An import and export commodity intelligent classification method based on knowledge graph metadata topology

Abstract

The application discloses an import and export commodity intelligent classification method based on knowledge graph metadata topology, relates to the technical field of reinforcement learning, and comprises the following steps: inputting initial data packets into a dynamic interaction model, performing explicit association mining in a semantic enhancement layer, optimizing rule matching paths in a rule evolution layer, and constructing a dynamic commodity knowledge graph; performing topology structure derivation on the dynamic commodity knowledge graph, generating a graph topology analysis report and a metadata list, integrating knowledge reasoning on the graph topology analysis report and the metadata list, and generating an intelligent navigation engine; calling the intelligent navigation engine to perform multi-path semantic query and rule verification on the dynamic knowledge graph, generating a candidate classification scheme set; performing multi-objective collaborative optimization on the candidate classification scheme set, generating a sorting scheme sequence, performing traceability packaging on the sorting scheme sequence, and outputting an intelligent classification scheme. The application improves the rule adaptation efficiency in complex scenarios through the dynamic interaction model and multi-objective collaborative optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of reinforcement learning, and particularly relates to an import and export commodity intelligent classification method based on knowledge graph metadata topology. BACKGROUND

[0002] The import and export commodity intelligent classification is a core technical link in the international trade customs clearance process, and the essence is to map the commodity to a standardized classification system by analyzing commodity attributes, declaration records and external knowledge. The traditional classification method mainly relies on artificial experience and static rule base. In the early stage, classification is mainly based on artificial interpretation, relying on the single-dimensional judgment of customs officers on commodity name, material, purpose and other text information. With the development of reinforcement learning technology, automatic methods based on rule engine have gradually become popular, and preliminary automation is achieved through preset keyword matching.

[0003] The main shortcomings of the prior art are that the fusion of multi-modal information only stays at the feature splicing level, a dynamic association network of cross-modal knowledge is not constructed, and it is difficult to capture implicit attributes; rule updating relies on manual intervention, and cannot adapt to the rapid changes of trade policies and commodity characteristics; there is a lack of deep mining of causal relationships, and it is difficult to extract dynamic rules from historical data, resulting in limited classification accuracy in complex scenarios. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an import and export commodity intelligent classification method based on knowledge graph metadata topology to solve the problems of shallow multi-modal information fusion and insufficient causal relationship mining.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The present application provides an import and export commodity intelligent classification method based on knowledge graph metadata topology, which comprises,

[0008] Collecting import and export commodity information and preprocessing, performing multi-modal feature fusion on the preprocessed import and export commodity information, and generating an initial data packet;

[0009] Inputting the initial data packet into a dynamic interaction model, performing explicit association mining in the semantic enhancement layer, optimizing the rule matching path in the rule evolution layer, and constructing a dynamic commodity knowledge graph;

[0010] Performing topological structure derivation on the dynamic commodity knowledge graph, generating a graph topology analysis report and a metadata list, integrating the graph topology analysis report and the metadata list into knowledge reasoning, and generating an intelligent navigation engine;

[0011] Call the intelligent navigation engine to perform multi-path semantic query and rule checking on the dynamic knowledge graph, and generate a candidate classification scheme set;

[0012] Multi-objective collaborative optimization is performed on the candidate classification scheme set to generate a sorted scheme sequence, and the sorted scheme sequence is traced and packaged to output an intelligent classification scheme.

[0013] As a preferred scheme of the import and export commodity intelligent classification method based on knowledge graph metadata topology, wherein: the import and export commodity information includes text information, multimedia information and historical declaration record;

[0014] The preprocessing includes data cleaning, format conversion, deduplication, normalization and outlier processing.

[0015] As a preferred scheme of the import and export commodity intelligent classification method based on knowledge graph metadata topology, wherein: the initial data packet is generated, and the specific steps are as follows,

[0016] The preprocessed text information is subjected to entity recognition and semantic coding to generate a commodity semantic vector;

[0017] The preprocessed multimedia information is subjected to structured analysis to generate an image representation vector;

[0018] The preprocessed historical declaration record is subjected to historical pattern mining to generate a declaration mode vector;

[0019] The commodity semantic vector, image representation vector and declaration mode vector are spliced to generate an initial data packet.

[0020] As a preferred scheme of the import and export commodity intelligent classification method based on knowledge graph metadata topology, wherein: the dynamic interaction model is constructed, and the specific steps are as follows,

[0021] Call and initialize the full connection network and multi-head Transformer to build a semantic enhancement layer and a rule evolution layer;

[0022] The semantic enhancement layer and the rule evolution layer are subjected to back propagation and gradient update using mean square error loss function to construct a dynamic interaction model.

[0023] As a preferred scheme of the import and export commodity intelligent classification method based on knowledge graph metadata topology, wherein: the dynamic commodity knowledge graph is constructed, and the specific steps are as follows,

[0024] The initial data packet is input into the dynamic interaction model, the semantic enhancement layer performs explicit association mining through the remote supervision method to form an attribute rule association matrix;

[0025] The rule evolution layer uses a near-end policy optimization algorithm to optimize the rule matching path and generate an optimized rule set;

[0026] The attribute rule association matrix and the optimized rule set are integrated through the feature splicing channel to construct a dynamic product knowledge graph.

[0027] As a preferred embodiment of the intelligent classification method for import and export commodities based on knowledge graph metadata topology as described in this invention, the specific steps for generating the knowledge graph topology analysis report and metadata list are as follows:

[0028] Perform global topological metric analysis on the dynamic product knowledge graph to generate an original set of topological indicators;

[0029] Perform community structure detection and hub node identification on the original topological indicator set to generate a graph structure insight report;

[0030] Based on the graph structure insight report, perform pattern traversal on the dynamic product knowledge graph to generate a metadata list;

[0031] Integrate the graph structure insight report and metadata list to generate a graph topology analysis report.

[0032] As a preferred embodiment of the intelligent classification method for import and export commodities based on knowledge graph metadata topology described in this invention, the specific steps for generating an intelligent navigation engine are as follows:

[0033] Parse the data schema of the metadata manifest to generate entity relationship type specifications;

[0034] Based on the graph topology analysis report, the structured business rules of the dynamic product knowledge graph are extracted and logically encoded to generate an executable rule set;

[0035] Perform structured query transformation on entity relationship type specifications to generate a parameterized query instruction template library;

[0036] The parameterized query instruction template library and executable rule set are logically coordinated and integrated to generate a rule and query coordinator.

[0037] The rules and query coordinator are encapsulated and their protocols are adapted to generate a standardized classification engine.

[0038] The standardized classification engine undergoes performance stress testing and stability optimization to generate an intelligent navigation engine.

[0039] As a preferred embodiment of the intelligent classification method for import and export commodities based on knowledge graph metadata topology described in this invention, the specific steps for generating a candidate classification scheme set are as follows:

[0040] The intelligent navigation engine is invoked to initiate multi-path semantic queries on the dynamic knowledge graph in parallel, generating an original candidate encoding set.

[0041] Perform compliance checks on the original candidate coding set to generate a valid candidate coding set;

[0042] Perform a multi-dimensional utility evaluation on the effective candidate coding set to generate a candidate classification scheme set.

[0043] As a preferred embodiment of the intelligent classification method for import and export commodities based on knowledge graph metadata topology described in this invention, the specific steps of the sorting scheme sequence are as follows:

[0044] Perform Pareto front search and non-dominated sorting on the candidate classification scheme set to generate a multi-objective optimization solution set;

[0045] Decision ranking and robustness verification for multi-objective optimization solution sets are performed, and a sequence of ranking schemes is generated.

[0046] As a preferred embodiment of the intelligent classification method for import and export commodities based on knowledge graph metadata topology described in this invention, the specific steps for outputting the intelligent classification scheme are as follows:

[0047] A priority truncation operation is performed on the sorting scheme sequence to generate the optimal recommendation conclusion, and a multi-source decision evidence association is performed on the optimal recommendation conclusion to generate a preliminary source evidence chain;

[0048] The optimal recommendation conclusion and the preliminary source evidence chain are subjected to cross-modal information fusion and visualization structure encapsulation to output an intelligent classification scheme.

[0049] The beneficial effects of this invention are as follows: by deeply mining the semantic associations of multimodal data through a dynamic interaction model, dynamic optimization of rule matching is achieved, which improves the rule adaptation efficiency in complex scenarios; and multi-objective collaborative optimization is carried out simultaneously to generate a sequence of sorting schemes, which improves the scenario adaptability of the schemes and the transparency and interpretability of the decision. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Fig. 1 This is a flowchart of an intelligent classification method for import and export commodities based on knowledge graph metadata topology.

[0052] Fig. 2 A flowchart for building a dynamic product knowledge graph.

[0053] Fig. 3 A flowchart generated for the intelligent navigation engine.

[0054] Fig. 4 A flowchart generated for the intelligent classification scheme. Detailed Implementation

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0058] Reference Figs. 1-4 This is one embodiment of the present invention, which provides a method for intelligent classification of import and export commodities based on knowledge graph metadata topology, including the following steps:

[0059] S1. Collect import and export commodity information and preprocess it. Perform multimodal feature fusion on the preprocessed import and export commodity information to generate an initial data packet.

[0060] S1.1 Import and export commodity information includes text information, multimedia information, and historical declaration records;

[0061] It should be noted that the text information includes the product name, specifications, and material, which is obtained through the enterprise's ERP (Enterprise Resource Planning) and electronic data exchange of customs declarations; the multimedia information includes product images and packaging videos, which are obtained through importing test reports, product image data provided by suppliers, and real-time collection by IoT devices; and the historical declaration records are extracted from the customs database and the enterprise's historical customs declaration files.

[0062] S1.2 Preprocessing includes data cleaning, format conversion, deduplication, normalization, and outlier handling;

[0063] It should be noted that the collected import and export commodity information, including text information, multimedia information, and historical declaration records, undergoes data cleaning to remove incomplete, erroneous, or irrelevant data records. After data cleaning, the import and export commodity information undergoes format conversion to ensure that all data sources adhere to a unified format standard. Deduplication is used to eliminate duplicate data entries to ensure the uniqueness of the dataset. Normalization is used to map data values ​​from different sources to the same scale, avoiding undue influence on the results due to significant differences in magnitude of certain features. Outlier handling identifies and corrects data points that deviate from the normal range, ensuring the accuracy and reliability of the dataset.

[0064] S1.3 Perform entity recognition and semantic encoding on the preprocessed text information to generate product semantic vectors;

[0065] It should be noted that the text information is segmented and standardized to convert it into a word sequence, and special markers are added to identify the start and end of the text. A multi-layer bidirectional Transformer encoder is used to capture the semantic relationships between all words in the word sequence in the context (such as the semantic adaptation relationship between "ceramics" and "heating"). The semantic relationships of the whole text are aggregated to generate a product semantic vector.

[0066] S1.4 Perform structured parsing on the preprocessed multimedia information to generate image representation vectors;

[0067] It should be noted that the product images and packaging videos in the multimedia information are uniformly converted to a standard format and resolution. Grayscale conversion and noise filtering are performed on each frame of the packaging video and the image areas of the product images to remove shooting noise. Any pixel in the image area containing the product image is selected as a seed point. The processing queue is initialized, and the seed point is added to the processing queue. The initial state of the seed point is marked as pending processing. When there is a seed point to be processed in the processing queue, the seed point is retrieved, and adjacent pixels within the seed point's neighborhood with a grayscale difference less than a preset similarity threshold are marked as new seed points and added to the processing queue. This process is repeated until the processing queue is empty. All pixels marked as seed points are integrated to form a binary region. The coordinate sequence of the seed points in the binary region is obtained to form a closed product outline. The geometric features, color distribution, and texture features of the product outline are extracted. The geometric features include aspect ratio, area, and perimeter, and the texture features include smoothness and edge density. The geometric features, color distribution, and texture features are numerically encoded and weighted to generate an image representation vector.

[0068] It should also be noted that the similarity threshold is defined based on the gray value distribution of the image region, and an exemplary value range is (5, 30).

[0069] S1.5 Perform historical pattern mining on the preprocessed historical declaration records to generate declaration pattern vectors;

[0070] It should be noted that historical declaration records undergo data cleaning and field alignment to extract classification codes, declaration dates, declaration quantities, declaration prices, customs clearance status, and inspection results from past declarations of the same or similar commodities. The frequency of classification code changes and dwell time are statistically analyzed chronologically to obtain the fluctuation range and price range of declaration prices. Patterns of customs clearance status changes (such as frequent order cancellations and modifications) are identified, and common problem types and their frequency of occurrence in inspection results are analyzed. Statistical characteristics such as classification stability, price volatility, declaration success rate, and inspection anomaly rate are quantified and encoded to generate a declaration pattern vector.

[0071] S1.6. Concatenate the product semantic vector, image representation vector, and declaration pattern vector to generate an initial data packet;

[0072] It should be noted that the commodity semantic vector, image representation vector, and declaration mode vector are aligned to unify the vector dimensions and form a high-dimensional joint feature vector; the high-dimensional joint feature vector is normalized to eliminate dimensional differences; and the normalized high-dimensional joint feature vector is structured to generate an initial data packet in a unified format.

[0073] S2. Input the initial data package into the dynamic interaction model, the semantic enhancement layer performs explicit association mining, the rule evolution layer optimizes the rule matching path, and constructs a dynamic product knowledge graph.

[0074] S2.1 Call and initialize the fully connected network and multi-head Transformer to build the semantic enhancement layer and rule evolution layer;

[0075] It should be noted that the number of input layer nodes in the fully connected network is set according to the dimension of the initial data packet, and the hidden layers are set according to the number of input layer nodes. Each hidden layer uses the ReLU activation function for nonlinear transformation, and the weights are initialized using the Xavier uniform method. The number of output layer nodes in the fully connected network matches the number of input layer nodes, thus completing the construction of the semantic enhancement layer. Simultaneously, the multi-head Transformer (a neural network architecture based on a self-attention mechanism) component is initialized. Specifically, the query, key, and value weight matrices of the attention head are initialized using a normal distribution, and the position encoding uses sine and cosine functions to generate fixed vectors. The Kaiming initialization method is used. The method generates a feedforward network weight matrix and sets the bias term to a zero vector; adjusts the output dimension of the fully connected network hidden layer to be compatible with the input dimension of the multi-head Transformer; merges the outputs of the fully connected network and the multi-head Transformer through residual connection and layer normalization, and inputs them into the feedforward network for reinforcement learning; analyzes the correlation strength between the commodity semantic vector, image representation vector and declaration pattern vector in the initial data packet; dynamically adjusts the rule matching conditions (such as expanding the inclusiveness of material classification and correcting the price reference range); generates the initial rule path (such as "when the material is ceramic, the declaration code is 3926"); and completes the construction of the rule evolution layer.

[0076] S2.2 Utilize the mean squared error loss function to perform backpropagation and gradient update on the semantic enhancement layer and rule evolution layer to construct a dynamic interaction model;

[0077] It should be noted that the mean squared error loss function is used to perform forward propagation on the semantic enhancement layer and the rule evolution layer to obtain the prediction error matrix. The prediction error matrix is ​​then backpropagated to generate the gradient tensor of each layer. The gradient tensor of each layer is updated using the momentum gradient descent method to obtain the updated network weights. Based on the updated network weights, depthwise separable convolution is used to perform hyperparameter tuning on the semantic enhancement layer and the rule evolution layer. The model is then stacked across layers through residual connections to output the constructed dynamic interaction model.

[0078] The initial data packets are divided into a sample set, a training set, and a validation set according to their data type (e.g., a 7:2:1 ratio). Features are augmented in the sample set using linear interpolation, and data is standardized using batch normalization to form enhanced standard samples. The Adam optimizer is used to dynamically adjust the learning rate of the enhanced standard samples, and early stopping is applied simultaneously for training monitoring to obtain intermediate parameters. The root mean square error loss function is used to quantify the loss of the intermediate parameters to obtain the prediction error. When the prediction error exceeds the convergence threshold, training terminates, and the trained dynamic interactive model is output.

[0079] It should be noted that the convergence threshold is defined based on the rate of change of the prediction error on the validation set, and its value ranges from 0.0001 to 0.01.

[0080] S2.3 Input the initial data packet into the dynamic interaction model. The semantic enhancement layer performs explicit association mining through remote supervision to form an attribute rule association matrix.

[0081] It should be noted that attribute association pairs (such as "material is ceramic" and "image texture is white glaze" frequently co-occurring and successfully clearing customs in historical declaration records) are extracted from historical declaration records to construct an initial supervision signal library. Each attribute association pair is labeled with its historical co-occurrence frequency and customs clearance status as the basis for association. Multi-dimensional vectors such as commodity semantic vectors, image representation vectors, and declaration mode vectors in the initial data package are input into the semantic enhancement layer. Deep local features of each multi-dimensional vector (such as the semantic strength of the keyword "material" in the text) are extracted through a fully connected network. Based on the initial supervision signal library, the cosine similarity between deep local features and attribute associations is used for matching. The matched attribute association pairs are organized into an attribute rule association matrix according to attribute dimensions (such as material, image features, and declaration parameters). The rows of the attribute rule association matrix represent "source attributes" (such as "material words in the product name"), the columns represent "target attributes" (such as "texture features"), and the element values ​​are existence descriptions of the attribute association pairs (such as "frequent co-occurrence").

[0082] S2.4 The rule evolution layer uses the near-end strategy optimization algorithm to optimize the rule matching path and generate an optimized rule set;

[0083] It should be noted that, based on the declaration pattern vector and attribute rule association matrix, the initial data packet is input into the initial rule path for rule matching. The performance data after the rule is triggered in the initial rule path is recorded, including the declaration success rate, the verification anomaly rate, and the actual co-occurrence frequency of attribute association pairs. Using the near-end policy optimization algorithm, with the performance data as the objective function, the initial rule parameters (such as the range of price volatility) for rule matching are iteratively adjusted. In specific operations, based on the historical statistical declaration success rate and co-occurrence frequency of association pairs, the expected performance data under the current condition parameters is obtained through the policy network. At the same time, the policy gradient is obtained and the update step size is limited (such as setting ε=0.2 to avoid drastic rule changes). After iterative optimization, condition parameters are generated, the condition parameters are standardized, and the optimized rule set is generated.

[0084] It should also be noted that the initial rule parameters are defined based on the statistical features in the declaration mode vector. The initial rule parameters include numerical parameters, categorical parameters and logical parameters. The specific steps for obtaining them are as follows: the numerical parameters are set based on the quantitative results of the statistical features. For example, if the historical price fluctuation range is concentrated in ±10%, the initial rule parameter is set to "price fluctuation range ±10%".

[0085] Category parameters are obtained by statistically analyzing high-frequency categories of statistical features. For example, if 90% of the declaration codes for a certain type of product have the first two digits "73" (corresponding to steel products), then the initial parameter is set to "when the first two digits of the declaration code are 73, it is classified as steel products by default".

[0086] Logical parameters are generated by inferring from historical statistical correlations. For example, if historical declaration records show that products containing the keyword "fuzzy material" in their product descriptions have a higher-than-average inspection anomaly rate, the initial rule parameter can be set to "trigger automatic review when the product description contains the keyword fuzzy material";

[0087] The application success rate is obtained by statistically analyzing the proportion of successfully matched applications in historical application records to the total number of application records.

[0088] The inspection anomaly rate is obtained by calculating the proportion of inspection cases marked as abnormal during the reporting process to the total number of inspection cases.

[0089] The actual co-occurrence frequency of attribute association pairs is obtained by statistically analyzing the proportion of the number of times attribute association pairs appear simultaneously in historical declaration records to the total number of declaration records.

[0090] S2.5 Integrate the attribute rule association matrix and the optimized rule set through the feature splicing channel to construct a dynamic product knowledge graph;

[0091] It should be noted that in the feature splicing channel, dynamic dimensional alignment and feature fusion are performed on the attribute rule association matrix and the optimized rule set. Specifically, an attribute dimension mapping protocol is used to perform semantic similarity matching between the "source attribute" row of the attribute rule association matrix and the condition parameters of the optimized rule set (e.g., matching the "ceramic" lexical with the condition "material contains ceramic components") to complete bidirectional index alignment and eliminate dimensional bias. Statistical weights of the declaration pattern vector are applied to the association strength of the attribute rule association matrix (e.g., a 90% declaration success rate is used as a correction coefficient for co-occurrence frequency). At the same time, the condition parameters of the optimized rule set are linearly embedded and mapped to the same feature space as the attribute rule association matrix. The association weights between the attribute rule association matrix nodes (e.g., the "ceramic" material word node) and the optimized rule set nodes (e.g., the "HS code 3926 rule") are obtained through a gated attention mechanism. Weighted fusion is performed based on the association weights to generate a fused feature tensor. The fused feature tensor is then subjected to nonlinear feature transformation and dimensionality enhancement using the GeLU activation function to construct a dynamic commodity knowledge graph.

[0092] S3. Perform topological structure deduction on the dynamic product knowledge graph, generate a graph topological analysis report and metadata list, integrate the graph topological analysis report and metadata list with knowledge reasoning, and generate an intelligent navigation engine.

[0093] S3.1 Perform global topological metric analysis on the dynamic product knowledge graph to generate an original set of topological indicators;

[0094] It should be noted that the process involves traversing the dynamic product knowledge graph to obtain basic network features, including the total number of graph nodes, the total number of edges, the average degree value, and the graph density; obtaining node-level centrality indicators in parallel, including degree centrality, betweenness centrality, and proximity centrality; performing global path analysis simultaneously to obtain the average length of the shortest path between any two graph nodes in the dynamic product knowledge graph and to evaluate information transmission efficiency; and then structurally integrating the basic network features, node-level centrality indicators, and global path analysis results to generate an original set of topological indicators.

[0095] It should also be noted that degree centrality refers to the number of neighboring nodes of a graph node in a dynamic product knowledge graph.

[0096] Betweenness centrality refers to the percentage of times a graph node appears in the shortest path between all pairs of graph nodes in a dynamic product knowledge graph.

[0097] Proximity centrality refers to the reciprocal of the sum of the shortest path lengths from a graph node to other nodes in a dynamic product knowledge graph.

[0098] S3.2 Perform community structure detection and hub node identification on the original topological index set to generate a graph structure insight report;

[0099] It should be noted that the community structure detection of the original topological index set is performed as follows: each graph node in the original topological index set is treated as an independent community. The graph nodes in the original topological index set are traversed, and the increase in density score when each graph node moves to an adjacent community is obtained. Graph nodes that can increase the density score are moved to the corresponding community. The operation is repeated until the density score no longer increases, and the community division is completed. The number of graph nodes contained in each community is counted as the community size. The ratio of the actual number of edges in the community to the theoretical maximum possible number of edges is used as the internal connection density. The degree centrality and betweenness centrality of the graph nodes are used. Hub nodes are selected according to the degree centrality ranking result, and bridge nodes are selected according to the betweenness centrality ranking result. The community division results are integrated, including the graph nodes contained in each community, community size, internal connection density, hub nodes and bridge nodes, to generate a graph structure insight report.

[0100] It should also be noted that the density score is a quantitative indicator for measuring the rationality of grouping in a dynamic product knowledge graph. It evaluates the grouping quality by comparing the difference between the "actual density of connections between graph nodes within a group" and the "expected degree of connection under random conditions". The density score will be higher when the connections between graph nodes within a group are denser and the connections between graph nodes between groups are sparser.

[0101] S3.3 Based on the graph structure insight report, perform pattern traversal on the dynamic product knowledge graph and generate a metadata list;

[0102] It should be noted that, based on the community division results in the graph structure insight report, the dynamic product knowledge graph is divided into multiple structured functional blocks (such as "material association community" and "declaration rule community"). A node relationship pattern traversal is performed. Specifically, starting from the hub node, adjacent nodes are traversed along the hub node's associated edges (relationship types such as "belong to", "associate", and "depend on"), and node type combinations (such as "product node + material node + declaration code node"), relationship path patterns (such as "product → material → texture feature → rule association"), and co-occurrence frequencies are recorded. Simultaneously, metadata attributes are recorded. Specifically, the data source and constraints of each graph node discovered during the traversal are recorded (such as "white glaze association rule applies only when the material is ceramic"). The node relationship patterns and metadata attributes are then structured and arranged to form a metadata list.

[0103] S3.4 Integrate the graph structure insight report and metadata list to generate a graph topology analysis report;

[0104] It should be noted that the node relationship patterns in the metadata list are mapped to the community division results in the graph structure insight report, and the community boundaries and hubs and bridge nodes involved in each node relationship pattern are marked to generate a community structure overview. The macro-structural features at the community level (including community division results, community size, internal connection density, hub nodes and bridge nodes) and micro-interaction features (including node relationship patterns, co-occurrence frequency and metadata attributes) are correlated and analyzed to identify the close connection patterns within the community (such as the high-frequency co-occurrence of ceramic material nodes and white glaze texture nodes) and cross-community bridge nodes (such as key nodes connecting material communities and regular communities). The community structure overview, close connection patterns and cross-community bridge nodes are structured to generate a graph topology analysis report.

[0105] S3.5. Parse the data schema of the metadata list and generate entity relationship type specifications;

[0106] It should be noted that the types of all graph nodes (such as product nodes, material nodes, and declaration code nodes) and their corresponding relationship types (such as "belongs to", "associated with", "dependent on", and "compliant with") are extracted from the metadata list to form a set of node types and a set of relationship types. Based on the node relationship patterns and co-occurrence frequencies recorded in the metadata list, the interaction logic in the dynamic product knowledge graph is identified (such as the attribute and rule mapping relationship reflected in "material nodes are associated with declaration code nodes through texture features"). The set of node types, the set of relationship types, and the interaction logic are structured and arranged to form an entity relationship type specification.

[0107] S3.6 Based on the graph topology analysis report, extract the structured business rules of the dynamic product knowledge graph and perform logical encoding to generate an executable rule set;

[0108] It should be noted that the community division boundaries (such as "material-related communities") and the distribution characteristics of hub nodes and bridging nodes within the communities are extracted from the graph topology analysis report as the corresponding business functions for each community. Based on the tightly connected patterns and associated metadata attributes identified in the graph topology analysis report, frequently co-occurring graph node combinations and interaction paths (such as "product → material → texture → rule") are transformed into business logic (such as "when the product material is ceramic, prioritize associating the white glaze texture and matching the declaration code range 3926"). For cross-community bridge nodes and their corresponding association paths (such as "material characteristics → declaration condition mapping"), combined with the constraints in the metadata list, the logical conditions for cross-community interaction are extracted (such as "after the material attribute is passed to the rule community through the bridging node, the compatibility of the declaration code needs to be verified"). The business functions, business logic, and logical conditions are used as business rules and transformed into structured logical expressions to generate an executable rule set.

[0109] S3.7 Perform structured query transformation on entity relationship type specifications, generate parameterized query instruction template library, and logically coordinate and integrate the parameterized query instruction template library with executable rule set to generate rule and query coordinator;

[0110] It should be noted that the process involves parsing the interaction logic in the entity relationship type specification, identifying common query scenarios (such as "querying the declaration code range corresponding to a certain material"), input parameters (such as product ID), and output requirements (such as association rules), converting the relationship path pattern into a parameterized query instruction structure, and transforming metadata attributes into logical constraints (such as "there is a mapping between material texture features and declaration codes"). This process integrates and generates a parameterized query instruction template library containing fields such as "query scenario name," "input parameters," "query instruction," and "expected output." By analyzing the logical connections between the parameterized query instruction template library and the executable rule set (such as the need to query entity relationships before rule triggering), a rule query interaction protocol is constructed (such as parameter preprocessing before the parameterized query instruction template library calls the executable rule set). This enables query instructions in the parameterized query instruction template library to trigger relevant execution rules in the executable rule set (such as automatically associating declaration rule verification when querying materials), and generates a rule and query coordinator.

[0111] S3.8. Perform interface encapsulation and protocol adaptation on the rules and query coordinator to generate a standardized classification engine. Perform performance stress testing and stability optimization on the standardized classification engine to generate an intelligent navigation engine.

[0112] It should be noted that a unified standardized input / output interface is defined for the rule and query coordinator, and the internal processes of the rule and query coordinator are encapsulated as general services. At the same time, it adapts to external communication protocols to ensure the compatibility of data formats, authentication methods and transmission standards between different devices, thereby generating a standardized classification engine. By simulating high-concurrency query requests (such as thousands of product category queries per second), complex rule nesting (such as multi-level material-declaration rule association verification), and large-scale data interaction (such as querying the relationship path of tens of thousands of nodes), the response time, throughput and resource consumption of the standardized classification engine under extreme load are tested. For the performance bottlenecks (such as database query latency) and stability issues (such as long-term memory leaks) found in the test, targeted optimization is carried out by indexing the parameterized query instruction template library to accelerate it, simplifying the rule execution logic (such as pre-compiled condition judgment), and enhancing the fault tolerance mechanism (such as retrying failed requests), thereby generating an intelligent navigation engine.

[0113] S4. Call the intelligent navigation engine to perform multi-path semantic query and rule verification on the dynamic knowledge graph, and generate a set of candidate classification schemes;

[0114] S4.1, Call the intelligent navigation engine to initiate multi-path semantic queries on the dynamic knowledge graph in parallel to generate the original candidate encoding set;

[0115] It should be noted that the intelligent navigation engine is invoked to initiate multiple semantic path queries in parallel against the dynamic knowledge graph (such as "product → material → texture feature → declaration code", "product → function / use → classification rule → declaration code", and "product → directly associated declaration code node"). Each semantic path is based on the node relationship pattern in the dynamic knowledge graph and the relevant execution rule definitions in the executable rule set. The intelligent navigation engine's rules and query coordinator dynamically match and execute the optimal query path. The associated declaration codes and other relevant rule identifiers (such as classification basis and applicable conditions) are extracted from the results of each query path, and these are aggregated to form an original candidate code set containing duplicates and potential conflicts.

[0116] S4.2 Perform compliance checks on the original candidate coding set to generate a valid candidate coding set;

[0117] It should be noted that the navigation engine is invoked to match each application code in the original candidate code set with the executable rule set one by one, verifying whether it meets the triggering conditions (such as whether the material attribute is ceramic) and execution logic (such as whether it belongs to the specified code range 3926) of the relevant execution rules in the executable rule set. At the same time, it checks whether the node type associated with the application code conforms to the business rules in the dynamic knowledge graph, removes application codes that do not conform to the business rules, and summarizes the application codes that pass the verification to form a valid candidate code set.

[0118] S4.3 Perform a multi-dimensional utility evaluation on the effective candidate coding set to generate a candidate classification scheme set;

[0119] It should be noted that, based on the historical business logic and executable rule set in the dynamic knowledge graph, multi-dimensional evaluation indicators are determined, including the matching accuracy of declaration codes, historical declaration success rate, customs clearance timeliness, and inspection anomaly rate. These multi-dimensional evaluation indicators are obtained from the dynamic knowledge graph. Specifically, the matching accuracy is obtained by comparing the semantic similarity between valid candidate codes and the product semantic vector; the historical declaration success rate is obtained by the success rate of declaration codes for the same product and similar products in past records within the statistical declaration pattern vector; the customs clearance timeliness is obtained by the average customs clearance time of similar historical declaration codes within the statistical declaration pattern vector; and the inspection anomaly rate is obtained by the proportion of inspection anomaly cases corresponding to the declaration codes within the statistical declaration pattern vector. The valid candidate codes and the multi-dimensional evaluation indicators are weighted and summed to generate a comprehensive utility value. Each valid candidate code is then associated with its corresponding comprehensive utility value to form a candidate classification scheme set.

[0120] S5. Perform multi-objective collaborative optimization on the candidate classification scheme set to generate a sorting scheme sequence. Then, perform source tracing and encapsulation on the sorting scheme sequence to output an intelligent classification scheme.

[0121] S5.1 Perform Pareto front search and non-dominated sorting on the candidate classification scheme set to generate a multi-objective optimization solution set;

[0122] It should be noted that the multi-dimensional evaluation indicators in the candidate classification scheme set are transformed into optimization objectives, including maximizing matching accuracy, maximizing historical declaration success rate, maximizing customs clearance timeliness, and minimizing inspection anomaly rate. A "dominance relationship" is defined for each candidate scheme in the candidate classification scheme set. Specifically, if a candidate scheme in the candidate classification scheme set is not inferior to another candidate scheme in all optimization objectives, and is superior to another candidate scheme in at least one optimization objective, then the candidate scheme is determined to dominate another candidate scheme. All candidate schemes in the candidate classification scheme set are compared pairwise, and candidate schemes that are not dominated by any other candidate scheme are selected as the first-level non-dominated solution set. This process continues to select the second-level non-dominated solution set, which is dominated only by the first-level non-dominated solution set, and so on, forming a hierarchical ranking structure. This structured integration results in a multi-objective optimized solution set.

[0123] S5.2 Execute decision ranking and multi-scenario robustness verification on the multi-objective optimization solution set, and generate a sequence of ranking schemes;

[0124] It should be noted that the optimization objectives of candidate solutions in the multi-objective optimization solution set are ranked using Pareto hierarchical analysis and objective complementarity analysis. Specifically, the first-level non-dominated solution set is selected as the high-priority candidate. If there are multiple candidate solutions in the first-level non-dominated solution set, the complementarity of the optimization objectives of the candidate solutions is further analyzed (for example, a candidate solution with high clearance timeliness is better overall if it has a lower inspection anomaly rate). For the second-level and subsequent non-dominated solution sets, they are ranked according to the relative advantage of the optimization objectives (e.g., the higher the clearance timeliness, the lower the inspection anomaly rate, and the higher the historical declaration success rate, the higher the ranking of the candidate solution). Each candidate solution in the multi-objective optimization solution set is simulated and verified. Specifically, for each candidate solution, the changing trend of the corresponding optimization objective under historical extreme scenarios is extracted (e.g., during a promotional period, the clearance timeliness of a certain candidate solution is extended from the usual 2 days to 3 days, and the inspection anomaly rate increases from 5% to 8%), and the fluctuation range of the optimization objective is statistically analyzed. (e.g., a 50% increase in customs clearance timeliness), and the stability of candidate solutions is judged based on the fluctuation range. Specifically, if a candidate solution still meets the common conditions of historical successful declarations after fluctuations in the optimization objective under extreme scenarios (the inspection anomaly rate does not exceed the average upper limit of the anomaly rate of similar products in the same period), it is marked as high stability. If some target values ​​in the optimization objective after fluctuation are close to but do not break through the historical common conditions, it is marked as medium stability. If the optimization objective deviates from the historical common conditions after fluctuation (e.g., the inspection anomaly rate exceeds the average upper limit in the same period), it is marked as low stability. The multi-objective optimization solution set is sorted by combining the ranking results and the historical fluctuation stability markings. Specifically, high-stability candidate solutions in the first-level non-dominated solution set are selected first, followed by high-stability candidate solutions in the second-level non-dominated solution set, and so on. At the same time, within the same stability level, they are further sorted according to the objective advantages of the optimization objective (e.g., higher customs clearance timeliness); thus forming a ranking scheme sequence.

[0125] S5.3 Perform a priority truncation operation on the sorting scheme sequence to generate the optimal recommendation conclusion, and perform multi-source decision evidence association on the optimal recommendation conclusion to generate a preliminary source evidence chain;

[0126] It should be noted that the candidate solution with the highest priority in the sorting scheme sequence is taken as the optimal recommendation conclusion. The optimal solution is traced back to the optimization objective in the multi-objective optimization solution set, Pareto hierarchical results (such as belonging to the first level non-dominated solution set), stability verification markers (such as high stability), related data in historical declaration records (such as historical successful declaration cases corresponding to the current candidate solution), and constraints in the set of executable rules (such as the current candidate solution meeting the special commodity declaration rules), etc. The multi-source data is then structured and correlated to form a preliminary traceability evidence chain.

[0127] S5.4 Perform cross-modal information fusion and visualization structure encapsulation on the optimal recommendation conclusion and the preliminary source evidence chain, and output an intelligent classification scheme.

[0128] It should be noted that semantic alignment is used to unify the multi-source data in the traceability evidence chain to the declaration code in the optimal recommendation conclusion, and to perform structured integration to form complete data information of "declaration code + multi-dimensional target value + multi-source decision basis"; the complete data information is then encapsulated into a readable visual structure (such as structured JSON) and an intelligent classification scheme is output.

[0129] This embodiment also provides a computer device applicable to the intelligent classification method for import and export commodities based on knowledge graph metadata topology, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent classification method for import and export commodities based on knowledge graph metadata topology proposed in the above embodiment.

[0130] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0131] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the intelligent classification method for import and export commodities based on knowledge graph metadata topology as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0132] In summary, this invention achieves dynamic optimization of rule matching by using a dynamic interaction model to deeply mine the semantic associations of multimodal data, thereby improving the efficiency of rule adaptation in complex scenarios; and simultaneously performs multi-objective collaborative optimization to generate a sequence of ranking schemes, thereby improving the scenario adaptability of the schemes and the transparency and interpretability of the decisions.

[0133] 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, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent classification of import and export commodities based on knowledge graph metadata topology, characterized in that: include, Information on import and export commodities is collected and preprocessed. Multimodal feature fusion is then performed on the preprocessed information to generate an initial data packet. The initial data package is input into the dynamic interaction model, the semantic enhancement layer performs explicit association mining, the rule evolution layer optimizes the rule matching path, and a dynamic product knowledge graph is constructed. The specific steps are as follows: The initial data packet is input into the dynamic interaction model, and the semantic enhancement layer performs explicit association mining through remote supervision to form an attribute rule association matrix. The rule evolution layer uses a near-end policy optimization algorithm to optimize the rule matching path and generate an optimized rule set; The attribute rule association matrix and the optimized rule set are integrated through the feature splicing channel to construct a dynamic product knowledge graph; Perform topological structure derivation on the dynamic product knowledge graph to generate a graph topology analysis report and metadata list. The specific steps are as follows: Perform global topological metric analysis on the dynamic product knowledge graph to generate an original set of topological indicators; Perform community structure detection and hub node identification on the original topological indicator set to generate a graph structure insight report; Based on the graph structure insight report, perform pattern traversal on the dynamic product knowledge graph to generate a metadata list; The graph-based insight report performs pattern traversal on the dynamic product knowledge graph to generate a metadata list. The specific steps are as follows: Based on the community division results in the graph structure insight report, the dynamic product knowledge graph is divided into multiple structured functional blocks. Node relationship pattern traversal is then performed. Specifically, starting from the hub node, adjacent nodes are traversed along the hub node's associated edges, and node type combinations, relationship path patterns, and co-occurrence frequencies are recorded. Simultaneously, the metadata attributes of the graph nodes are extracted. Specifically, the data source and constraints of each graph node discovered during the traversal are recorded. The node relationship schema and metadata attributes are structured and organized to form a metadata list; Integrate the graph structure insight report and metadata list to generate a graph topology analysis report; The following steps are taken to integrate the topology analysis report and metadata list into a knowledge-based reasoning system to generate an intelligent navigation engine: Parse the data schema of the metadata manifest to generate entity relationship type specifications; Based on the graph topology analysis report, the structured business rules of the dynamic product knowledge graph are extracted and logically encoded to generate an executable rule set; Perform structured query transformation on entity relationship type specifications to generate a parameterized query instruction template library; The parameterized query instruction template library and executable rule set are logically coordinated and integrated to generate a rule and query coordinator. The rules and query coordinator are encapsulated and their protocols are adapted to generate a standardized classification engine. The standardized classification engine undergoes performance stress testing and stability optimization to generate an intelligent navigation engine. The intelligent navigation engine is invoked to perform multi-path semantic queries and rule validations on the dynamic knowledge graph, generating a set of candidate classification schemes. Multi-objective collaborative optimization is performed on the candidate classification scheme set to generate a sorting scheme sequence. The sorting scheme sequence is then encapsulated through source tracing to output an intelligent classification scheme.

2. The intelligent classification method for import and export commodities based on knowledge graph metadata topology as described in claim 1, characterized in that: The import and export commodity information includes text information, multimedia information, and historical declaration records; The preprocessing includes data cleaning, format conversion, deduplication, normalization, and outlier handling.

3. The intelligent classification method for import and export commodities based on knowledge graph metadata topology as described in claim 2, characterized in that: The specific steps for generating the initial data packet are as follows: Entity recognition and semantic encoding are performed on the preprocessed text information to generate product semantic vectors; The preprocessed multimedia information is subjected to structured parsing to generate image representation vectors; Historical pattern mining is performed on the preprocessed historical declaration records to generate declaration pattern vectors; The product semantic vector, image representation vector, and declaration pattern vector are concatenated to generate an initial data packet.

4. The intelligent classification method for import and export commodities based on knowledge graph metadata topology as described in claim 3, characterized in that: The specific construction steps of the dynamic interaction model are as follows: Call and initialize the fully connected network and multi-head Transformer to build the semantic enhancement layer and rule evolution layer; A dynamic interaction model is constructed by performing backpropagation and gradient update on the semantic enhancement layer and the rule evolution layer using the mean squared error loss function.

5. The intelligent classification method for import and export commodities based on knowledge graph metadata topology as described in claim 4, characterized in that: The specific steps for generating the candidate classification scheme set are as follows: The intelligent navigation engine is invoked to initiate multi-path semantic queries on the dynamic knowledge graph in parallel, generating an original candidate encoding set. Perform compliance checks on the original candidate coding set to generate a valid candidate coding set; Perform a multi-dimensional utility evaluation on the effective candidate coding set to generate a candidate classification scheme set.

6. The intelligent classification method for import and export commodities based on knowledge graph metadata topology as described in claim 5, characterized in that: The specific steps for the sorting scheme sequence are as follows: Perform Pareto front search and non-dominated sorting on the candidate classification scheme set to generate a multi-objective optimization solution set; Decision ranking and robustness verification for multi-objective optimization solution sets are performed, and a sequence of ranking schemes is generated.

7. The intelligent classification method for import and export commodities based on knowledge graph metadata topology as described in claim 6, characterized in that: The specific steps of the intelligent classification scheme are as follows: A priority truncation operation is performed on the sorting scheme sequence to generate the optimal recommendation conclusion, and a multi-source decision evidence association is performed on the optimal recommendation conclusion to generate a preliminary source evidence chain; The optimal recommendation conclusion and the preliminary source evidence chain are subjected to cross-modal information fusion and visualization structure encapsulation to output an intelligent classification scheme.

Images