preferably, the portrait update mechanism specifically includes:

[0046] Define the sliding window size as T and obtain the user behavior data within the last T time;

[0047] Assign a different weight coefficient k to each behavior type in the sliding window b , where behavior type b includes search behavior, browsing behavior, transaction behavior, and social behavior;

[0048] Calculate the weighted frequency of label i in the sliding window :

[0049] ,

[0050] In the formula, is the number of occurrences of label i in behavior type b;

[0051] Update time weight:

[0052] ,

[0053] In the formula, The time interval between the latest appearance of tag i and the current one;

[0054] Calculate new label weights:

[0055] ,

[0056] The new label weight Replace the original label weight and update the user portrait.

[0057] On the basis of the above technical solution, preferably, the matching module includes:

[0058] The synonym expansion unit is used to use the chemical plastic product synonym library to perform synonym expansion on the search conditions input by the user and generate an expanded search condition set;

[0059] The preliminary matching unit is used to match in the chemical and plastic product database according to the expanded search condition set to generate a preliminary product candidate set.

[0060] Based on the above technical solution, preferably, the synonym expansion unit includes:

[0061] Collect synonyms of various plastic products in the plastic industry, including aliases, common names, and trademark names of plastic products, and build an initial synonym database;

[0062] Collect the abbreviations of chemical and plastic products and their corresponding full names, including element symbols, molecular formula abbreviations, and industry abbreviations;

[0063] Add abbreviations and their corresponding full names as synonyms to the thesaurus;

[0064] The search conditions entered by the user are segmented, and the synonym set corresponding to each segmented word is searched in the synonym library, including abbreviations and full names. If there are unmatched segmented words, the edit distance algorithm is used for fuzzy matching;

[0065] For each word segmentation, all synonyms found in the synonym library are collected, and all the collected synonyms are merged to form an expanded search condition set.

[0066] Based on the above technical solution, preferably, the calculation formula for the product recommendation score is:

[0067] ,

[0068] In the formula, is the recommendation score of product p for user u, is the weight of user u on label i, is the characteristic value of product p on label i, is the heat factor of product p, is the adjustment parameter of the heat factor, and n is the number of tags in the user portrait.

[0069] On the other hand, the present invention also provides a method for personalized recommendation of plastic products in combination with user portraits, the method being executed in any of the above-mentioned systems, and the method comprising:

[0070] S1. Receive search conditions input by the user and obtain the user's historical data;

[0071] S2. Extract multi-dimensional features from the user's historical data, map them to a multi-level tag system, generate a user tag set, and assign weights to each tag through a weight generation method to build a user portrait;

[0072] S3, using the synonym database of chemical plastic products, performing synonym expansion on the search conditions input by the user to generate an expanded search condition set;

[0073] S4. Matching is performed in the chemical and plastic product database according to the expanded search condition set to generate a preliminary product candidate set;

[0074] S5. Based on the user portrait and the preliminary product candidate set, the recommendation score of each product is calculated, and the products are sorted according to the recommendation score to generate the final personalized recommendation result;

[0075] S6. Presenting personalized recommendation results to the user;

[0076] S7. Based on the user's latest behavior and feedback, a sliding window and attenuation factor are used to dynamically adjust the labels and weights in the user portrait to update the user portrait.

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] (1) The present invention achieves accurate understanding and matching of user needs by combining multi-dimensional user portraits, chemical and plastic product knowledge base and personalized recommendation algorithm, improves the accuracy and personalization of chemical and plastic product recommendations, and effectively solves the shortcomings of existing recommendation systems in dealing with the diversity of professional terms and dynamic changes in user needs;

[0079] (2) A comprehensive and detailed user portrait is constructed by adopting a multi-level labeling system and a multi-dimensional feature extraction method, which can more accurately describe user characteristics and improve the accuracy and comprehensiveness of user portraits;

[0080] (3) Introducing a portrait update mechanism, dynamically adjusting the labels and weights in the user portrait through sliding windows and attenuation factors, so that the system can capture changes in user interests and preferences in a timely manner, improving the adaptability and timeliness of the recommendation system;

[0081] (4) Using the synonym database of chemical and plastic products to expand the synonyms of search conditions effectively solves the problem of diversity of professional terms in the chemical and plastic industry, improves the coverage and accuracy of product matching, and enables users to find the required products more conveniently;

[0082] (5) The user profile matching degree and product popularity factor are comprehensively considered in the calculation of recommendation scores, which not only ensures the personalization of recommendations, but also takes into account the market performance of the product, thereby improving the overall quality of the recommendation results and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0084] Figure 1 It is a system framework diagram of the present invention;

[0085] Figure 2 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0086] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0087] like Figure 1 As shown, the present invention provides a personalized recommendation system for plastic products combined with user portraits, including:

[0088] The data acquisition module is used to receive the search conditions input by the user and obtain the user's historical data;

[0089] The chemical and plastic product knowledge base module includes a chemical and plastic product database for storing and managing chemical and plastic product information;

[0090] User portrait module, used to extract multi-dimensional features of users, build user portraits, and dynamically update user portraits according to the portrait update mechanism;

[0091] A matching module, used to match the search conditions in a database including chemical and plastic products to generate a preliminary set of product candidates;

[0092] The personalized recommendation module is used to calculate the recommendation score of the product based on the user portrait and the preliminary product candidate set, sort the products according to the recommendation score, and generate the final personalized recommendation results.

[0093] Specifically, in one embodiment of the present invention, the data acquisition module serves as the front-end interface of the system and is responsible for receiving search conditions input by the user through the interface. These search conditions include keywords or phrases such as the name, type, specification, and performance parameters of the plastic product.

[0094] In this embodiment, the user's historical data includes historical browsing records, historical search records, historical purchase records, historical collection records, historical evaluation records, historical consultation records, and historical recommendation records. Historical browsing records: chemical and plastic product pages that users have browsed on the platform; historical search records: search keywords that users have previously entered; historical purchase records: chemical and plastic product transaction information completed by users on the platform; historical collection records: chemical and plastic products that users have collected or marked as interested in; historical evaluation records: user evaluation information on purchased chemical and plastic products; historical consultation records: user communication records with customer service or sellers; historical recommendation records: chemical and plastic product information that the system has previously recommended to users.

[0095] These historical data cover various behaviors and interactions of users on the platform, providing a rich data foundation for building a comprehensive user profile. The data acquisition module extracts these historical data from various relevant databases or logs of the system, and performs preliminary sorting and organization to prepare for subsequent user profile construction and personalized recommendations.

[0096] Specifically, in one embodiment of the present invention, a chemical plastic product knowledge base module is configured in the system, which mainly includes a chemical plastic product database for storing and managing chemical plastic product information. The functions of this module are as follows:

[0097] The chemical and plastic product knowledge base covers various types of chemical and plastic products, including raw materials, recycled materials, modified materials, degradable materials, additives / pigments and fillers, etc. In this embodiment, the system provides more than 120,000 spot product information in real time.

[0098] Store detailed product information. For each plastic product, the stored information includes: product name and type, specifications and model (such as PC-1100, SC-1100R, etc.), manufacturer, performance parameters, price information (such as 15700.00 yuan / ton), inventory status, product description and application areas.

[0099] The knowledge base not only stores static product information, but also contains dynamic market data: real-time price updates, historical price trends, market dynamics, crude oil dynamics, and device dynamics. As well as regional and manufacturer information, in this embodiment, the system covers multiple regional markets (such as the East China market, the South China market, etc.) and many manufacturers, including information on more than 40 cities and more than 400 manufacturers.

[0100] Grade management, knowledge base provides users with detailed grade information and comparison.

[0101] The knowledge base has a timely update mechanism that can reflect the latest product information and market trends in real time.

[0102] Store and manage supplier related information such as company name, contact information, credit rating, etc.

[0103] Specifically, in one embodiment of the present invention, the user portrait module includes:

[0104] The label system construction unit is used to establish the user and product label system of the chemical and plastic industry. The label system adopts a multi-level label system, which is divided into four levels, including basic labels, behavioral labels, interest preference labels and value labels;

[0105] A feature extraction unit is used to extract multi-dimensional features from the user's historical data, map them to the tag system, and generate a user tag set;

[0106] A weight distribution unit, used to generate a weight for each tag in the user tag set according to a weight generation method to obtain a weight set;

[0107] A portrait construction unit, used to construct a user portrait according to a user tag set and a weight set;

[0108] The portrait update mechanism is used to dynamically adjust the labels and weights in the user portrait based on the user's latest behavior and feedback, using a sliding window and attenuation factor to update the user portrait.

[0109] In this embodiment, the label system construction process is as follows: collect user data and product data related to the plastics industry; perform statistical analysis on the collected data to identify key features and attributes; define labels at each level based on the analysis results; classify the defined labels into four levels; determine the hierarchical relationship and association relationship between labels; ensure the accuracy and completeness of the label system through data verification and expert review; and establish a regular update mechanism to adapt to industry changes and emerging needs.

[0110] In a specific embodiment, the labeling system is exemplified as follows:

[0111] Basic tags: user ID, enterprise type (such as manufacturer, trader, end user), geographic location, enterprise size, registration time;

[0112] Behavioral tags: search frequency, browsing depth, transaction amount, social activity;

[0113] Interest preference tags:

[0114] Material type preference: PP, PE, ABS, PC, etc.; Product attribute preference: strength, toughness, heat resistance, transparency, etc.; Market region preference: East China, South China, North China, etc.; Industry dynamic preference: price trend, new product launch, technological innovation, etc.;

[0115] Value labels: customer value level (such as diamond, gold, silver), loyalty index, repurchase rate.

[0116] In this embodiment, the multi-dimensional features of the user include features in five dimensions, namely: basic attribute features, behavior features, interest preference features, time series features and external data features; among which:

[0117] Basic attribute features include user ID, enterprise type, geographic location, enterprise size, and registration time;

[0118] Behavioral characteristics include search behavior, browsing behavior, transaction behavior, and social behavior;

[0119] Interest preference characteristics include material type preference, product attribute preference, market area preference, industry dynamics preference, and content interaction preference;

[0120] Time series features include purchasing cycles, active time periods, seasonal demand patterns, and behavioral change trends;

[0121] External data characteristics include industry macro-indicator sensitivity, social media behavior, and customer feedback and satisfaction.

[0122] The process of generating a user tag set includes:

[0123] Preprocess user historical data, including data cleaning, deduplication, and format unification, and divide the preprocessed data into text data and numerical data;

[0124] Use the WordVec model to segment text data and extract keywords, and normalize numerical data;

[0125] Perform feature extraction on the processed data to obtain the preliminary feature set F={f1,f2,...,f n}; Among them, the time series features are extracted using the sliding window method, and the external data features are extracted using the statistical analysis method.

[0126] According to the constructed multi-level label system T, the preliminary feature set is mapped to the label system to generate the initial user label set S0={t1,t2,...,t m}, where t i ∈T; when mapping, the basic attribute features are mapped to the basic label level, the behavior features and time series features are mapped to the behavior label level, the interest preference features are mapped to the interest preference label level, and the external data features are mapped to the value label level. The decision tree algorithm is used for mapping, M:F→T.

[0127] Use the association rule mining algorithm to discover the association relationship between tags and supplement the relevant tags; specifically, the improved FP-Growth algorithm discovers the association rules between tags and sets the minimum support min sup and minimum confidence minconf , generate association rule set R={r|support(r)≥min sup ,confidence(r)≥min conf}, for each association rule r∈R, if r.antecedent S0 and r.consequent S0, then add r.consequent to S0 to get the supplemented tag set S1. Among them, support represents the association support, which calculates the association support between two tags.

[0128] The tag fusion algorithm is applied to merge tags with similar semantics and delete tags with low confidence to obtain the final user tag set. Specifically, the Word2Vec algorithm is used to calculate the semantic similarity between tags. For any two tags t i ,t j ∈S1, if sim(t i ,t j )≥sim threshold , then merge these two labels, the merging function: merge(t i ,t j )=(t i +t j ) / 2, and the fused tag set S2 is obtained. The confidence of each tag is calculated, tags with confidence lower than the threshold are deleted, and all processed tags are integrated to form the final user tag set.

[0129] In this embodiment, the weight generation method includes:

[0130] Based on the frequency of occurrence of tags in the user's historical data , calculate the initial weight of each label :

[0131] ,

[0132] Where n is the total number of user tags;

[0133] Calculate the time weight of the label :

[0134] ,

[0135] In the formula, is the time interval from the last appearance of tag i to the current one, is the time attenuation coefficient;

[0136] Calculate the inverse user frequency IUF of a tag:

[0137] ,

[0138] Where N is the total number of stored users, is the number of users with label i;

[0139] Generate the weight of the tag based on the initial weight, time weight and inverse user frequency:

[0140] ,

[0141] According to the above steps, a weight is generated for each tag in the user tag set to obtain a weight set.

[0142] Combine each label and its corresponding weight into a label-weight pair, design a multi-level portrait structure, correspond to the five levels of the label system, fill the label-weight pairs into the corresponding portrait structure level, normalize the weight of each level to ensure that the sum of the weights is 1, use principal component analysis (PCA) or autoencoder to reduce the dimensionality of the high-dimensional portrait, and store the constructed user portrait in the database.

[0143] In this embodiment, in order to ensure the timeliness of the user portrait, a dynamic portrait update mechanism is provided, and the portrait update mechanism specifically includes:

[0144] Define the sliding window size as T and obtain the user behavior data within the last T time;

[0145] Assign a different weight coefficient k to each behavior type in the sliding window b , where behavior type b includes search behavior, browsing behavior, transaction behavior, and social behavior;

[0146] Calculate the weighted frequency of label i in the sliding window :

[0147] ,

[0148] In the formula, is the number of occurrences of label i in behavior type b;

[0149] Update time weight:

[0150] ,

[0151] In the formula, The time interval between the latest appearance of tag i and the current one;

[0152] Calculate new label weights:

[0153] ,

[0154] The new label weight Replace the original label weights. If new labels appear, add them to the user portrait. If the weights of some labels drop below the threshold, consider removing them from the portrait to update the user portrait.

[0155] Specifically, in one embodiment of the present invention, the matching module includes:

[0156] The synonym expansion unit is used to use the chemical plastic product synonym library to perform synonym expansion on the search conditions input by the user and generate an expanded search condition set;

[0157] The preliminary matching unit is used to match in the chemical and plastic product database according to the expanded search condition set to generate a preliminary product candidate set.

[0158] In this embodiment, the synonym expansion unit includes:

[0159] Collect synonyms of various plastic products in the plastic industry, including aliases, common names, and trademark names of plastic products, and build an initial synonym database;

[0160] Collect the abbreviations of chemical and plastic products and their corresponding full names, including element symbols, molecular formula abbreviations, and industry abbreviations;

[0161] Add abbreviations and their corresponding full names as synonyms to the thesaurus.

[0162] In a specific example, the specific construction process of the chemical plastic product synonym library is as follows:

[0163] Data collection: Collect chemical and plastic product terms from authoritative documents, professional dictionaries, and industry standard documents in the chemical and plastic industry. Crawl the official websites of major chemical and plastic product suppliers and manufacturers to obtain product names and descriptions. Collect patent documents and academic papers related to the chemical and plastic industry and extract professional terms from them.

[0164] Data preprocessing: Clean the collected data, remove duplicates and irrelevant information, perform word segmentation, and extract keywords and phrases.

[0165] Synonym identification: Use Word2Vec or BERT model to calculate the semantic similarity between words. Set a similarity threshold and treat word pairs above the threshold as potential synonyms. Use existing general synonym dictionaries (such as WordNet) to assist in identification.

[0166] Expert review: Chemical and plastics industry experts review and correct automatically identified synonyms and supplement industry-specific synonym relationships, such as the correspondence between chemical names and trade names.

[0167] Abbreviation processing: Collect common abbreviations of plastic products and their corresponding full names, including element symbols (such as "C" for "carbon"), molecular formula abbreviations (such as "PE" for "polyethylene"), and industry abbreviations (such as "ABS" for "acrylonitrile-butadiene-styrene copolymer").

[0168] Structured storage: Design a suitable data structure to store synonym relationships, such as using a graph database or a relational database. Create a unique identifier for each entry to facilitate subsequent rapid retrieval and update.

[0169] Continuous update mechanism: Set up a regular update mechanism, such as monthly or quarterly updates. Establish user feedback channels to collect new synonyms or error correction information provided by users.

[0170] In another embodiment of the present invention, in the synonym expansion unit, the synonym library can be a synonym network map of chemical plastic products, and the unit uses the following method to expand synonyms:

[0171] (a) Construct a synonym network map of chemical and plastic products. Take the various aliases, chemical names, trade names, common names, etc. of chemical and plastic products as nodes, and use the associations between products to build connecting edges to form a network map G(V,E), where V is the node set and E is the edge set.

[0172] Specifically, this embodiment takes into account the diversity of chemical and plastic product names, including chemical names, trade names, common names, etc. When constructing a synonym network map of chemical and plastic products, the chemical and plastic product knowledge base is used to enhance the construction of the synonym network map. The process is as follows:

[0173] 1. The chemical and plastic product knowledge base is recorded as KB.

[0174] 2. Extract term set V from multi-source data.

[0175] 3. For each term , use KB for entity linking and obtain enhanced entity representation:

[0176] ,

[0177] in, is the textual representation of the term, is the representation of the corresponding entity in the knowledge base.

[0178] 4. Calculate the cosine similarity SIM between the enhanced terms.

[0179] 5. Construct the edge set E based on similarity to form the initial graph G(V,E).

[0180] 6. Design a multi-level graph structure and define the hierarchy L={L1,L2,...,Ln}, where L1 is the smallest particle size, as in the IUPAC name, L n is the maximum granularity, such as a common name.

[0181] 7. For each node v in the initial graph G(V,E), determine the level L(v) to which it belongs.

[0182] 8. Construct the inner edge E of the layer intra and the interlayer edge E inter :

[0183] for ,if ,but ;

[0184] for , if there is a corresponding relationship, then .

[0185] 9. Repeat steps 7 and 8 to get the final graph structure .

[0186] (b) Extract keywords from the user's search conditions to form the original query word q, locate the node of query word q in graph G, take the node of query word q as the starting point, perform a random walk on graph G, generate a sequence of candidate nodes, use the Skip-gram model to learn the embedding vector of each node, and calculate the similarity between query word q and candidate node c. The formula is as follows:

[0187] ,

[0188] In the formula, is the similarity between query word q and candidate node c, is the node embedding vector, , is the degree of the node. The degree of a node refers to the number of edges directly connected to the node in the graph G.

[0189] (c) Select the top-m words with the highest similarity as the synonym expansion results.

[0190] The search conditions entered by the user are segmented, and the synonym set corresponding to each segmented word is searched in the synonym library, including abbreviations and full names. If there are unmatched segmented words, the edit distance algorithm is used for fuzzy matching.

[0191] Use the word segmenter of the chemical and plastic industry to segment the user input. Perform an exact match in the synonym library for each word segment. Use a hash table or prefix tree Trie data structure to accelerate the matching process. For word segments that fail to match exactly, use an improved edit distance algorithm for fuzzy matching. The traditional edit distance algorithm may not be accurate enough when processing chemical names and professional terms. Therefore, the present invention improves the edit distance algorithm as follows: assign special weights to chemical elements so that the cost of element replacement is higher. For example, the cost of mistakenly replacing "C" with "O" should be higher than the replacement of ordinary letters. Perform special processing on the numbers in the chemical formula to allow the increase and decrease of numbers. For example, the edit distance of "C2H4" and "C2H5" should be small. Identify common roots in chemical names, such as "ethyl", "phenyl", etc., and maintain the integrity of the roots when calculating the edit distance. Before calculating the edit distance, use the abbreviation expansion rule to expand the abbreviation to the full name. For example, expand "PE" to "polyethylene" and then compare.

[0192] The steps of the improved edit distance algorithm are as follows:

[0193] 1. Initialize the distance matrix: create a matrix of size (m+1)*(n+1), where m and n are the lengths of the two strings respectively.

[0194] 2. Set weights: For alphabetic characters, set the weight to 1. For numeric characters, set the weight to 0.5. For chemical symbols (such as "-", "()"), set the weight to 0.8.

[0195] 3. Filling matrix: Calculate the weighted distance of insertion, deletion, and substitution operations. If the current character is a chemical keyword, the weight is halved when matching.

[0196] 4. Get the edit distance: The value in the lower right corner of the matrix is ​​the weighted minimum edit distance.

[0197] 5. Normalization: Normalize the edit distance to a similarity score ranging between 0 and 1.

[0198] For each word segmentation, all synonyms found in the synonym library are collected, and all the collected synonyms are merged to form an expanded search condition set.

[0199] Afterwards, the preliminary matching unit matches the chemical and plastic product database according to the expanded search condition set to generate a preliminary product candidate set. The specific process is as follows:

[0200] 1. Index construction: Use inverted index to create an index for each product in the chemical and plastic product database.

[0201] 2. Multi-dimensional matching: Match each condition in the expanded search condition set in the index.

[0202] 3. Relevance score: Calculate a relevance score for each matching result.

[0203] 4. Filtering mechanism: Set the relevance score threshold to filter out low-relevance matching results.

[0204] 5. Result sorting: Sort the matching results according to the relevance score.

[0205] 6. Candidate set generation: Select the top N results after sorting as the preliminary product candidate set. The value of N can be dynamically adjusted based on system performance and user experience.

[0206] Specifically, in one embodiment of the present invention, the personalized recommendation module calculates the recommendation score of the product based on the user portrait and the preliminary product candidate set. The calculation formula is as follows:

[0207] ,

[0208] In the formula, is the recommendation score of product p for user u, is the weight of user u on label i, is the characteristic value of product p on label i, is the heat factor of product p, is the adjustment parameter of the heat factor, and n is the number of tags in the user portrait.

[0209] In this embodiment, The weight is the label weight. Each label in the user portrait has a corresponding weight, which reflects the importance of the feature to the user. The calculation of the weight takes into account the frequency of occurrence of the label in the user's historical data, the time decay factor and the inverse user frequency. Labels with high weights have a greater impact on the recommendation results. The system analyzes the features of each candidate product and matches it with the label in the user portrait. The feature value of product p on label i can be a binary value (0 or 1) or a continuous value (between 0 and 1). The higher the feature value, the higher the match between the product and the user profile on the label. Calculate the degree of match between the user portrait and product features. For each label in the user portrait, multiply the user's weight on the label by the product's feature value on the label. The sum of the matching degrees of all labels represents the overall user-product matching degree. Indicates the popularity of product p, which is calculated based on factors such as the product's total sales, recent sales, and user reviews.

[0210] In this embodiment, the cold start problem is handled as follows: for new users or users with less data, the system will rely more on the product's popularity factor. ,With the accumulation of user behavior data, the weight of personalized tags is gradually increased.

[0211] For each user u, after calculating the recommendation score of each product relative to user u, the preliminary product candidate set is sorted in descending order of the recommendation score, and the Top-K products are selected as the personalized recommendation results for user u this time.

[0212] After the final personalized recommendation results are formed into a recommendation list and displayed to the user, the user's response to the recommendation results (click, purchase, favorite, etc.) is recorded. Based on user feedback, the label weights in the user portrait are dynamically adjusted. Based on a large amount of user feedback data, the parameters in the recommendation score calculation formula are continuously optimized.

[0213] In addition, if Figure 2 As shown, the present invention also provides a personalized recommendation method for plastic products combined with user portraits, the method is executed in any of the above-mentioned systems, and the method includes:

[0214] S1. Receive search conditions input by the user and obtain the user's historical data.

[0215] Design a user-friendly search interface that supports multiple input methods (such as keywords, product categories, attribute filtering, etc.). Implement real-time search suggestion functions to improve user input efficiency. Extract user historical data from the system database, including: historical browsing records, historical search records, historical purchase records, historical collection records, historical evaluation records, historical consultation records, and historical recommendation records.

[0216] S2. Extract multi-dimensional features from the user's historical data, map them to a multi-level tag system, generate a user tag set, and assign weights to each tag through a weight generation method to build a user portrait.

[0217] Multi-dimensional feature extraction:

[0218] Basic attribute features: extracted from user registration information. Behavioral features: analyze user historical data to extract search, browsing, transaction and social behavior patterns. Interest preference features: based on user historical interaction data, identify user preferences for different material types, product attributes, market regions, etc. Time series features: analyze the time pattern of user behavior, identify procurement cycles, active time periods, etc. External data features: integrate external information such as industry macro indicators and social media data.

[0219] Use a predefined multi-level label system, including basic labels, behavior labels, interest preference labels, and value labels, to map the extracted features to the corresponding labels.

[0220] Use the WordVec model to segment text data and extract keywords. Normalize numerical data. Apply the association rule mining algorithm to discover the association between tags and add relevant tags. Use the tag fusion algorithm to merge tags with similar semantics and delete tags with low confidence.

[0221] Based on the frequency of occurrence of tags in the user's historical data , calculate the initial weight of each label :

[0222] ,

[0223] Where n is the total number of user tags;

[0224] Calculate the time weight of the label :

[0225] ,

[0226] In the formula, is the time interval from the last appearance of tag i to the current one, is the time attenuation coefficient;

[0227] Calculate the inverse user frequency IUF of a tag:

[0228] ,

[0229] Where N is the total number of stored users, is the number of users with label i;

[0230] Generate the weight of the tag based on the initial weight, time weight and inverse user frequency:

[0231] ,

[0232] According to the above steps, a weight is generated for each tag in the user tag set to obtain a weight set.

[0233] Combine the tag set and the corresponding weight set to form a user profile.

[0234] S3. Using the synonym database of chemical plastic products, the search conditions input by the user are expanded with synonyms to generate an expanded search condition set.

[0235] Collect synonyms for products in the plastics industry, including aliases, common names, and trademark names. Collect abbreviations for plastics products and their corresponding full names.

[0236] Segment the search conditions entered by the user. Search the synonym set corresponding to each segmentation in the thesaurus. For unmatched segmentations, use the edit distance algorithm for fuzzy matching.

[0237] All found synonyms are merged to form an expanded set of search terms.

[0238] S4. According to the expanded search condition set, matching is performed in the chemical and plastic product database to generate a preliminary product candidate set.

[0239] Use the expanded search criteria set to query the chemical and plastic product database. Use efficient database indexing and query optimization technology to improve matching speed. For each matched product, calculate its relevance score with the search criteria. Based on the relevance score, select the top-N products as the preliminary product candidate set.

[0240] S5. Based on the user portrait and the preliminary product candidate set, calculate the recommendation score of each product, sort the products according to the recommendation score, and generate the final personalized recommendation results.

[0241] Based on the user portrait and the preliminary product candidate set, the product recommendation score is calculated using the following formula:

[0242] ,

[0243] In the formula, is the recommendation score of product p for user u, is the weight of user u on label i, is the characteristic value of product p on label i, is the heat factor of product p, is the adjustment parameter of the heat factor, and n is the number of tags in the user portrait. This includes: extracting the features of each candidate product, matching them with the tags in the user portrait, and calculating the feature value of each tag. , calculate the popularity factor based on the product's total sales, recent sales, user reviews, etc. .

[0244] Sort the products in descending order according to their ratings. Select the top K products after sorting as the final personalized recommendation results.

[0245] S6. Present personalized recommendation results to the user.

[0246] Design an intuitive user interface to display a list of recommended products. Display key information (such as name, image, price, and short description) for each recommended product. Provide personalized reasons for each recommended product based on relevant tags in the user portrait. Implement functions such as viewing product details, adding to favorites, and adding to shopping carts. Provide a feedback mechanism to allow users to evaluate the recommendation results.

[0247] S7. Based on the user's latest behavior and feedback, a sliding window and attenuation factor are used to dynamically adjust the labels and weights in the user portrait to update the user portrait.

[0248] Record users’ responses to recommendation results (clicks, purchases, favorites, etc.) in real time.

[0249] Define the sliding window size T, and only consider the user behavior data within the last T time. Assign different weight coefficients k to different types of behaviors (search, browse, trade, social) b , calculate the weighted frequency of the label in the sliding window : , update time weight: , calculate the new label weight: , replace the original weights with the newly calculated tag weights and update the user profile. Set up a regular comprehensive update mechanism, such as a comprehensive reconstruction of the user profile once a month.

[0250] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A personalized recommendation system for chemical and plastic products based on user portraits, characterized by: include: The data acquisition module is used to receive the search conditions input by the user and obtain the user's historical data; The chemical and plastic product knowledge base module includes a chemical and plastic product database for storing and managing chemical and plastic product information; User portrait module, used to extract multi-dimensional features of users, build user portraits, and dynamically update user portraits according to the portrait update mechanism; The user portrait module includes: The label system construction unit is used to establish the user and product label system of the chemical and plastic industry. The label system adopts a multi-level label system, which is divided into four levels, including basic labels, behavioral labels, interest preference labels and value labels; A feature extraction unit is used to extract multi-dimensional features from the user's historical data, map them to the tag system, and generate a user tag set; A weight distribution unit, used to generate a weight for each tag in the user tag set according to a weight generation method to obtain a weight set; A portrait construction unit, used to construct a user portrait according to a user tag set and a weight set; The portrait update mechanism is used to dynamically adjust the labels and weights in the user portrait based on the user's latest behavior and feedback, using a sliding window and attenuation factor to update the user portrait; Weight generation methods include: Based on the frequency of occurrence of tags in the user's historical data , calculate the initial weight of each label : ， In the formula, n is the total number of user tags; Calculate the time weight of the label : ， In the formula, is the time interval from the last appearance of tag i to the current one, is the time attenuation coefficient; Calculate the inverse user frequency IUF of a tag: ， Where N is the total number of stored users, is the number of users with label i; Generate the weight of the tag based on the initial weight, time weight and inverse user frequency: ， According to the above steps, a weight is generated for each tag in the user tag set to obtain a weight set; The portrait update mechanism specifically includes: Define the sliding window size as T and obtain the user behavior data within the last T time; Assign a different weight coefficient k to each behavior type in the sliding window b , where behavior type b includes search behavior, browsing behavior, transaction behavior, and social behavior; Calculate the weighted frequency of label i in the sliding window : ， In the formula, is the number of occurrences of label i in behavior type b; Update time weight: ， In the formula, The time interval between the latest appearance of tag i and the current one; Calculate new label weights: ， The new label weight Replace the original label weight and update the user portrait; A matching module is used to match the search conditions in the chemical and plastic product database to generate a preliminary set of product candidates; The matching modules include: The synonym expansion unit is used to use the chemical plastic product synonym library to perform synonym expansion on the search conditions input by the user and generate an expanded search condition set; A preliminary matching unit, used to match in the chemical and plastic product database according to the expanded search condition set to generate a preliminary product candidate set; Among them, the synonym database of chemical and plastic products is a synonym network map of chemical and plastic products. When constructing the synonym network map of chemical and plastic products, the chemical and plastic product knowledge base is used to enhance the construction of the synonym network map. The process is as follows:

1. The chemical and plastic product knowledge base is recorded as KB; 2. Extract term set V from multi-source data; 3. For each term , use KB for entity linking and obtain enhanced entity representation: ， in, is the textual representation of the term, is the representation of the corresponding entity in the knowledge base; 4. Calculate the cosine similarity SIM between the enhanced terms; 5. Construct edge set E based on similarity to form the initial graph G(V,E); 6. Design a multi-level graph structure and define the hierarchy L={L1,L2,...,L n }, where L1 is the smallest particle size, as in the IUPAC name, L n is the maximum granularity, such as the common name; 7. For each node v in the initial graph G(V,E), determine the level L(v) to which it belongs; 8. Construct the inner edge E of the layer intra and the interlayer edge E inter : for ,if ,but ; for , if there is a corresponding relationship, then ; 9. Repeat steps 7 and 8 to get the final graph structure ; The personalized recommendation module is used to calculate the recommendation score of the product based on the user portrait and the preliminary product candidate set, sort the products according to the recommendation score, and generate the final personalized recommendation results.

2. The personalized recommendation system for chemical and plastic products combined with user portraits as claimed in claim 1, characterized in that: The user's historical data includes historical browsing records, historical search records, historical purchase records, historical collection records, historical evaluation records, historical consultation records and historical recommendation records; the user's multi-dimensional characteristics include characteristics of five dimensions, namely: basic attribute characteristics, behavioral characteristics, interest preference characteristics, time series characteristics and external data characteristics; among them: Basic attribute features include user ID, enterprise type, geographic location, enterprise size, and registration time; Behavioral characteristics include search behavior, browsing behavior, transaction behavior, and social behavior; Interest preference characteristics include material type preference, product attribute preference, market area preference, industry dynamics preference, and content interaction preference; Time series features include purchasing cycles, active time periods, seasonal demand patterns, and behavioral change trends; External data characteristics include industry macro-indicator sensitivity, social media behavior, and customer feedback and satisfaction.

3. The personalized recommendation system for chemical and plastic products combined with user portraits as claimed in claim 1, characterized in that: The process of generating a user tag set includes: Preprocess user historical data, including data cleaning, deduplication, and format unification, and divide the preprocessed data into text data and numerical data; Use the WordVec model to segment text data and extract keywords, and normalize numerical data; Perform feature extraction on the processed data to obtain a preliminary feature set; According to the constructed multi-level label system, the preliminary feature set is mapped with the label system to generate the initial user label set; Use association rule mining algorithms to discover the association between tags and add relevant tags; Apply the tag fusion algorithm to merge tags with similar semantics and delete tags with low confidence to obtain the final user tag set.

4. The personalized recommendation system for chemical and plastic products combined with user portraits according to claim 1 is characterized in that: Synonym expansion units include: Collect synonyms of various plastic products in the plastic industry, including aliases, common names, and trademark names of plastic products, and build an initial synonym database; Collect the abbreviations of chemical and plastic products and their corresponding full names, including element symbols, molecular formula abbreviations, and industry abbreviations; Add abbreviations and their corresponding full names as synonyms to the thesaurus; The search conditions entered by the user are segmented, and the synonym set corresponding to each segmented word is searched in the synonym library, including abbreviations and full names. If there are unmatched segmented words, the edit distance algorithm is used for fuzzy matching; For each word segmentation, all synonyms found in the synonym library are collected, and all the collected synonyms are merged to form an expanded search condition set.

5. The personalized recommendation system for chemical and plastic products combined with user portraits according to claim 1 is characterized in that: The formula for calculating the recommendation score of a product is: ， In the formula, is the recommendation score of product p for user u, is the weight of user u on label i, The feature value of product p on label i, is the heat factor of product p, is the adjustment parameter of the heat factor, and n is the number of tags in the user portrait.

6. A personalized recommendation method for plastic products based on user portraits, the method being executed in the system according to any one of claims 1 to 5, the method comprising: S1. Receive search conditions input by the user and obtain the user's historical data; S2. Extract multi-dimensional features from the user's historical data, map them to a multi-level tag system, generate a user tag set, and assign weights to each tag through a weight generation method to build a user portrait; S3, using the synonym database of chemical plastic products, performing synonym expansion on the search conditions input by the user to generate an expanded search condition set; S4. Matching is performed in the chemical and plastic product database according to the expanded search condition set to generate a preliminary product candidate set; S5. Based on the user portrait and the preliminary product candidate set, the recommendation score of each product is calculated, and the products are sorted according to the recommendation score to generate the final personalized recommendation result; S6. Presenting personalized recommendation results to the user; S7. Based on the user's latest behavior and feedback, a sliding window and attenuation factor are used to dynamically adjust the labels and weights in the user portrait to update the user portrait.

Images