Plasticizing product mixed recommendation system combining user behavior and product content

Abstract

The invention discloses a mixed recommendation system for plasticized products in combination with user behaviors and product contents, which relates to the field of big data analysis, and comprises an acquisition and analysis module used for connecting a plasticized product sales platform, acquiring browsing, clicking, collecting and trading behavior data of platform users on plasticized products in the sales platform, and sending the acquired behavior data to a user terminal; carrying out structured analysis on the data; the extraction module is used for extracting material components, physical properties, application scenes and specification parameter data of the on-sale plastic products in the sales platform so as to construct feature vectors of the plastic products; according to the method, browsing, clicking, collecting and trading behaviors of a user on a plastic product are accurately captured and converted into behavior scores, meanwhile, material components, physical properties, application scenes and specification parameters of the product are integrated to construct feature vectors, and a user-product mixed feature basis meeting requirements is formed through association mapping; user preference coefficients are updated in real time based on dynamic matching logic, and a precise initial recommendation sequence is generated.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, specifically to a hybrid recommendation system for chemical and plastic products that combines user behavior and product content. Background Technology

[0002] Chemicals and Plastics' product sales cover basic resins such as PE and PP, as well as engineering plastics and environmentally friendly materials, meeting the needs of multiple fields such as packaging, automotive, and medical and health. Supported by production, sales, research and application, it provides customized solutions, ensures supply through online platforms and logistics networks, and helps the industrial chain reduce costs and increase efficiency.

[0003] The invention patent application with application number 202411555131.6 discloses a personalized recommendation system for chemical and plastic products that combines user profiles. This application aims to solve the problem that "some chemical and plastic product recommendation systems still have shortcomings in handling the diversity of professional terms, dynamic changes in user needs, and accurate matching of product characteristics with user needs, making it difficult to provide users with truly personalized and efficient recommendation services."

[0004] However, in the chemical and plastic products sales industry, users find it difficult to quickly find products they are interested in on sales platforms that contain massive amounts of product information. Therefore, they are eager to use intelligent systems to conduct refined comparisons and screenings of chemical and plastic products.

[0005] To address this, we propose a hybrid recommendation system for chemical and plastic products that combines user behavior and product content. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a hybrid recommendation system for chemical and plastic products that combines user behavior and product content, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions; This invention discloses a hybrid recommendation system for chemical and plastic products that combines user behavior and product content, comprising: The system comprises the following modules: a data acquisition and analysis module, which connects to a chemical and plastic product sales platform to collect user browsing, clicking, collection, and transaction data related to chemical and plastic products, and performs structured analysis on this data; an extraction module, which extracts material composition, physical properties, application scenarios, and specifications of chemical and plastic products sold on the platform to construct feature vectors for these products; a fusion module, which receives user behavior analysis results from the acquisition and analysis module and chemical and plastic product feature vectors from the extraction module, performs correlation mapping between the two, and constructs a user-product hybrid feature matrix; a generation module, which receives the user-product hybrid feature matrix and, based on this matrix, generates a chemical and plastic product recommendation model by constructing a dynamic matching logic between user preferences and product features; a correction module, which dynamically corrects the initial recommendation results of the recommendation model based on inventory fluctuations, real-time changes in user behavior, and product relevance; and an output module, which converts the corrected initial recommendation results into a document file according to a preset format and pushes it to a preset user-held mobile device. The acquisition and analysis module is interconnected with an extraction module via a wireless network. The extraction module is interconnected with a fusion module and a generation module via a wireless network. The fusion module and the generation module are interconnected with a correction module via a wireless network. The correction module is interconnected with an output module via a wireless network.

[0008] Furthermore, the operation process of the acquisition and parsing module when performing structured parsing of user behavior data is as follows: Dynamic behavior weighting coefficients are assigned to four user behaviors: browsing, clicking, favoriting, and trading. Using a preset time window as a unit, the frequency of each user's four behaviors towards each chemical and plastic product within that time window is counted, and the behavior frequency is multiplied by the corresponding weight coefficient. The product result is recorded as the user's behavior score for each product. A two-dimensional matrix is ​​constructed based on user ID and product ID, with the matrix elements representing the user's behavior score for the corresponding product, in order to form a structured user behavior analysis result. Among them, the weight coefficient of transaction behavior is greater than that of collection behavior, the weight coefficient of collection behavior is greater than that of click behavior, the weight coefficient of click behavior is greater than that of browsing behavior, and the sum of the four weights is 1.

[0009] Furthermore, during the extraction module's operation phase, when constructing the feature vector of the chemical and plastic products, the following applies: The material composition data is classified and coded, and different materials are mapped to unique classification values. Normalize the physical performance data and specification parameter data: ; In the formula: These are the normalized physical properties or specification parameters. The original measured values ​​of the features; , The minimum and maximum values ​​of this feature among all commercially available chemical and plastic products; Multi-label encoding is performed on application scenario data, mapping each product's corresponding application scenario to a binary label; The material composition after classification and coding, the normalized physical properties and specifications, and the application scenarios after multi-label coding are combined in a preset order to complete the construction of the feature vector of chemical plastic products.

[0010] Furthermore, when the fusion module associates and maps the user behavior analysis results with the feature vectors of chemical and plastic products, it converts the user behavior scores in the user behavior analysis results into user behavior vectors. The vector dimension is set to be consistent with the number of chemical and plastic products on sale, and the vector elements are the user's behavior scores for the corresponding products. Then, a product feature matrix is ​​constructed, with a dimension of m×n, where m represents the number of chemical and plastic products on sale and n represents the total dimension of the chemical and plastic product features. Each row of the matrix corresponds to the feature vector of one product, which is consistent with the feature vectors of chemical and plastic products constructed by the extraction module. A feature space alignment weight matrix with a dimension of n×n is further introduced. The elements of the weight matrix are dynamically optimized based on the historical conversion value of each feature dimension, using the formula... , user behavior vector Mapping to the product feature space yields the user feature preference vector. Its dimension is 1×n, where These represent the product feature matrix and the feature space alignment weight matrix, respectively. Calculate the correlation between user feature preference vectors and feature vectors of each product. ,in, Represents the product feature vector transpose, These represent the magnitude of the user feature preference vector and the magnitude of the product feature vector, respectively. Construct a user-product hybrid feature matrix with user ID as the row, product ID as the column, and matrix elements representing the correlation between the corresponding user and product.

[0011] Furthermore, the dynamic matching logic between user preferences and product features is as follows: Based on the user-product hybrid feature matrix, the preference coefficients of each user for each feature dimension of chemical and plastic products are obtained. The preference coefficients are the average correlation of the user corresponding to the current calculated preference coefficient with all products in the target dimension of each feature dimension of chemical and plastic products. When new user behavior data is generated, the behavior vector of the user to which the new behavior data belongs is updated in real time, and the user's preference coefficient is recalculated based on the updated behavior vector to complete dynamic iteration; For each user, the current user's preference coefficient is multiplied by the feature value of the corresponding dimension in each product feature vector, and the sum is used to obtain the product matching score. Then, the products are sorted from high to low according to the matching score to generate the initial recommendation sequence, which is the output result of the chemical and plastic product recommendation model. Among them, the feature dimensions are the material composition dimension, physical performance dimension, application scenario dimension, and specification parameter dimension.

[0012] Furthermore, when the correction module performs dynamic correction based on the inventory fluctuations of chemical and plastic products, it obtains the inventory quantity of each chemical and plastic product in real time and calculates the ratio of the current inventory quantity to the preset safety stock quantity, which is recorded as the inventory adequacy. Set an inventory impact factor. When the inventory adequacy is greater than a preset threshold, the inventory impact factor is 1. When the inventory adequacy is less than or equal to the preset threshold and greater than 0, the inventory impact factor decreases linearly as the inventory adequacy decreases. When the inventory adequacy is 0, the inventory impact factor is 0. The matching score of each product in the initial recommendation sequence is multiplied by the corresponding inventory influence factor to obtain the corrected matching score, which is then reordered to update the initial recommendation results.

[0013] Furthermore, when the correction module dynamically corrects user behavior based on real-time changes, it uses a preset time window (not exceeding 2 hours) as a unit to count the frequency of real-time user behavior within that window and compares it with the average frequency of user behavior in the same historical time window to calculate the behavior change rate. , These represent the frequency of user u's behavior in the current time window and the average frequency of user u's behavior in the same historical time windows, respectively. Simultaneously, calculate the distribution drift of the user's current behavior in the product feature space: Extract the set of feature vectors of the interactive products within the user's current time window, denoted as Calculate the eigencenter vector of this set. ,in, This indicates the number of products the user is currently interacting with within the window. Let be the feature vector of the i-th interactive product; Extract the feature vector set of products interacted with by users within the same historical time window. Calculate the eigencenter vector of this set. ,in, The average number of products interacted with by users within the historical window. Let be the feature vector of the j-th historical interactive product; Then the distribution drift , This represents the transpose of the historical feature center vector. This represents the magnitude of the current feature center vector relative to the historical feature center vector; When the absolute value of the rate of change in behavior is greater than a preset change threshold, or When the drift rate exceeds a preset threshold, a significant change in user behavior is determined. The user's behavior vector and preference coefficients are then recalculated, and the matching scores of each product in the initial recommendation sequence are adjusted based on the newly calculated preference coefficients. When the absolute value of the behavior change rate is less than or equal to a preset change threshold, and... When the frequency is less than or equal to the preset drift threshold, the initial recommended sequence remains unchanged, and only the frequency of the behavior is updated cumulatively.

[0014] Furthermore, when the correction module dynamically corrects based on product correlation, it calculates the product correlation between any two chemical and plastic products. , , , These represent the weighting coefficients for the material association score, the scene association score, and the performance association score, respectively. , , The values ​​of all are within the range (0,1), and , , The sum is 1, and < < , Indicates the material correlation score. Indicates the scene association score. Indicates the performance correlation score; In the initial recommendation sequence, if a product's matching score is higher than that of its associated products, but the associated products have higher inventory sufficiency or higher historical conversion rate of users towards the associated products, then the matching score of the associated products will be increased to within the preset ratio range of the product's matching score. Based on the adjusted matching score, the recommendation sequence will be sorted again to complete the correction of the product relevance dimension.

[0015] Furthermore, the preset formats in the output module include an initial recommended sequence list of chemical and plastic products and a list of links to detailed information about chemical and plastic products; The output module has built-in push logic: Set push priority: for users whose historical transaction frequency is higher than the preset frequency, the push frequency is once every x days; for users whose historical transaction frequency is less than or equal to the preset frequency but higher than 0, the push frequency is once every y days; for newly registered users who have not made any transactions, the push frequency is once every z days, and x < y < z. Add a brief description of the product matching score to the push content. The brief description is manually edited by the system user and includes: This product has a high degree of matching with your material preferences; The document contains a link that redirects to the product details page on the sales platform.

[0016] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects: This invention provides a hybrid recommendation system for chemical and plastic products that combines user behavior and product content. During operation, the system accurately captures users' browsing, clicking, collection, and transaction behaviors of chemical and plastic products and converts them into behavior scores. At the same time, it integrates product material composition, physical properties, application scenarios, and specification parameters to construct feature vectors. Through correlation mapping, it forms a user-product hybrid feature base that fits the needs. Based on dynamic matching logic, it updates the user preference coefficient in real time and generates an accurate initial recommendation sequence. At the same time, it can adjust the recommendation priority based on inventory fluctuations to avoid recommending products out of stock, iterate preferences in real time based on changes in user behavior to meet dynamic needs, optimize the sequence based on product relevance to improve rationality, and push recommendations in a tiered manner according to user transaction frequency, with additional matching instructions and detailed links. This not only improves the recommendation conversion rate but also enhances the user experience, helping the chemical and plastic product sales platform improve operational efficiency and user stickiness. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0018] Figure 1 This is a schematic diagram of a hybrid recommendation system for chemical and plastic products that combines user behavior and product content. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] The present invention will be further described below with reference to embodiments.

[0021] Example: This embodiment presents a hybrid recommendation system for chemical and plastic products that combines user behavior and product content, such as... Figure 1 As shown, it includes: The data collection and analysis module is used to connect to the sales platform for chemical and plastic products, collect data on users' browsing, clicking, collection, and transaction behavior of chemical and plastic products on the sales platform, and perform structured analysis on the data. The operation process of the data collection and parsing module when performing structured parsing of user behavior data is as follows: Dynamic behavior weighting coefficients are assigned to four user behaviors: browsing, clicking, favoriting, and trading. ; In the formula: Dynamic weights for transaction behavior, collection behavior, click behavior, and browsing behavior; The conversion coefficients are for transaction behavior, collection behavior, click behavior, and browsing behavior. The current system time; The time of the most recent action i; The time of the most recent transaction, favorites, click, or browsing activity; Quantify the historical transformation value of behavior i; This is the time decay coefficient; It should be noted that: based on the fact that the transaction itself has already been converted, let's assume... =1, The ratio of the number of transactions that occurred within a preset period after a user added a item to their favorites to the total number of times the user added the item to their favorites. The ratio of the number of transactions that occurred within a preset period after a user clicked to the total number of user clicks. The ratio of the number of transactions that occur within a preset period after a user browses the page to the total number of page views. ∈[0.01, 0.05], set by the system based on the average period of platform user behavior; the shorter the period, the better. The larger, the smaller, and vice versa. The smaller; The above formula introduces a time decay factor. This addresses the issue of inconsistent weighting between historical and recent behaviors. For example, browsing behavior from one month ago carries less weight than browsing behavior from one day ago. This approach better aligns with the characteristics of chemical and plastic product procurement decisions, where the decision-making cycle is long but recent behavior better reflects current needs. Furthermore, it incorporates behavioral conversion coefficients. This makes the weight strongly correlated with the "commercial value" of the behavior. For example, if the conversion coefficient of a user's collection behavior is higher than the industry average, the user's collection weight will be adaptively increased, which is more accurate than a fixed weight. Using a preset time window as a unit, the frequency of each user's four behaviors towards each chemical and plastic product within that time window is counted, and the behavior frequency is multiplied by the corresponding weight coefficient. The product result is recorded as the user's behavior score for each product. A two-dimensional matrix is ​​constructed based on user ID and product ID, with the matrix elements representing the user's behavior score for the corresponding product, in order to form a structured user behavior analysis result. Among them, the weight coefficient of transaction behavior is greater than that of collection behavior, the weight coefficient of collection behavior is greater than that of click behavior, the weight coefficient of click behavior is greater than that of browsing behavior, and the sum of the four weights is 1. The conversion rate of users' historical behavior is the core basis for determining the weight coefficients of four types of user behavior: browsing, clicking, collecting, and trading. The conversion rate is defined as the ratio of the number of times a transaction is triggered within a preset time interval after the occurrence of the behavior to the total number of times the behavior occurs. The transaction behavior weight coefficient is the maximum value among all user behavior conversion rates, and is not less than 0.4; the collection behavior weight coefficient is the product of the transaction behavior weight coefficient and the collection-transaction conversion rate; the click behavior weight coefficient is the product of the collection behavior weight coefficient and the click-collection conversion rate; the browsing behavior weight coefficient is the product of the click behavior weight coefficient and the browsing-click conversion rate. The extraction module is used to extract the material composition, physical properties, application scenarios and specification parameters of chemical and plastic products sold on the sales platform in order to construct the feature vector of chemical and plastic products. During the extraction module's runtime phase, when constructing the feature vector of chemical and plastic products, the following rules apply: The material composition data is classified and coded, and different materials are mapped to unique classification values. Normalize the physical performance data and specification parameter data: ; In the formula: These are the normalized physical properties or specification parameters. The original measured values ​​of the features; , The minimum and maximum values ​​of this feature among all commercially available chemical and plastic products; Multi-label encoding is performed on application scenario data, mapping each product's corresponding application scenario to a binary label; The material composition after classification and coding, the normalized physical properties and specifications, and the application scenarios after multi-label coding are combined in a preset order to complete the construction of the feature vector of chemical plastic products. The fusion module is used to receive the user behavior analysis results from the acquisition and analysis module and the chemical and plastic product feature vectors from the extraction module, perform correlation mapping between the two, and construct a user-product hybrid feature matrix. When the fusion module maps user behavior analysis results to feature vectors of chemical and plastic products, it transforms user behavior scores in the analysis results into user behavior vectors. The vector dimension is set to match the number of chemical and plastic products on sale, and the vector elements represent the user's behavior scores for the corresponding products. A product feature matrix is ​​then constructed, with dimensions denoted as m×n, where m represents the number of chemical and plastic products on sale and n represents the total dimension of the chemical and plastic product features. Each row of the matrix corresponds to the feature vector of one product, which is consistent with the feature vectors of chemical and plastic products constructed by the extraction module. A feature space alignment weight matrix with dimensions n×n is further introduced. The elements of this weight matrix are dynamically optimized based on the historical conversion value of each feature dimension, using the formula... , user behavior vector Mapping to the product feature space yields the user feature preference vector. Its dimension is 1×n, where These represent the product feature matrix and the feature space alignment weight matrix, respectively. Calculate the correlation between user feature preference vectors and feature vectors of each product. ,in, Represents the product feature vector transpose, These represent the magnitude of the user feature preference vector and the magnitude of the product feature vector, respectively. Construct a user-product hybrid feature matrix with user ID as the row, product ID as the column, and matrix elements representing the correlation between the corresponding user and product; Among them, conversion value is represented by the contribution of this feature dimension to user transaction behavior. The total feature dimension of chemical and plastic products is the sum of the dimensions of material composition, physical properties, application scenarios, and specification parameters. The above management formula transforms user behavior scores into user behavior vectors and product features into product feature vectors. By calculating the correlation between the two, the matching degree between users and products is measured. This calculation logic transforms user behavior and product features into quantifiable and comparable vector forms. With the help of vector correlation calculation, it can comprehensively consider users' behavior towards all products and the features of each dimension of the products, and more comprehensively reflect the intrinsic connection between users and products, providing a key basis for constructing a user-product hybrid feature matrix. The generation module receives a user-product hybrid feature matrix and, based on the user-product hybrid feature matrix, generates a chemical and plastic product recommendation model by constructing a dynamic matching logic between user preferences and product features. The dynamic matching logic between user preferences and product features is as follows: Based on the user-product hybrid feature matrix, the preference coefficients of each user for each feature dimension of chemical and plastic products are obtained. The preference coefficients are the average correlation of the user corresponding to the current calculated preference coefficient with all products in the target dimension of each feature dimension of chemical and plastic products. When new user behavior data is generated, the behavior vector of the user to which the new behavior data belongs is updated in real time, and the user's preference coefficient is recalculated based on the updated behavior vector to complete dynamic iteration; For each user, the current user's preference coefficient is multiplied by the feature value of the corresponding dimension in each product feature vector, and the sum is used to obtain the product matching score. Then, the products are sorted from high to low according to the matching score to generate the initial recommendation sequence, which is the output result of the chemical and plastic product recommendation model. Among them, the feature dimensions are material composition dimension, physical performance dimension, application scenario dimension, and specification parameter dimension; The correction module is used to dynamically correct the initial recommendation results of chemical and plastic products output by the recommendation model based on inventory fluctuations of chemical and plastic products, real-time changes in user behavior, and product relevance. When the correction module dynamically corrects the inventory fluctuations of chemical and plastic products, it obtains the inventory quantity of each chemical and plastic product in real time and calculates the ratio of the current inventory quantity to the preset safety stock quantity, which is recorded as the inventory adequacy. An inventory impact factor is set. When inventory adequacy is greater than a preset threshold, the inventory impact factor is 1. When inventory adequacy is less than or equal to the preset threshold but greater than 0, the inventory impact factor decreases linearly as inventory adequacy decreases. The slope used in the linear decrease process is... When the inventory adequacy ratio is 0, the inventory impact factor is set to 0. The above formula uses the preset safety stock quantity as the denominator and combines the product priority coefficient and historical inventory turnover coefficient to calculate the slope, so that the inventory impact factor decreases linearly as the inventory adequacy decreases. This design takes into account both the importance of the product itself (such as profit contribution and scarcity) and the historical inventory turnover efficiency, so that inventory adjustment is no longer a simple "have or not have" judgment, but dynamically adjusts the recommendation priority according to the actual value of the product, balancing user needs and platform inventory management efficiency. The matching score of each product in the initial recommendation sequence is multiplied by the corresponding inventory influence factor to obtain the corrected matching score, which is then reordered to update the initial recommendation results; in, The preset safety stock quantity for this chemical and plastic product; , For preset weighting coefficients, , Both are positive numbers and their sum is 1, and their values ​​are user-defined on the system side. > ; This is the product priority coefficient; This refers to the historical inventory turnover coefficient. Based on the profit contribution rate and market scarcity of chemical and plastic products, priority levels are dynamically preset, and each priority level corresponds to a different coefficient value. The product's inventory turnover days are calculated based on the average inventory turnover days over the past three preset periods. The ratio of the product's historical average turnover days to the maximum historical average turnover days of all chemical and plastic products on sale is taken. When the correction module dynamically adjusts based on real-time changes in user behavior, it uses a preset time window (not exceeding 2 hours) as the unit. It counts the frequency of user behavior within this window and compares it with the average frequency of behavior in the same historical time window to calculate the rate of change in behavior. , These represent the frequency of user u's behavior in the current time window and the average frequency of user u's behavior in the same historical time windows, respectively. Simultaneously, calculate the distribution drift of the user's current behavior in the product feature space: Extract the set of feature vectors of the interactive products within the user's current time window, denoted as Calculate the eigencenter vector of this set. ,in, This indicates the number of products the user is currently interacting with within the window. Let be the feature vector of the i-th interactive product; Extract the feature vector set of products interacted with by users within the same historical time window. Calculate the eigencenter vector of this set. ,in, The average number of products interacted with by users within the historical window. Let be the feature vector of the j-th historical interactive product; Then the distribution drift , This represents the transpose of the historical feature center vector. This represents the magnitude of the current feature center vector relative to the historical feature center vector; When the absolute value of the rate of change in behavior is greater than a preset change threshold, or When the drift rate exceeds a preset threshold, a significant change in user behavior is determined. The user's behavior vector and preference coefficients are then recalculated, and the matching scores of each product in the initial recommendation sequence are adjusted based on the newly calculated preference coefficients. When the absolute value of the behavior change rate is less than or equal to a preset change threshold, and... When the frequency is less than or equal to the preset drift threshold, the initial recommendation sequence remains unchanged, and only the frequency of the behavior is cumulatively updated. When the correction module dynamically corrects based on product correlation, it calculates the product correlation between any two chemical and plastic products. , , , These represent the weighting coefficients for the material association score, the scene association score, and the performance association score, respectively. , , The values ​​of all are within the range (0,1), and , , The sum is 1, and < < , Indicates the material correlation score. Indicates the scene association score. Indicates the performance correlation score; in , , The specific values ​​of follow: ; In the formula: The relevance weights are dynamically adjusted. The system presets basic weights, and the system user defines the initial values. This represents the influence coefficient of historical behavior. This represents the user's historical preference index for dimension d. This is the dimensional interaction decay coefficient; The time interval between the current time and the user's most recent interaction with dimension d; The initial value range is set to [0.1, 0.3]. A larger value is used when the influence of user history on the weight is desired to be more significant, and a smaller value is used when the influence of history needs to be weakened. The initial value range is set to [0,1]. When the user presets a period, such as the percentage of interactions with dimension d in the past 90 days, the value is larger; when the percentage of interactions is lower, the value is smaller. The initial value range is set to [0.01, 0.05]. When it is desired that the decay rate of the current weight of the long-term interaction behavior is faster, the value is larger, and when it is necessary to retain the long-term interaction effect for a longer period of time, the value is smaller. The above formula quantifies the product correlation from core product attributes to application scenarios and basic materials in multiple dimensions. This allows the recommendation system to prioritize suitable products based on the actual correlation logic between products when adjusting the recommendation sequence, avoiding fragmented recommendation results. At the same time, it highlights the impact of key correlation dimensions on user decisions through weight allocation. In the initial recommendation sequence, if a product's matching score is higher than that of its associated products, but the associated products have higher inventory sufficiency or higher historical conversion rate of users towards the associated products, then the matching score of the associated products will be increased to the preset ratio range of the product's matching score. Based on the adjusted matching score, the recommendation sequence will be sorted again to complete the correction of the product relevance dimension. If the material composition of two products is exactly the same, then The value is the preset base score for the material, initially set to 1. If the material composition is different, then... =0; Take the ratio of the number of application scenarios jointly covered by the two products to the total number of application scenarios covered by each product individually. The ratio of the number of physical performance features of two products that are similar within a preset range to the total number of feature dimensions of the physical performance of the chemical and plastic products is taken. The improvement in the matching score of the associated products is the minimum value of the preset ratio range. The output module is used to convert the corrected initial recommendation results of chemical and plastic products into a document file according to a preset format and push it to the preset mobile device held by the user. The output module includes preset formats such as an initial recommended sequence list of chemical and plastic products and a list of links to detailed information about chemical and plastic products. The output module has built-in push logic: Set push priority: for users whose historical transaction frequency is higher than the preset frequency, the push frequency is once every x days; for users whose historical transaction frequency is less than or equal to the preset frequency but higher than 0, the push frequency is once every y days; for newly registered users who have not made any transactions, the push frequency is once every z days, and x < y < z. Add a brief description of the product matching score to the push content. The brief description is manually edited by the system user and includes: This product has a high degree of matching with your material preferences; The document contains a link that redirects to the product details page on the sales platform; The acquisition and analysis module is interconnected with the extraction module via a wireless network. The extraction module is interconnected with the fusion module and the generation module via a wireless network. The fusion module and the generation module are interconnected with the correction module via a wireless network. The correction module is interconnected with the output module via a wireless network.

[0022] In this embodiment, the data acquisition and analysis module connects to the chemical and plastic product sales platform, collects data on users' browsing, clicking, collection, and transaction behavior of chemical and plastic products on the platform, and performs structured analysis on the data. The extraction module then extracts the material composition, physical properties, application scenarios, and specification parameters of the chemical and plastic products sold on the sales platform to construct chemical and plastic product feature vectors. The fusion module further receives the user behavior analysis results from the data acquisition and analysis module and the chemical and plastic product feature vectors from the extraction module, performs correlation mapping between the two, and constructs a user-product hybrid feature matrix. The generation module then receives the user-product hybrid feature matrix and, based on the user-product hybrid feature matrix, generates a chemical and plastic product recommendation model by constructing a dynamic matching logic between user preferences and product features. The correction module dynamically corrects the initial chemical and plastic product recommendation results output by the recommendation model based on inventory fluctuations, real-time changes in user behavior, and product relevance. Finally, the output module converts the corrected initial chemical and plastic product recommendation results into a document file according to a preset format and pushes it to the preset mobile device terminal held by the user.

[0023] In the above embodiments, in the sales scenario of chemical and plastic products, this solution can dynamically match user preferences to generate recommendations by combining user browsing and transaction behaviors with information such as product materials, performance, and application scenarios. It can also adjust the results based on inventory fluctuations, real-time changes in user behavior, and product relevance. The push format includes a recommendation list and a details link, with a frequency adapted to users with different transaction frequencies, and includes matching instructions. This not only helps users quickly find suitable products but also helps the platform optimize inventory turnover, effectively improving recommendation accuracy and user experience.

[0024] The following is an application example simulating the system described in the above embodiments: Taking the "XX Chemicals & Plastics E-commerce Platform" and platform user A (registered for 1 year, with 3 transaction records) as an example: 1. Data Acquisition and Analysis After the data collection and analysis module connects to the platform, it first determines the weight coefficients for four behaviors—browsing, clicking, favoriting, and trading—based on the user's historical behavior conversion rate: 0.1, 0.2, 0.25, and 0.45, respectively (the sum of the four is 1, with trading having the highest weight). Using a preset 7-day time window, user A's behavior within this window is statistically analyzed: PP food packaging plastic (product P1) was browsed 5 times, clicked 3 times, favorited 1 time, and traded no times; PE pipes (product P2) were browsed 2 times, clicked 4 times, and traded no times. User A's behavior scores for the two products are calculated: P1 score is 1.35, and P2 score is 1.0. Finally, a two-dimensional matrix is ​​constructed with user A's ID as the row and the IDs of products sold on the platform (P1 / P2, etc.) as columns. The matrix elements are filled with the corresponding product behavior scores, forming a structured analysis result.

[0025] 2. Product Feature Vector Construction The extraction module extracts data from products sold on the platform and constructs feature vectors. Taking product P1 as an example, its "PP material" is first coded as 101; in terms of physical properties, "melting point 160℃" is normalized to 0.67, since the minimum melting point of all products sold on the platform is 120℃ and the maximum is 180℃; the application scenario "food packaging, toy manufacturing" is coded with multiple tags according to the platform's four core scenarios (food packaging, toy manufacturing, building pipes, and electronic accessories), resulting in 1100; the specification parameter "thickness 2mm" is normalized to 0.5. The above classification coding, normalization, and multi-tag coding results are combined in the order of "material-physical properties-application scenario-specific parameters" to complete the construction of P1's feature vector; other products such as product P2 are also constructed using the same logic.

[0026] 3. Construction of Hybrid Feature Matrix The fusion module first converts user A's behavior score into a behavior vector. The vector dimension matches the number of 100 products sold on the platform, and the vector elements represent user A's behavior score for each product (e.g., 1.35 for P1 and 1.0 for P2). Next, the correlation between this behavior vector and the feature vectors of each product is calculated. The correlation between user A and P1 is 0.82, and the correlation with P2 is 0.75. Finally, a user-product hybrid feature matrix is ​​constructed with user A's ID as the row and all product IDs as the column. The matrix elements are filled with the correlation between the corresponding user and product (e.g., 0.82 at the intersection of user A and P1).

[0027] 4. Recommendation Model Generation The generation module, based on a hybrid feature matrix, first calculates the preference coefficients for user A across various feature dimensions: material dimension 0.75 (the average of user A's material associations across all products), physical performance dimension 0.8, application scenario dimension 0.78, and specification parameter dimension 0.72. The next day, user A clicks on product P3 (PVC material board). The system updates its behavior vector in real time and recalculates the preference coefficients, adjusting the material dimension coefficient to 0.76. Subsequently, user A's preference coefficients across all dimensions are multiplied by the corresponding dimension values ​​in each product's feature vector and summed to obtain product matching scores: P1 score 2.241, P2 score 2.31, and P3 score 2.18. Sorted from highest to lowest score, an initial recommendation sequence is generated: P2 (2.31), P5 (2.28, other products), P1 (2.241), P3 (2.18)... 5. Recommendation results are dynamically adjusted. The correction module corrects the initial sequence in three ways: Inventory Adjustment: Obtain the inventory of each product. P1 has a preset safety stock of 500 units and a current inventory of 300 units, with an inventory adequacy of 0.6 (higher than the preset threshold of 0.5). The inventory impact factor is set to 1, and the matching score of P1 remains at 2.241. Product P5 currently has an inventory of 200 units (safety stock of 500 units, adequacy of 0.4). The impact factor is set to 0.8, and the original score of 2.28 is adjusted to 1.82, causing the ranking to drop to 6th place.

[0028] Behavior change correction: With a preset time window of 1 hour, the system counts that user A clicks 3 times in the current window and the average frequency of the same window in the past is 1 time, and the behavior change rate is 200% (higher than the preset threshold of 150%). The system recalculates its preference coefficient (the physical performance dimension is adjusted to 0.81) and updates the P1 matching score to 2.26 and the P2 matching score to 2.33.

[0029] Product Relevance Correction: The correlation between P1 and P4 (another PP food packaging plastic) is calculated to be 0.599, and P4's inventory sufficiency is 0.9 (higher than P1's 0.6). The system increases P4's matching score from 2.0 to 80% of P1's current score (i.e., 1.808), and after correction, P4's ranking rises to 5th place. The final corrected recommended sequence is: P2 (2.33), P1 (2.26), P3 (2.18), P6 (2.12, other products), P4 (1.808)... 6. Recommendation Results Output The output module converts the revised recommendation sequence into a document file containing a list of recommended products (arranged in the revised sequence) and links to platform details for each product. Since User A's historical transaction frequency is 5 times per month (higher than the platform's preset threshold of 3 times per month), the push notification frequency is set to once every 3 days according to the push logic. A brief description is added to the push content: "Product P2 (PE material pipe) has a high match with your physical performance preferences," and all product names in the document include links to the platform details page. The document is ultimately pushed to User A's mobile device.

[0030] In summary, during operation, the system in the above embodiments accurately captures users' browsing, clicking, collection, and transaction behaviors related to chemical and plastic products and converts them into behavioral scores. Simultaneously, it integrates product material composition, physical properties, application scenarios, and specifications to construct feature vectors. Through correlation mapping, it forms a user-product hybrid feature base that aligns with user needs. Relying on dynamic matching logic, it updates user preference coefficients in real time, generating an accurate initial recommendation sequence. Furthermore, it adjusts recommendation priorities based on inventory fluctuations to avoid recommending products out of stock. It iterates preferences in conjunction with real-time changes in user behavior to dynamically match demand, optimizes the sequence based on product relevance to improve its rationality, and then pushes recommendations in a tiered manner according to user transaction frequency, adding matching instructions and detailed links. This not only improves recommendation conversion rates but also enhances user experience, assisting chemical and plastic product sales platforms in improving operational efficiency and user stickiness.

[0031] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hybrid recommendation system for chemical and plastic products that combines user behavior and product content, characterized in that, include: The data collection and analysis module is used to connect to the sales platform for chemical and plastic products, collect data on users' browsing, clicking, collection, and transaction behavior of chemical and plastic products on the sales platform, and perform structured analysis on the data. The extraction module is used to extract the material composition, physical properties, application scenarios and specification parameters of chemical and plastic products sold on the sales platform in order to construct the feature vector of chemical and plastic products. The fusion module is used to receive the user behavior analysis results from the acquisition and analysis module and the chemical and plastic product feature vectors from the extraction module, perform correlation mapping between the two, and construct a user-product hybrid feature matrix. The generation module receives a user-product hybrid feature matrix and, based on the user-product hybrid feature matrix, generates a chemical and plastic product recommendation model by constructing a dynamic matching logic between user preferences and product features. The correction module is used to dynamically correct the initial recommendation results of chemical and plastic products output by the recommendation model based on inventory fluctuations of chemical and plastic products, real-time changes in user behavior, and product relevance. The output module is used to convert the corrected initial recommendation results of chemical and plastic products into a document file according to a preset format and push it to the preset mobile device held by the user.

2. The hybrid recommendation system for chemical and plastic products combining user behavior and product content according to claim 1, characterized in that, The operation process of the acquisition and parsing module when performing structured parsing of user behavior data is as follows: Dynamic behavior weighting coefficients are assigned to four user behaviors: browsing, clicking, favoriting, and trading. Using a preset time window as a unit, the frequency of each user's four behaviors towards each chemical and plastic product within that time window is counted, and the behavior frequency is multiplied by the corresponding weight coefficient. The product result is recorded as the user's behavior score for each product. A two-dimensional matrix is ​​constructed based on user ID and product ID, with the matrix elements representing the user's behavior score for the corresponding product, in order to form a structured user behavior analysis result. Among them, the weight coefficient of transaction behavior is greater than that of collection behavior, the weight coefficient of collection behavior is greater than that of click behavior, the weight coefficient of click behavior is greater than that of browsing behavior, and the sum of the four weights is 1.

3. The hybrid recommendation system for chemical and plastic products combining user behavior and product content according to claim 1, characterized in that, During the operation of the extraction module, when constructing the feature vector of chemical and plastic products, the following applies: The material composition data is classified and coded, and different materials are mapped to unique classification values. Normalize the physical performance data and specification parameter data: ； In the formula: These are the normalized physical properties or specification parameters. The original measured values ​​of the features; , The minimum and maximum values ​​of this feature among all commercially available chemical and plastic products; Multi-label encoding is performed on application scenario data, mapping each product's corresponding application scenario to a binary label; The material composition after classification and coding, the normalized physical properties and specifications, and the application scenarios after multi-label coding are combined in a preset order to complete the construction of the feature vector of chemical plastic products.

4. The hybrid recommendation system for chemical and plastic products combining user behavior and product content according to claim 1, characterized in that, When the fusion module associates and maps user behavior analysis results with chemical and plastic product feature vectors, it converts user behavior scores in the analysis results into user behavior vectors. The vector dimension is set to match the number of chemical and plastic products on sale, and the vector elements are the user's behavior scores for the corresponding products. A product feature matrix is ​​then constructed, with dimensions denoted as m×n, where m represents the number of chemical and plastic products on sale and n represents the total dimension of the chemical and plastic product features. Each row of the matrix corresponds to the feature vector of one product, which is consistent with the feature vectors of chemical and plastic products constructed by the extraction module. A feature space alignment weight matrix with dimensions n×n is further introduced. The elements of this weight matrix are dynamically optimized based on the historical conversion value of each feature dimension, using the formula... , user behavior vector Mapping to the product feature space yields the user feature preference vector. Its dimension is 1×n, where These represent the product feature matrix and the feature space alignment weight matrix, respectively. Calculate the correlation between user feature preference vectors and feature vectors of each product. ,in, Represents the product feature vector transpose, These represent the magnitude of the user feature preference vector and the magnitude of the product feature vector, respectively. Construct a user-product hybrid feature matrix with user ID as the row, product ID as the column, and matrix elements representing the correlation between the corresponding user and product.

5. A hybrid recommendation system for chemical and plastic products combining user behavior and product content according to claim 1, characterized in that, The dynamic matching logic between user preferences and product features is as follows: Based on the user-product hybrid feature matrix, the preference coefficients of each user for each feature dimension of chemical and plastic products are obtained. The preference coefficients are the average correlation of the user corresponding to the current calculated preference coefficient with all products in the target dimension of each feature dimension of chemical and plastic products. When new user behavior data is generated, the behavior vector of the user to which the new behavior data belongs is updated in real time, and the user's preference coefficient is recalculated based on the updated behavior vector to complete dynamic iteration; For each user, the current user's preference coefficient is multiplied by the feature value of the corresponding dimension in each product feature vector, and the sum is used to obtain the product matching score. Then, the products are sorted from high to low according to the matching score to generate the initial recommendation sequence, which is the output result of the chemical and plastic product recommendation model. Among them, the feature dimensions are the material composition dimension, physical performance dimension, application scenario dimension, and specification parameter dimension.

6. A hybrid recommendation system for chemical and plastic products combining user behavior and product content according to claim 1, characterized in that, When the correction module performs dynamic correction based on the inventory fluctuations of chemical and plastic products, it obtains the inventory quantity of each chemical and plastic product in real time and calculates the ratio of the current inventory quantity to the preset safety stock quantity, which is recorded as the inventory adequacy. Set an inventory impact factor; when the inventory adequacy is greater than a preset threshold, the inventory impact factor is set to 1. When the inventory adequacy is less than or equal to the preset threshold and greater than 0, the inventory impact factor decreases linearly as the inventory adequacy decreases; when the inventory adequacy is 0, the inventory impact factor is 0. The matching score of each product in the initial recommendation sequence is multiplied by the corresponding inventory influence factor to obtain the corrected matching score, which is then reordered to update the initial recommendation results.

7. A hybrid recommendation system for chemical and plastic products combining user behavior and product content according to claim 6, characterized in that, When the correction module dynamically corrects based on real-time changes in user behavior, it uses a preset time window (not exceeding 2 hours) as a unit to count the frequency of real-time user behavior within that window and compares it with the average frequency of behavior in the same historical time window to calculate the behavior change rate. , These represent the frequency of user u's behavior in the current time window and the average frequency of user u's behavior in the same historical time windows, respectively. Simultaneously, calculate the distribution drift of the user's current behavior in the product feature space: Extract the set of feature vectors of the interactive products within the user's current time window, denoted as Calculate the eigencenter vector of this set. ,in, This indicates the number of products the user is currently interacting with within the window. Let be the feature vector of the i-th interactive product; Extract the feature vector set of products interacted with by users within the same historical time window. Calculate the eigencenter vector of this set. ,in, The average number of products interacted with by users within the historical window. Let be the feature vector of the j-th historical interactive product; Then the distribution drift , This represents the transpose of the historical feature center vector. This represents the magnitude of the current feature center vector relative to the historical feature center vector; When the absolute value of the rate of change in behavior is greater than a preset change threshold, or When the drift rate exceeds a preset threshold, a significant change in user behavior is determined. The user's behavior vector and preference coefficients are then recalculated, and the matching scores of each product in the initial recommendation sequence are adjusted based on the newly calculated preference coefficients. When the absolute value of the behavior change rate is less than or equal to a preset change threshold, and... When the frequency is less than or equal to the preset drift threshold, the initial recommended sequence remains unchanged, and only the frequency of the behavior is updated cumulatively.

8. A hybrid recommendation system for chemical and plastic products combining user behavior and product content according to claim 6, characterized in that, When the correction module performs dynamic correction based on product correlation, it calculates the product correlation between any two chemical and plastic products. , , , These represent the weighting coefficients for the material association score, the scene association score, and the performance association score, respectively. , , The values ​​of all are within the range (0,1), and , , The sum is 1, and < < , Indicates the material correlation score. Indicates the scene association score. Indicates the performance correlation score; In the initial recommendation sequence, if a product's matching score is higher than that of its associated products, but the associated products have higher inventory sufficiency or higher historical conversion rate of users towards the associated products, then the matching score of the associated products will be increased to within the preset ratio range of the product's matching score. Based on the adjusted matching score, the recommendation sequence will be sorted again to complete the correction of the product relevance dimension.

9. A hybrid recommendation system for chemical and plastic products combining user behavior and product content according to claim 1, characterized in that, The output module includes preset formats such as an initial recommended sequence list of chemical and plastic products and a list of links to detailed chemical and plastic products. The output module has built-in push logic: Set push priority: for users whose historical transaction frequency is higher than the preset frequency, the push frequency is once every x days; for users whose historical transaction frequency is less than or equal to the preset frequency but higher than 0, the push frequency is once every y days; for newly registered users who have not made any transactions, the push frequency is once every z days, and x < y < z. Add a brief description of the product matching score to the push content. The brief description is manually edited by the system user and includes: This product has a high degree of matching with your material preferences; The document contains a link that redirects to the product details page on the sales platform.

10. A hybrid recommendation system for chemical and plastic products combining user behavior and product content according to claim 1, characterized in that, The acquisition and analysis module is interconnected with an extraction module via a wireless network. The extraction module is interconnected with a fusion module and a generation module via a wireless network. The fusion module and the generation module are interconnected with a correction module via a wireless network. The correction module is interconnected with an output module via a wireless network.

Images