Method and system for realizing intelligent screening of multi-level commodities based on AI large model

Abstract

The application relates to the field of artificial intelligence technology and discloses a method and system for realizing intelligent screening of multilevel commodities based on an AI large model, which comprises the following steps: collecting a multilevel commodity query instruction of a user, a commodity screening dimension of the query instruction, and constructing a commodity query graph of the user; performing node query degree assignment on the commodity query graph to obtain an assigned graph, constructing a commodity screening candidate pool, performing primary screening of commodities on the commodity screening candidate pool to obtain primary screening commodities; when the primary screening commodities do not meet the commodity demand of the user, querying an interaction behavior to construct a commodity preference structure tree of the user, performing structure reconstruction on the commodity preference structure tree to obtain a reconstructed structure tree; and when the reconstructed screening commodities meet the commodity demand of the user, taking the reconstructed screening commodities as final screening commodities. The application can improve the matching degree of screening commodity results and real demands of users.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent screening of multi-level goods based on a large AI model, belonging to the field of artificial intelligence technology. Background Technology

[0002] In the current booming development of e-commerce and retail industries, intelligent decision-making to achieve efficient and accurate product selection in various scenarios such as e-commerce platforms, social e-commerce, and cross-border shopping has become a key technology for improving user conversion rates and shopping experience, and is also crucial for the sales volume of product sales platforms.

[0003] Currently, multi-level product filtering often combines rule engines with collaborative filtering algorithms. This method typically first constructs a product tagging system, then combines it with user profile data, and finally uses predefined filtering conditions to categorize and display products. For example, in e-commerce platforms, interest models are built based on user click behavior to achieve "you might like" product recommendations. However, this traditional recommendation method struggles to adapt to rapidly changing consumer trends and ignores user preferences, often resulting in many products that do not meet user needs during the filtering process. Summary of the Invention

[0004] This invention provides a method and system for intelligent filtering of multi-level products based on a large AI model. Its main purpose is to improve the match between the filtered product results and the user's actual needs.

[0005] To achieve the above objectives, the present invention provides a method for intelligent selection of multi-level products based on a large AI model, comprising:

[0006] Collect users' multi-level product query commands, use a trained AI model to identify the product filtering dimensions of the multi-level product query commands, and use the product filtering dimensions to construct the user's product query graph;

[0007] The system queries the user's historical filtering information, assigns node query degree values ​​to the product query graph based on the historical filtering information to obtain an assigned graph, constructs a product filtering candidate pool corresponding to the user using the assigned graph, calculates the convergence value of the product filtering candidate pool, and performs primary product filtering on the product filtering candidate pool based on the convergence value to obtain primary filtered products.

[0008] When the products selected in the initial screening meet the user's product needs, the products selected in the initial screening will be used as the final products selected.

[0009] When the primary filtered products do not meet the user's product needs, the user's interaction behavior with the primary filtered products is queried. Based on the interaction behavior, the user's product preference structure tree is constructed. Cross-dimensional correlation feature mining is performed on the product preference structure tree to obtain multi-level correlation features. The product preference structure tree is restructured using the multi-level correlation features to obtain the restructured structure tree.

[0010] Using the reconstructed structure tree, the reconstructed filtered products corresponding to the user are analyzed. When the reconstructed filtered products do not meet the user's product needs, the user's interaction behavior with the primary filtered products is returned until the reconstructed filtered products meet the user's product needs, at which point the reconstructed filtered products are used as the final filtered products.

[0011] Optionally, the user's product query graph is constructed using the product filtering dimensions, including:

[0012] Using the product filtering dimensions, construct structured query conditions for multi-level products corresponding to the user's query;

[0013] The structured query conditions are then matched using a knowledge graph to obtain an extended dimension set.

[0014] The extended dimension set is initialized with graph nodes to obtain the basic graph nodes;

[0015] Static relationship modeling is performed on the basic graph nodes to obtain the product query graph.

[0016] Optionally, based on the historical filtering information, node query degrees are assigned to the product query graph to obtain an assigned graph, including:

[0017] The historical filtering information is subjected to user behavior standardization processing to obtain standardized behavior data;

[0018] Using the standardized behavioral data, construct the user's dimensional utility matrix;

[0019] The time decay weighting of the aforementioned dimensional utility matrix is ​​applied to obtain the time-effect weight table;

[0020] Using the aforementioned timeliness weight table, node weight mapping is performed on the product query graph to obtain the assigned graph.

[0021] Optionally, constructing the user's dimensional utility matrix using the standardized behavioral data includes:

[0022] Using the standardized behavioral data, calculate the node exposure weight and node conversion rate weight of the product query graph;

[0023] Based on the node exposure weight and the node conversion rate weight, a comprehensive weight assignment is performed on the product query graph to obtain a comprehensive weight assignment graph.

[0024] Based on the comprehensive weight assignment graph, the user's dimensional utility matrix is ​​constructed.

[0025] Optionally, the step of using the standardized behavioral data to calculate the node exposure weight and node conversion rate weight of the product query graph includes:

[0026] Identify the behavioral nodes and behavioral types in the standardized behavioral data that correspond to the product query graph;

[0027] Based on the behavioral nodes, the node exposure weight of the product query graph is calculated using the following formula:

[0028]

[0029] Among them, A i E represents the exposure percentage of behavior node i. i E represents the historical exposure count of the i-th behavior node, e represents the smoothing factor, and n represents the total number of behavior nodes at the same level in the product query graph. j Let α represent the historical exposure count of the j-th behavior node, and let α represent the node exposure weight. min(A i ) represents A i Maximum value, max(A) i ) represents A i The minimum value of ∈ represents the denominator protection parameter;

[0030] Based on the behavior type, the node conversion rate weight of the product query graph is calculated using the following formula:

[0031]

[0032] Among them, B k Let m represent the node conversion rate for the k-th behavior, and w represent the total number of behavior types. k C represents the weight of the k-th action. k Let μ represent the total number of times the k-th behavior occurs, β represent the node conversion rate weight, and μ represent the total number of times the behavior occurs. k σ represents the average conversion rate at the same level as k, σ represents the standard deviation, and ∈ represents the denominator protection parameter.

[0033] Optionally, calculating the convergence value of the product screening candidate pool includes:

[0034] The category concentration of the product screening candidate pool is calculated using the following formula;

[0035] Calculate the attribute similarity of products in the product screening candidate pool;

[0036] Based on the category concentration and the attribute similarity, the convergence value of the product screening candidate pool is calculated.

[0037] Optionally, based on the interaction behavior, a product preference structure tree for the user is constructed, including:

[0038] Extract the real-time features of the interaction behavior to obtain a behavior feature set;

[0039] Using the aforementioned behavioral characteristics, construct the user's product preference nodes;

[0040] Using the product preference nodes, construct the user's product preference tree architecture;

[0041] Incrementally update the product preference tree structure to obtain the product preference tree structure.

[0042] Optionally, cross-dimensional correlation feature mining is performed on the product preference structure tree to obtain multi-level correlation features, including:

[0043] The product preference structure tree is transformed into a graph structure to obtain a product dimension association graph;

[0044] Frequent subgraph mining is performed on the product dimension association graph to obtain a set of high-frequency association patterns;

[0045] Path calculation is performed on the high-frequency association pattern set to obtain weighted association paths;

[0046] Graph neural network learning is performed on the weighted association paths to identify potential association features in the product preference structure tree;

[0047] Similarity clustering is performed on potential association features to obtain multi-level association features.

[0048] Optionally, using the reconstructed structure tree, analyze the reconstructed filtered products corresponding to the user, including:

[0049] The reconstructed structure tree is encoded using conditional vectors to obtain the filtering condition vectors;

[0050] Using the filtering condition vector, construct a candidate product score table corresponding to the user;

[0051] The candidate product score table is sorted in a multi-dimensional mixed order to obtain a preliminary list of products;

[0052] The initial product list is checked for inventory in real time to obtain the reconstructed product list.

[0053] To address the aforementioned problems, this invention also provides an intelligent product selection system based on a large AI model, the system comprising:

[0054] The product graph construction module is used to collect users' multi-level product query commands, use a trained AI model to identify the product filtering dimensions of the multi-level product query commands, and use the product filtering dimensions to construct the user's product query graph.

[0055] The product primary screening module is used to query the user's historical screening information, assign node query degree values ​​to the product query graph based on the historical screening information to obtain an assigned graph, construct a product screening candidate pool corresponding to the user using the assigned graph, calculate the product convergence value of the product screening candidate pool, and perform primary screening of the product screening candidate pool based on the product convergence value to obtain primary screened products.

[0056] The satisfaction judgment module is used to select the primary selected products as the final selected products when the primary selected products meet the user's product needs.

[0057] The product structure reconstruction module is used to query the user's interaction behavior with the primary filtered products when the primary filtered products do not meet the user's product needs, construct the user's product preference structure tree based on the interaction behavior, perform cross-dimensional correlation feature mining on the product preference structure tree to obtain multi-level correlation features, and use the multi-level correlation features to reconstruct the product preference structure tree to obtain the reconstructed structure tree.

[0058] The final filtering module is used to analyze the reconstructed filtering products corresponding to the user using the reconstructed structure tree. When the reconstructed filtering products do not meet the user's product needs, it returns to the step of querying the user's interaction behavior with the primary filtering products until the reconstructed filtering products meet the user's product needs, at which point the reconstructed filtering products are used as the final filtering products.

[0059] This invention first uses technology to analyze user text / voice commands, extracting basic dimensions such as category and price, as well as fuzzy requirements such as "high cost-performance ratio." Then, it maps the analysis results to a product knowledge graph, generating a structured query tree with multi-level nodes. Next, based on user historical behavior data, it assigns weights to the query tree nodes using a collaborative filtering algorithm, constructing an initial multi-level filtering decision tree. This transforms the user's natural language requirements into machine-processable structured conditions, and combines historical preferences to generate a personalized filtering framework, laying the foundation for subsequent dynamic filtering. Furthermore, this invention dynamically removes invalid filtering nodes (such as attributes corresponding to out-of-stock products) by accessing real-time inventory and price data, ensuring that filtering conditions are effective in real time. It also analyzes potential relationships between product attributes (such as the relationship between processor and battery life), providing a basis for completing fuzzy requirements (such as converting "high cost-performance ratio" into specific price and rating conditions). Finally, it optimizes the filtering order at each level by combining user preferences, real-time status, and completion conditions (such as prioritizing the display of highly rated products with limited inventory), improving filtering efficiency. Even further, this invention automatically generates lower-level candidate dimensions and constraints based on the upper-level filtering results, achieving intelligent transmission of conditions. Finally, by comprehensively considering factors such as user preferences, product ratings, and inventory, the filtered products are ranked, and a candidate set that balances personalization and practicality is output. Therefore, this invention can improve the alignment between the filtered product results and users' actual needs. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating an embodiment of the intelligent screening method for multi-level products based on a large AI model provided by the present invention.

[0061] Figure 2 This is a schematic diagram of a module for implementing an intelligent selection method for multi-level products based on a large AI model, as provided in an embodiment of the present invention.

[0062] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0063] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0064] This application provides a method for intelligently filtering multi-level products based on a large AI model. The executing entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for intelligently filtering multi-level products based on a large AI model can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0065] Example 1:

[0066] Reference Figure 1 The diagram shown is a flowchart illustrating a method for intelligent selection of multi-level products based on a large AI model, according to an embodiment of the present invention. In this embodiment, the method for intelligent selection of multi-level products based on a large AI model includes:

[0067] S1. Collect the user's multi-level product query commands, use a trained AI model to identify the product filtering dimensions of the multi-level product query commands, and use the product filtering dimensions to construct the user's product query graph.

[0068] This invention can transform the complex needs expressed by users in natural language into structured information that can be processed by AI models by collecting users' multi-level product query commands, thus providing the core analytical basis for intelligent product selection.

[0069] The multi-level product query instruction refers to a natural language instruction input by the user that includes multiple levels of filtering dimensions. For example, a user might input the following filtering instruction on an e-commerce platform: "Budget 5000-8000 yuan, want a phone with good camera and long battery life, preferably supporting wireless charging."

[0070] Specifically, the multi-level product query instructions can be obtained by collecting the user's product query statements through the interactive interface corresponding to the product.

[0071] Furthermore, by utilizing a trained AI model to identify the product filtering dimensions of the multi-level product query instructions, this embodiment of the invention can identify key filtering dimensions (such as price, material, or function) in the user's query, transforming ambiguous natural language into structured filtering conditions.

[0072] The trained AI model refers to an AI model obtained by using a pre-trained Transformer model (such as BERT) for sequence labeling and intent classification, combined with a product domain dictionary to enhance terminology recognition accuracy. It can perform semantic parsing on user-input text / voice commands and extract user needs. The trained AI model includes a text / voice conversion layer, a text recognition layer, and a parsing information output layer.

[0073] Optionally, the product filtering dimension can be obtained by converting the user's input text or voice command into a text sequence, and then using the text recognition layer in the trained AI model to identify entities (such as "Apple mobile phone" corresponding to "brand-Apple" and "category-mobile phone") and modifiers (such as "budget within 5000 yuan" corresponding to "price ≤ 5000 yuan") in the command.

[0074] Furthermore, by utilizing the product filtering dimensions, this embodiment of the invention constructs a product query graph for the user, thereby transforming the user's multi-dimensional filtering conditions into a product query graph and realizing the visual association of product attributes.

[0075] The product query graph refers to a hierarchical graph structure model formed by mapping the results of user filtering intent analysis to the product knowledge graph. Its nodes contain multi-level filtering elements (such as the first-level category node "electronic products", the second-level attribute node "battery life > 10 hours", and the third-level strategy node "promotional products"), and the edges represent the semantic relationships between nodes (such as "category → subcategory" and "attribute → product instance").

[0076] As an embodiment of the present invention, constructing the user's product query graph using the product filtering dimensions includes:

[0077] Using the product filtering dimensions, construct structured query conditions for multi-level products corresponding to the user's query;

[0078] The structured query conditions are then matched using a knowledge graph to obtain an extended dimension set.

[0079] The extended dimension set is initialized with graph nodes to obtain the basic graph nodes;

[0080] Static relationship modeling is performed on the basic graph nodes to obtain the product query graph.

[0081] The structured query conditions refer to the hierarchical organization of user filtering requirements (such as category, price, function, etc.) into a formal combination of conditions that can be processed by machines. The extended dimension set refers to the implicit filtering dimensions supplemented from the explicit user requirements through knowledge graph matching, forming a more complete set of conditions. The basic graph node refers to the atomic-level graph structure unit created after mapping the filtering dimensions (including explicit and extended conditions) to the graph database.

[0082] In the specific implementation process, the extracted product filtering dimensions (such as category "mobile phone", brand "Apple", price "≤5000 yuan") can be divided into a multi-level structure according to business logic. The first-level node is the category (such as "mobile phone"), the second-level node is the core attribute (such as "brand - Apple" and "price ≤5000 yuan"), and the third-level node is the fuzzy requirement to be filled in (such as "good at taking pictures"), forming structured query conditions. The structured query conditions are input into the product knowledge graph, and then the semantic matching algorithm (string matching + cosine similarity) is used to find related entities, supplementing the implicit dimensions not explicitly mentioned by the user (such as "lightweight laptop" matching the required attributes such as "weight ≤1.5kg" and "thickness ≤18mm" in the knowledge graph), generating extended dimensions. Degree set; Map each entity in the extended dimension set (e.g., "Category - Laptop", "Attribute - Weight ≤ 1.5kg") to an independent node in the graph database, assign it a unique identifier (e.g., Node Category_Laptop, Node Attribute_Weight ≤ 1.5kg), and label the node type (Category / Attribute / Constraint); According to the hierarchical logic of the product data (Category → Subcategory → Attribute), add semantic relationship edges (e.g., "belongs to", "contains", "constraint") between nodes to form a directed acyclic graph (DAG). For example, the "Subcategory - Lightweight Laptop" node is connected to the "Category - Laptop" node through the "belongs to" edge, and connected to the "Weight ≤ 1.5kg" node through the "required attribute" edge, forming a product query graph.

[0083] S2. Query the user's historical filtering information, assign node query degree values ​​to the product query graph based on the historical filtering information to obtain an assigned graph, construct a product filtering candidate pool corresponding to the user using the assigned graph, calculate the convergence value of the product filtering candidate pool, and perform primary product filtering on the product filtering candidate pool based on the convergence value to obtain primary filtered products.

[0084] This invention can help understand the historical query behavior or data of users by querying the user's historical filtering information, thereby facilitating the quantification of the implicit preference weight of users for product attributes, and also providing data support for subsequent personalized graph assignment.

[0085] Optionally, the historical filtering information can be obtained by collecting users' past operation behavior logs (e.g., clicks / skips / purchases) in real time and synchronizing them to the behavior analysis database.

[0086] In this embodiment of the invention, the node query degree of the product query graph is assigned based on the historical filtering information. The resulting assigned graph can have user-personalized preference characteristics and can distinguish the priority of filtering dimensions (such as automatically increasing the weight of the "brand" node that users often pay attention to).

[0087] The assigned graph refers to a personalized graph structure that assigns query weights to each node based on the product query graph and the user's historical filtering behavior.

[0088] As an embodiment of the present invention, the product query graph is assigned node query degree values ​​based on the historical filtering information to obtain an assigned graph, including:

[0089] The historical filtering information is subjected to user behavior standardization processing to obtain standardized behavior data;

[0090] Using the standardized behavioral data, construct the user's dimensional utility matrix;

[0091] The time decay weighting of the aforementioned dimensional utility matrix is ​​applied to obtain the time-effect weight table;

[0092] Using the aforementioned timeliness weight table, node weight mapping is performed on the product query graph to obtain the assigned graph.

[0093] The standardized behavioral data refers to the data that is converted into numerical data with uniform dimensions after the historical filtering behavior of users is grouped and statistically analyzed by dimension and then normalized. The dimension utility matrix is ​​a two-dimensional data table constructed with the filtering dimension as the column and the user as the row. The time-weighted table is a table generated by weighting the historical behavioral data with a time decay function based on the dimension utility matrix.

[0094] In the specific implementation process, user historical filtering behavior data (such as clicks, favorites, and purchase records) can be converted into unified quantitative indicators, weighted according to behavior type (e.g., purchase = 3, favorites = 2, clicks = 1, no interaction = 0), and normalized (e.g., divided by the maximum behavior value) to obtain standardized behavior data. For each behavior data in the dimensional utility matrix, a decay weight is calculated according to the time of the behavior (e.g., weight for the last 30 days = 1, for 60 days = 0.8, for 90 days = 0.6). The matrix elements are then weighted using the exponential decay formula (wt = original value × λd, λ = 0.95, d is the number of days between intervals) to generate a time-effect weight table. For example, if a user clicked on "Xiaomi phone" 1 year ago (original value 0.5), after a period of time... After time decay, the weight decreases by 0.3, while the weight of "Xiaomi Redmi series" collected one month ago (original value 0.6) remains at 0.57, showing a significant difference in the time-sensitive weight table. The dimension weight values ​​in the time-sensitive weight table are mapped to the nodes of the product query graph. Through the node attribute update statement of the graph database, each node is assigned a query degree weight, forming a value assignment graph that includes personalized preferences. For example, if a user has repeatedly filtered for phones with "long battery life + 256GB storage", the weight of the node "attribute-battery capacity ≥ 4000mAh" in the value assignment graph is set to 0.9, and the weight of the node "attribute-storage capacity = 256GB" is set to 0.8. Products containing these two nodes will be given priority in the candidate set during the filtering process.

[0095] Preferably, constructing the user's dimensional utility matrix using the standardized behavioral data includes:

[0096] Using the standardized behavioral data, calculate the node exposure weight and node conversion rate weight of the product query graph;

[0097] Based on the node exposure weight and the node conversion rate weight, a comprehensive weight assignment is performed on the product query graph to obtain a comprehensive weight assignment graph.

[0098] Based on the comprehensive weight assignment graph, the user's dimensional utility matrix is ​​constructed.

[0099] The node exposure weight refers to the percentage of times a certain graph node (such as category, brand, attribute) is exposed in the user's historical filtering. The node conversion rate weight refers to the probability that a user will take effective action (such as click, favorite, purchase) after being exposed to a certain graph node. The comprehensive weight assigned graph refers to a new personalized graph structure obtained by assigning comprehensive weights to the graph node attributes based on the product query graph.

[0100] In the specific implementation process, priority can be assigned to exposure weight and conversion rate weight according to business needs (e.g., exposure accounts for 40%, conversion accounts for 60%). The comprehensive weight of each node is calculated using a linear weighting formula (comprehensive weight = exposure weight × 0.4 + conversion weight × 0.6). Then, the node attributes are updated in batches through the graph database to generate a valued graph with comprehensive weights. For example, the "attribute - long battery life" node has an exposure weight of 0.5 and a conversion weight of 0.9. The comprehensive weight is calculated as 0.5 × 0.4 + 0.9 × 0.6 = 0.74 using a 6:4 weighting. The preference attribute of this node in the valued graph is set to 0.74 to indicate the user's preference intensity for "long battery life". Using the nodes in the product query graph (e.g., category, brand, attribute) as columns and users as rows, the comprehensive weight of the nodes is filled into the corresponding cells to form a two-dimensional table (dimensional utility matrix).

[0101] Furthermore, as another optional embodiment of the present invention, the step of using the standardized behavioral data to calculate the node exposure weight and node conversion rate weight of the product query graph includes:

[0102] Identify the behavioral nodes and behavioral types in the standardized behavioral data that correspond to the product query graph;

[0103] Based on the behavioral nodes, the node exposure weight of the product query graph is calculated using the following formula:

[0104]

[0105] Among them, A i E represents the exposure percentage of behavior node i. i E represents the historical exposure count of the i-th behavior node, e represents the smoothing factor, and n represents the total number of behavior nodes at the same level in the product query graph. j Let α represent the historical exposure count of the j-th behavior node, and let α represent the node exposure weight. min(A i ) represents A i Maximum value, max(A) i ) represents A i The minimum value of ∈ represents the denominator protection parameter;

[0106] Based on the behavior type, the node conversion rate weight of the product query graph is calculated using the following formula:

[0107]

[0108] Among them, B k Let m represent the node conversion rate for the k-th behavior, and w represent the total number of behavior types. k C represents the weight of the k-th action. kLet μ represent the total number of times the k-th behavior occurs, β represent the node conversion rate weight, and μ represent the total number of times the behavior occurs. k σ represents the average conversion rate at the same level as k, σ represents the standard deviation, and ∈ represents the denominator protection parameter.

[0109] It should be further explained that the node exposure weight calculation formula quantifies user attention by measuring the historical exposure frequency of a node, introduces a smoothing mechanism to prevent new nodes from having zero weight, and then performs hierarchical normalization to make the exposure weights of different types of nodes comparable. This can accurately reflect users' browsing preferences for the filtering dimensions, solve the cold start problem of new product dimensions being ignored, and provide a basic priority basis for the filtering strategy based on "what users have viewed". The node conversion rate weight calculation formula calculates the node conversion rate by weighting clicks, purchases, and other behaviors according to their importance, and highlights the preference intensity above the average level through standardization. This comprehensively evaluates the actual degree of user recognition of the filtering dimensions, thereby distinguishing the purpose of shallow browsing and deep interaction behaviors, capturing users' true preferences (such as a significant increase in the weight of core dimensions for high-frequency purchases), and providing the filtering system with a more reliable deep preference signal of "what users recognize".

[0110] Furthermore, by utilizing the assigned graph, this embodiment of the invention can construct a product selection pool corresponding to the user, thereby filtering out product dimension combinations that meet the user's personalized needs, thus quickly narrowing down the range of candidate products and enabling the system to prioritize the set of products that the user may be interested in.

[0111] The product screening candidate pool refers to the set of priority candidates obtained by the system after narrowing down the screening scope based on the user's historical preferences, which includes products that the user may be interested in.

[0112] In practice, based on the historical behavior weights of nodes in the graph, high-weight dimension nodes can be selected and their corresponding product instances can be queried through the graph database to generate a candidate set containing products that users may be interested in.

[0113] This invention quantifies the degree of fit between the candidate pool and the user's historical preferences by calculating the convergence value of the candidate pool, ensuring that the screening results are neither too broad and unfocused nor too narrow and inaccurate. For example, if the candidate pool contains two categories of products, "Apple phone + long battery life" and "Apple phone + fast charging," and the variance of their matching degree at the "brand-Apple" node is 0.1 (low variance), then the convergence value is high, indicating that the candidate pool closely revolves around the user's core preferences. If "Xiaomi phone + long battery life" is mixed in, the variance rises to 0.6, the convergence value decreases, and the screening conditions need to be adjusted.

[0114] The convergence value refers to a quantitative indicator that measures the degree of fit between the candidate pool of selected products and the user's historical preferences.

[0115] As an embodiment of the present invention, calculating the convergence value of the product screening candidate pool includes:

[0116] The category concentration of the product screening candidate pool is calculated using the following formula:

[0117]

[0118] Where γ represents category concentration, R represents the total number of categories in the product screening candidate pool, and p(c i H represents the proportion of the i-th category in the candidate pool. max This represents the maximum possible entropy value of the category distribution;

[0119] The attribute similarity of products in the product screening candidate pool is calculated using the following formula:

[0120]

[0121] Where s represents attribute similarity, N represents the total number of products in the product filtering candidate pool, and T i Let T represent the set of attributes for the i-th item. core This represents the core attribute set of the product selection candidate pool;

[0122] Based on the category concentration and the attribute similarity, the convergence value of the product screening candidate pool is calculated using the following formula:

[0123]

[0124] Where ω represents the convergence value, γ represents the category concentration degree, and s represents the attribute similarity.

[0125] It should be explained that the category concentration refers to a quantitative indicator that measures the degree of focus of the product screening candidate pool at the target category level, and the attribute similarity refers to a quantitative indicator that measures the degree of matching between candidate products and user preference attribute combinations.

[0126] It should be further explained that the convergence value calculation formula quantifies the overall fit between the candidate pool and user preferences by integrating category concentration and attribute similarity. Category concentration is based on the principle of information entropy and calculates the degree of focus of the candidate pool at the target category level (the ratio of the actual entropy value to the maximum possible entropy value). The closer the value is to 1, the more singular the category is. Attribute similarity measures the fit of attribute combinations by statistically analyzing the average matching ratio between candidate products and the user's core attribute set. The closer the value is to 1, the more accurate the attribute matching is.

[0127] Furthermore, in this embodiment of the invention, based on the convergence value, a preliminary screening of the product selection candidate pool is performed. The preliminary screening results in products that match the user's historical preferences in terms of category focus and attribute matching, thus quickly filtering out invalid products that deviate from the core needs.

[0128] Optionally, the initial screening of products can be achieved by pre-setting a convergence value screening threshold (e.g., ≥0.7), then traversing the candidate pool of products, calculating the fit score (i.e., convergence value) between the category and attribute combination of each product and user preferences, and then retaining products with high convergence values. Only products with convergence values ​​≥ the threshold are included, while low-fit products with scattered categories (e.g., spanning "phone" and "tablet" categories) or missing core attributes (e.g., user preference for "long battery life" but product battery life <3000mAh) are removed. For example, if the user candidate pool contains 10 "Apple phone" related products, of which 8 meet the criteria of "brand - Apple" and "battery capacity ≥4000mAh" (convergence value 0.8-0.9), and 2 are "Apple Watch" (category mismatch, convergence value 0.4) or "Apple phone - short battery life" (attribute mismatch, convergence value 0.5), only the top 8 high convergence value products are retained after the initial screening.

[0129] S3. When the primary filtered products meet the user's product needs, the primary filtered products are used as the final filtered products.

[0130] It should be clarified that when the initial filtered products meet the user's product needs, it means that the user's needs can be determined when the user engages in deep interactive behaviors such as adding to favorites, adding to cart, or purchasing. For example, adding to favorites and cart are high-threshold determination methods, while purchasing is an extremely high-threshold determination method. Of course, it can also be determined based on deep browsing behavior (spending ≥60 seconds on the details page) or comparison selection (adding products to comparison tools). The specific settings need to be combined with the actual application scenario.

[0131] S4. When the primary filtered products do not meet the user's product needs, query the user's interaction behavior with the primary filtered products, construct the user's product preference structure tree based on the interaction behavior, perform cross-dimensional correlation feature mining on the product preference structure tree to obtain multi-level correlation features, and use the multi-level correlation features to reconstruct the product preference structure tree to obtain the reconstructed structure tree.

[0132] It should be noted that when the initial filtered products do not meet the user's product needs, it means that the user has not initiated any interactive behavior that meets the conditions, such as adding to favorites, adding to cart, or purchasing.

[0133] Furthermore, by querying the user's interaction behavior with the initially filtered products, this embodiment of the invention can capture the user's real-time preference intensity and interest changes in various dimensions of the products (such as brand and attributes), providing direct feedback data for the dynamic optimization of the filtering strategy.

[0134] The interactive behavior refers to the user's direct interactive operations on the initially filtered products, including but not limited to clicking on the product details page, adding products to favorites, adding products to the shopping cart, submitting an order (purchase), adding comments, and sharing product links.

[0135] Optionally, the interactive behavior can be obtained by recording user behavior data such as clicking, adding to favorites, adding to cart, or purchasing products in the initial filter through system logs or databases in real time.

[0136] This invention, through the construction of a product preference structure tree based on the interactive behavior, can intuitively present the user's preference intensity across different filtering dimensions (e.g., nodes with higher frequency of purchase have higher weight).

[0137] The product preference structure tree refers to a tree-like model constructed based on the user's historical interaction behavior (clicks, favorites, purchases, etc.) and the hierarchical relationship of product selection dimensions (category → brand → attribute → model).

[0138] As an embodiment of the present invention, based on the interaction behavior, a product preference structure tree for the user is constructed, including:

[0139] Extract the real-time features of the interaction behavior to obtain a behavior feature set;

[0140] Using the aforementioned behavioral characteristics, construct the user's product preference nodes;

[0141] Using the product preference nodes, construct the user's product preference tree architecture;

[0142] Incrementally update the product preference tree structure to obtain the product preference tree structure.

[0143] The product preference node refers to the entity unit of the filtering dimension extracted based on user interaction behavior (click / favorite / purchase), and the product preference tree architecture refers to the tree-like hierarchical model formed by organizing preference nodes according to the product hierarchy structure (such as category → subcategory → brand → attribute → model).

[0144] In practice, user actions such as clicking, adding to favorites, adding to cart, and purchasing of initially filtered products can be captured in real time. Behavior types (e.g., purchase = 5 points, adding to cart = 3 points, click = 1 point), timestamps, associated product IDs, and corresponding filtering dimensions (category / brand / attribute) are extracted to generate weighted behavior vectors, which are then aggregated into a behavior feature set. These behavior features are mapped to standard entities in the product knowledge graph (e.g., matching "Apple" to the "brand" node using the Neo4j graph database, and "long battery life" to the "battery capacity ≥ 4000mAh" attribute node). Initial node weights are calculated based on behavior weights (purchase > adding to cart > click) and a time decay function to generate weighted preference nodes. These nodes are then processed according to the product hierarchy (category → ...). The preference nodes are organized into subcategories (brands, attributes, and models). The parent-child relationship between nodes is determined by a decision tree algorithm (such as CART) (e.g., "Category - Mobile Phone" is the root node, and the child nodes are the frequently interacted subcategories "Foldable Screen" and "Candybar Phone"). This forms the product preference tree structure. When a new interaction occurs, the weight of the corresponding node is updated by an incremental learning algorithm (such as online ALS collaborative filtering) ("Brand - Xiaomi" from 0.3 to 0.5, "Subcategory - Foldable Screen" weight remains at 0.8). The path weight from the parent node to the child node is recalculated (e.g., "Mobile Phone → Foldable Screen → Xiaomi" path weight from 0.24 to 0.4). This ensures that the tree structure reflects the latest user preferences in real time, thus forming the product preference tree structure.

[0145] Furthermore, this embodiment of the invention performs cross-dimensional correlation feature mining on the product preference structure tree to obtain multi-level correlation features. This allows for the mining of hidden correlations between nodes of different dimensions from the tree structure of user preferences, integrating fragmented user preferences into a computable multi-level feature combination, and improving the coverage of potential needs by the screening conditions.

[0146] As an embodiment of the present invention, cross-dimensional correlation feature mining is performed on the product preference structure tree to obtain multi-level correlation features, including:

[0147] The product preference structure tree is transformed into a graph structure to obtain a product dimension association graph;

[0148] Frequent subgraph mining is performed on the product dimension association graph to obtain a set of high-frequency association patterns;

[0149] Path calculation is performed on the high-frequency association pattern set to obtain weighted association paths;

[0150] Graph neural network learning is performed on the weighted association paths to identify potential association features in the product preference structure tree;

[0151] Similarity clustering is performed on potential association features to obtain multi-level association features.

[0152] The product dimension association graph refers to a graph structure constructed with product attributes, categories, brands, etc. as nodes and the relationships between dimensions as edges. The high-frequency association pattern set refers to a set of frequently co-occurring dimension combinations extracted from the product dimension association graph through data mining algorithms. The weighted association path refers to the path connecting different nodes in the product dimension association graph.

[0153] In the specific implementation process, a graph database (such as Neo4j) can be used to transform nodes in the tree structure (such as "category-mobile phone" and "brand-Apple") into graph nodes. The hierarchical relationship between parent and child nodes is transformed into directed edges, and the edge weights are assigned the historical co-occurrence frequency between nodes (such as the number of times a user simultaneously filters "Apple" and "long battery life"), forming a weighted directed graph (product dimension association graph). A graph pattern matching algorithm is used to search for subgraph structures with support ≥ a threshold (such as 30%) in the graph, and close subgraphs containing cross-dimensional nodes (category + brand + attribute) are extracted as high-frequency association patterns. The shortest weighted path between any two nodes in the high-frequency subgraph is calculated based on Dijkstra's algorithm (minimum sum of edge weights), or the path weight (such as the weight of the contribution of purchase behavior) is calculated through an attention mechanism. The algorithm generates optimal transmission paths for cross-dimensional associations (e.g., the "brand → attribute" path has a higher weight than the "category → attribute" path) by weighting the association paths with 60% weight and click behavior accounting for 40%. It also uses graph neural networks (e.g., GCN) to learn node embeddings in the weighted association paths, capturing implicit preference connections that exist in the product dimension association graph but are not explicitly interacted with by users (e.g., the potential association between "Xiaomi" and "fast charging" formed by their co-occurrence in the "high-frequency screening by student users" scenario). Finally, it uses the K-means clustering algorithm to cluster the potential association features, dividing them into clusters according to association levels (first-level category association, second-level brand attribute association, and third-level model / specification association), generating multi-level features that can be directly used for combining screening conditions (e.g., first-level "electronic products", second-level "phone - Apple", and third-level "long battery life - 256GB").

[0154] This invention reconstructs the product preference structure tree by utilizing the multi-level association features, resulting in a reconstructed structure tree that integrates multi-level association features. It also dynamically adjusts the node levels and weights of the user preference tree to make the tree structure more closely reflect the user's actual needs.

[0155] In practice, multi-level related features (such as the frequently co-occurring "brand-attribute-subcategory" combination) can be mapped to the product preference structure tree, and the node hierarchy relationship can be dynamically adjusted (such as adding cross-dimensional direct related edges) and the node weights can be updated (assigned according to the strength of the association) to form a reconstructed tree structure that fits the user's potential preferences.

[0156] S5. Using the reconstructed structure tree, analyze the reconstructed filtered products corresponding to the user. When the reconstructed filtered products do not meet the user's product needs, return to the step of querying the user's interaction behavior with the primary filtered products, until the reconstructed filtered products meet the user's product needs, and then use the reconstructed filtered products as the final filtered products.

[0157] This invention utilizes the reconstructed structure tree to analyze the reconstructed filtered products corresponding to the user, which can accurately locate the combination of product features preferred by the user, quickly filter out products that deviate from the core preference dimension, and improve the matching of the filtered products with the user's actual needs.

[0158] As an embodiment of the present invention, the reconstructed structure tree is used to analyze the reconstructed filtered products corresponding to the user, including:

[0159] The reconstructed structure tree is encoded using conditional vectors to obtain the filtering condition vectors;

[0160] Using the filtering condition vector, construct a candidate product score table corresponding to the user;

[0161] The candidate product score table is sorted in a multi-dimensional mixed order to obtain a preliminary list of products;

[0162] The initial product list is checked for inventory in real time to obtain the reconstructed product list.

[0163] The filtering condition vector refers to a low-dimensional dense numerical vector transformed from nodes and their preference weights in the reconstructed tree structure by a graph encoder. The candidate product score table refers to a two-dimensional table formed by mapping the attribute data of candidate products to feature vectors, calculating the matching score with the filtering condition vector using cosine similarity, and superimposing historical preference weights (such as purchase behavior bonuses). The initial screening product list refers to an ordered set of products generated by using a multi-objective sorting algorithm (such as weighted summation) to sort the candidate product score table in descending order of total score, integrating dimensions such as matching score, product rating, and price fit.

[0164] In the specific implementation process, a graph encoder (such as GCN) can be used to convert the nodes of each screening dimension (category / brand / attribute) and their preference weights in the reconstructed structure tree into low-dimensional dense vectors. After aggregating cross-level related features, a numerical screening condition vector containing explicit and implicit user preferences is generated. The attribute data of candidate products are mapped to product feature vectors. The matching score (range 0-1) with the screening condition vector is calculated by cosine similarity. Combined with the historical preference weights in step (3) (an additional 0.2 points if there is a purchase behavior) to generate a comprehensive score (e.g., a matching degree of 0.8 for a certain product + weighted score of 0.15 = 0.95). A multi-objective ranking algorithm (such as weighted ranking algorithm) is used. The system integrates matching scores (60%), real-time product ratings (20%), and price fit (adding 0.1 points to products within the user-preset price range) from the score table. These scores are then sorted in descending order by total score (e.g., scores 0.95 > 0.85 > 0.7) to generate an initial screening list containing product ID, score, and dimension scores. The system then obtains the inventory status of the initial screening products via a real-time API interface (e.g., "Inventory > 0" is marked as 1, "Out of Stock" as 0), filters out products with zero inventory (e.g., a product with a score of 0.8 but zero inventory is removed), retains products with valid inventory, and re-sorts them by score to generate the final recommended reconstructed and screened product set.

[0165] It should be explained that the condition for determining when the reconstructed filtered products do not meet the user's product needs is the same as the condition for determining when the primary filtered products do not meet the user's product needs.

[0166] Furthermore, by revisiting the user's interaction with the initially filtered products and reconstructing the filtered products to meet the user's product needs, this invention continuously queries the user's real-time interaction with the filtered products (such as clicks and adding to cart) and dynamically updates the preference structure tree and filtering strategy, forming a closed loop of "interaction feedback - strategy optimization - result iteration". This ensures that the final output products accurately match the user's increasingly explicit / implicit needs.

[0167] Example 2:

[0168] like Figure 2 The diagram shown is a functional module diagram of the intelligent screening system for multi-level products based on a large AI model, as described in this invention.

[0169] The intelligent product selection system 200 based on an AI large-scale model described in this invention can be installed in an electronic device. Depending on the functions implemented, the intelligent product selection system based on an AI large-scale model may include a product map construction module 201, a primary product selection module 202, a satisfaction judgment module 203, a product structure reconstruction module 204, and a final selection module 205. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0170] In this embodiment of the invention, the functions of each module / unit are as follows:

[0171] The product graph construction module 201 is used to collect multi-level product query instructions from users, use a trained AI model to identify the product filtering dimensions of the multi-level product query instructions, and use the product filtering dimensions to construct the product query graph of the user.

[0172] The product primary screening module 202 is used to query the user's historical screening information, assign node query degree values ​​to the product query graph based on the historical screening information to obtain an assigned graph, construct a product screening candidate pool corresponding to the user using the assigned graph, calculate the product convergence value of the product screening candidate pool, and perform primary screening of the product screening candidate pool based on the product convergence value to obtain primary screened products.

[0173] The satisfaction judgment module 203 is used to select the primary screened products as the final screened products when the primary screened products meet the user's product needs.

[0174] The product structure reconstruction module 204 is used to query the user's interaction behavior with the primary filtered products when the primary filtered products do not meet the user's product needs, construct the user's product preference structure tree based on the interaction behavior, perform cross-dimensional correlation feature mining on the product preference structure tree to obtain multi-level correlation features, and use the multi-level correlation features to reconstruct the product preference structure tree to obtain a reconstructed structure tree.

[0175] The final filtering module 205 is used to analyze the reconstructed filtering products corresponding to the user using the reconstructed structure tree. When the reconstructed filtering products do not meet the user's product needs, it returns to the step of querying the user's interaction behavior with the primary filtering products until the reconstructed filtering products meet the user's product needs, and then uses the reconstructed filtering products as the final filtering products.

[0176] In detail, the modules in the AI-based large-scale intelligent product screening system 200 described in this embodiment of the invention employ the same methods as described above. Figure 1 The method described above uses the same technical means as the AI-based large model for intelligent screening of multi-level products and can produce the same technical effect, so it will not be elaborated here.

[0177] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0178] Finally, it should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for realizing intelligent screening of multi-level commodities based on an AI large model, characterized in that, The method comprises: Collecting a multi-level commodity query instruction of a user, identifying a commodity screening dimension of the multi-level commodity query instruction by using a trained AI large model, and constructing a commodity query graph of the user by using the commodity screening dimension; Querying historical screening information of the user, assigning a node query degree to the commodity query graph based on the historical screening information, obtaining an assigned graph, constructing a commodity screening candidate pool corresponding to the user by using the assigned graph, calculating a product convergence value of the commodity screening candidate pool, performing a commodity primary screening on the commodity screening candidate pool based on the product convergence value, and obtaining a primary screening commodity; When the primary screening commodity meets the commodity demand of the user, the primary screening commodity is taken as a final screening commodity; When the primary screening commodity does not meet the commodity demand of the user, querying an interaction behavior of the user on the primary screening commodity, constructing a commodity preference structure tree of the user based on the interaction behavior, performing cross-dimension associated feature mining on the commodity preference structure tree, obtaining a multi-level associated feature, and reconstructing the structure of the commodity preference structure tree by using the multi-level associated feature to obtain a reconstructed structure tree; Using the reconstructed structure tree to analyze a reconstructed screening commodity corresponding to the user, when the reconstructed screening commodity does not meet the commodity demand of the user, returning to the step of querying the interaction behavior of the user on the primary screening commodity, and taking the reconstructed screening commodity as a final screening commodity when the reconstructed screening commodity meets the commodity demand of the user. 2.The AI large model-based multi-level commodity intelligent screening method of claim 1, wherein, Constructing the commodity query graph of the user by using the commodity screening dimension comprises: Constructing a structured query condition of a multi-level commodity corresponding to the user by using the commodity screening dimension; Performing knowledge graph matching on the structured query condition to obtain an extended dimension set; Performing graph node initialization on the extended dimension set to obtain a basic graph node; Performing static relationship modeling on the basic graph node to obtain a commodity query graph. 3.The method of claim 1, wherein the method comprises: Assigning a node query degree to the commodity query graph based on the historical screening information to obtain an assigned graph comprises: Performing user behavior standardization processing on the historical screening information to obtain standardized behavior data; Constructing a dimension utility matrix of the user by using the standardized behavior data; Performing time decay weighting on the dimension utility matrix to obtain a time-efficiency weight table; Mapping node weights of the commodity query graph by using the time-efficiency weight table to obtain an assigned graph. 4.The method of claim 3, wherein the method further comprises: determining a first AI model based on a first category of the plurality of categories; determining a second AI model based on a second category of the plurality of categories; and determining a third AI model based on a third category of the plurality of categories. The method comprises: Calculating node exposure weight and node conversion rate weight of the commodity query graph by using the standardized behavior data; Assigning a comprehensive weight to the commodity query graph based on the node exposure weight and the node conversion rate weight to obtain a comprehensive weight assigned graph; Constructing a dimension utility matrix of the user based on the comprehensive weight assigned graph. 5.The method of claim 4, wherein the method further comprises: determining a first AI model based on a first category of the plurality of categories; determining a second AI model based on a second category of the plurality of categories; and determining a third AI model based on a third category of the plurality of categories. The method comprises: Identify the behavior node and behavior type corresponding to the commodity query graph in the standardized behavior data; Based on the behavior node, the node exposure weight of the commodity query graph is calculated using the following formula: wherein A i represents the exposure proportion of the behavior node i, E i represents the historical exposure times of the i-th behavior node, e represents a smoothing factor, n represents the total number of behavior nodes in the same level in the commodity query graph, E j represents the historical exposure times of the j-th behavior node, a represents a node exposure weight, min(A i ) represents the maximum value of A i , max(A i ) represents the minimum value of A i , and ∈ represents a denominator protection parameter. Based on the behavior type, the node conversion rate weight of the commodity query graph is calculated using the following formula: where B k denotes the conversion rate of the kth behavior, m denotes the total number of behavior types, w k denotes the weight of the kth behavior, C k denotes the total order of the kth behavior, β denotes the weight of the conversion rate, μ k denotes the average conversion rate of the same level as k, σ denotes the standard deviation, and ∈ denotes the denominator protection parameter. 6.The method of claim 1, wherein the method further comprises: determining a first AI model based on a first category of the plurality of categories; determining a second AI model based on a second category of the plurality of categories; and determining a third AI model based on a third category of the plurality of categories. The product convergence value of the commodity screening candidate pool is calculated, including: The category concentration of the commodity screening candidate pool is calculated; The attribute similarity of the commodities in the commodity screening candidate pool is calculated; Based on the category concentration and the attribute similarity, the product convergence value of the commodity screening candidate pool is calculated. 7.The method of claim 1, wherein the method further comprises: determining a first AI model based on a first category of the plurality of categories; determining a second AI model based on a second category of the plurality of categories; and determining a third AI model based on a third category of the plurality of categories. Based on the interaction behavior, the commodity preference structure tree of the user is constructed, including: Extracting real-time features of the interaction behavior to obtain a behavior feature set; Using the behavior features, the commodity preference node of the user is constructed; Using the commodity preference node, the commodity preference tree architecture of the user is constructed; Incrementally updating the commodity preference tree architecture to obtain a commodity preference structure tree.

8. The intelligent screening method for multi-level goods based on a large AI model as described in claim 1, characterized in that, Cross-dimension association feature mining is performed on the commodity preference structure tree to obtain multi-level association features, including: Converting the commodity preference structure tree into a graph structure to obtain a commodity dimension association graph; Performing frequent subgraph mining on the commodity dimension association graph to obtain a high-frequency association pattern set; Performing path calculation on the high-frequency association pattern set to obtain a weighted association path; Performing graph neural network learning on the weighted association path to identify the latent association features of the commodity preference structure tree; Performing similarity clustering on the latent association features to obtain multi-level association features. 9.The method of claim 1, wherein the method further comprises: determining a first AI model based on a first category of the plurality of categories; determining a second AI model based on a second category of the plurality of categories; and determining a third AI model based on a third category of the plurality of categories. Using the reconstructed structure tree, the reconstructed screening commodities corresponding to the user are analyzed, including: Condition vector encoding is performed on the reconstructed structure tree to obtain a screening condition vector; Using the screening condition vector, a candidate commodity score table corresponding to the user is constructed; Multi-dimensional mixed sorting is performed on the candidate commodity score table to obtain a preliminary screening commodity list; Real-time inventory verification is performed on the preliminary screening commodity list to obtain reconstructed screening commodities.

10. An intelligent screening system for multi-level goods based on an AI large model, characterized in that, The system includes: A commodity graph construction module for collecting multi-level commodity query instructions of a user, identifying the commodity screening dimensions of the multi-level commodity query instructions using a trained AI large model, and constructing a commodity query graph of the user using the commodity screening dimensions; A commodity primary screening module for querying historical screening information of the user, assigning node query degrees to the commodity query graph based on the historical screening information to obtain an assigned graph, constructing a commodity screening candidate pool corresponding to the user using the assigned graph, calculating the product convergence value of the commodity screening candidate pool, and performing commodity primary screening on the commodity screening candidate pool based on the product convergence value to obtain a primary screening commodity; A satisfaction judgment module for determining whether the primary screening commodity meets the user's commodity demand and selecting the primary screening commodity as the final screening commodity when the primary screening commodity meets the user's commodity demand. The commodity structure reconstruction module is configured to, when the primary screening commodity does not meet the commodity demand of the user, query an interaction behavior of the user on the primary screening commodity, construct a commodity preference structure tree of the user based on the interaction behavior, perform cross-dimension associated feature mining on the commodity preference structure tree to obtain a multi-level associated feature, reconstruct a structure of the commodity preference structure tree by using the multi-level associated feature, and obtain a reconstructed structure tree; The final screening module is configured to analyze a reconstructed screening commodity corresponding to the user by using the reconstructed structure tree, return to the step of querying the interaction behavior of the user on the primary screening commodity when the reconstructed screening commodity does not meet the commodity demand of the user, and take the reconstructed screening commodity as a final screening commodity when the reconstructed screening commodity meets the commodity demand of the user.

Images