Large-scale article classification method and system

Abstract

The invention discloses a large-scale article classification method and system, and relates to the field of intelligent supervision, and the method comprises the steps: extracting a visual feature vector, a semantic embedding vector and an attribute field vector from an original article set, carrying out the splicing fusion, obtaining a multi-modal vector, carrying out the dimension reduction, and obtaining a low-dimensional potential vector; clustering the low-dimensional potential vectors to obtain a hierarchical label of each article; constructing a reference name library; calculating the overlapping ratio of the structured attribute feature set Fp corresponding to the reference model and the structured attribute feature set Fc of the clustering cluster, and determining a category model matched with the clustering cluster; generating candidate names; checking whether the candidate names exist in a reference name library or not, and if yes, calculating a candidate name list closest to the candidate names for selection; and if the candidate name does not exist, directly using the candidate name to obtain a classification result of each article. According to the method and the device, structured classification and standardized naming of mass article data can be realized on the premise of low manual dependence.

Description

Technical Field

[0001] The present application relates to the field of intelligent supervision, and in particular to a method and system for large-scale object classification. Background Art

[0002] Equipment cataloging is a fundamental and prerequisite step in the development and management of equipment data resources. Due to the inefficiency and difficulty of scalability of traditional cataloging methods, as well as the bottlenecks of existing methods such as manual processing, threshold matching, and supervised learning, the cataloging process is plagued by efficiency bottlenecks, high misclassification rates, delayed updates to classification standards, inconsistent coding standards across systems, and poor generalization of automated methods. These issues lead to delayed resource allocation, disorganized inventory information, dramatically increased manual maintenance costs, and the inability to trace audit responsibilities.

[0003] With the emergence of massive amounts of heterogeneous equipment data in fields such as the Internet of Things, smart manufacturing, and defense logistics, traditional equipment cataloging methods, which rely primarily on manual operations, fixed rule matching, or supervised models, face increasing efficiency bottlenecks and limited scalability. These methods are ill-suited to today's large-scale, heterogeneous, and multi-source equipment data environments. Manual cataloging requires professionals to compare item names and attributes one by one. This is not only inefficient when faced with massive amounts of data from diverse sources and emergencies, but also prone to misjudgment and omissions. Relevant research indicates that in a sample from the medical device industry, manual cataloging methods spend over 80% of their work time searching for similar records and verifying standards, rather than actual data entry. Furthermore, equipment cataloging must strictly adhere to GBT 37936-2019 (Information Classification and Coding Requirements), GBT 37944-2019 (Data Model Compilation Requirements), and GBT 37948-2019 (Data Element Compilation Requirements). These standards clearly define requirements for material classification and coding, the construction of reference names, and the compilation of data models and data elements. Strictly implementing multiple standards under manual rules further increases the complexity and difficulty of the work.

[0004] Matching methods based on preset rules also have significant limitations. Rule sets struggle to cover all complex and varied descriptions. Once new, undefined expressions or data types emerge, rules must be added or adjusted, causing cataloging to lag behind data updates. In the classification of manufacturing parts, TF-IDF + cosine similarity can achieve 85% automatic classification accuracy, but the authors point out that support for new models and low-frequency terms is insufficient, and recall drops by over 20% in long-tail categories. In bioinformatics, PLOS ONE proposes a threshold determination method for minimizing the false positive / false negative ratio, which can provide an optimal threshold selection reference for any semantic similarity task. For data directory structure matching, Sheen's similarity flooding iterative algorithm can globally align pattern elements at the graph structure level, but relies on manual threshold setting for text attribute similarity, making it difficult to autonomously expand to new nodes.

[0005] Existing automated cataloging methods also have shortcomings: simple threshold matching algorithms lack dynamic adaptability and are difficult to adjust in real time as data changes; Joachims (1998) first proved that SVM can significantly outperform traditional KNN and naive Bayes in text classification tasks, is robust to high-dimensional sparse features, and is widely used in product title classification and document cataloging scenarios; Dalal and Zaveri (2011) reviewed automatic text classification technology, pointing out that supervised models lack generalization capabilities in cold start and long-tail categories, and need to be supplemented by incremental annotation and transfer learning strategies to cope with dynamic category expansion; recently, the hybrid neural network architecture of GCNN+BiLSTM+attention has surpassed the BERT baseline in short text classification, but still relies on large-scale high-quality annotated corpus and lacks generalization efforts for unseen categories. At the same time, its "black box" decision-making process is difficult to explain and lacks transparent interpretability, which in turn affects the credibility of the results.

[0006] Although the existing methods have their own advantages, overall, due to the lack of mature systematic solutions to support the classification and cataloging needs of "large-scale, large-scale, dynamic item classification", a more scalable, intelligent and interpretable new method and system is urgently needed in the cataloging process to meet the growing challenges of equipment data management. These problems have jointly led to a series of negative consequences in the cataloging process, such as delayed resource allocation, information redundancy conflicts, frequent coding conflicts, sharp increases in manual maintenance costs, and difficulty in holding people accountable for errors, which have seriously restricted the standardized management and efficient utilization of equipment data. To this end, this application proposes an automatic classification method system with "cluster analysis + multimodal semantic matching + attribute recommendation" as the core. It can achieve structured classification and standardized naming of massive item data under the premise of low manual dependence, has strong generalization ability and good structural interpretability, and aims to build a highly adaptable intelligent system for future complex cataloging tasks. Summary of the Invention

[0007] The purpose of this application is to provide a large-scale item classification method and system, which can achieve structured classification and standardized naming of massive item data with low manual dependence, has strong generalization ability and good structural interpretability, and aims to build a highly adaptable intelligent system for future complex cataloging tasks.

[0008] To achieve the above objectives, this application provides the following solutions: In a first aspect, the present application provides a method for large-scale object classification, comprising: Obtaining an original item set; the original item set includes: images, text descriptions, and structured attribute fields; Extracting visual feature vectors, semantic embedding vectors, and attribute field vectors from the original item set; Concatenating and fusing the visual feature vector, the semantic embedding vector, and the attribute field vector to obtain a multimodal vector of the object; Performing dimensionality reduction on the multimodal vector to obtain a low-dimensional latent vector; The low-dimensional potential vector is clustered into the following categories: K (1) Large clusters; For each large cluster, secondary clustering is performed again and subdivided into K (2) small clusters to obtain the hierarchical label of each item; the large clusters and small clusters constitute clusters; Constructing a benchmark name library; the benchmark name library includes benchmark models and structured attribute feature sets corresponding to the benchmark models F p ; Get the structured attribute feature set of the cluster F c ; Compute the structured attribute feature set corresponding to the baseline model F p and structured attribute feature sets of clusters F c and determining a category model that matches the cluster based on the overlap rate; the category model includes: a reuse model, an extended model, and a new model; For clusters that require new models, candidate names are generated based on the characteristic content or domain terms of the clusters; Check whether the candidate name already exists in the benchmark name library. If the candidate name already exists, it indicates a naming conflict. The system will prompt a name conflict and use the edit distance or synonym library to calculate a list of candidate names closest to the candidate name for selection; if the candidate name does not exist, directly use this candidate name, assign it a new unique ID and register it in the library to obtain the classification result of each item.

[0009] Optionally, the large-scale object classification method further comprises, after the last step: The classification result of each item is represented as a tuple.

[0010] Optionally, the following formula is specifically used to extract the visual feature vector, the semantic embedding vector, and the attribute field vector from the original item set: ; in, is the visual feature vector, is the semantic embedding vector, is the attribute field vector, For images, is a text description, is a structured attribute field, is the visual feature space, is the text semantic embedding space, Encoding space for structured attributes.

[0011] Optionally, the visual feature vector, semantic embedding vector, and attribute field vector are concatenated and fused to obtain a multimodal vector of the item using the following formula: ; in, is the vector splicing symbol, is the visual feature vector, is the semantic embedding vector, is the attribute field vector, is the Euclidean space where the three types of vectors are located, For the i A high-dimensional multimodal representation vector of an item, is the visual sub-vector dimension; is the semantic sub-vector dimension; is the attribute sub-vector dimension; is the sum of the dimensions of the three types of sub-vectors.

[0012] Optionally, the multimodal vector is subjected to dimensionality reduction to obtain a low-dimensional latent vector. Specifically, an autoencoder is introduced to reduce the dimensionality of the multimodal vector. The specific formula is as follows: coding: ; decoding: ; loss: ; in, is a low-dimensional latent vector; Represents the encoder network Enc (Encoder) with parameters For control, input sample , outputs a latent space vector , represents the compressed representation of the sample; is the potential mean vector output by the encoder; For the i A high-dimensional multimodal representation vector for each item; represents the standard deviation vector of the latent distribution of the encoder output, which is uncertainty scale mapped to the latent space; is a noise vector that obeys the standard normal distribution; represents a random vector sampled from a standard normal distribution; is the representation vector reconstructed by the decoder, if the result is the same as The closer they are, the more The more information you have; represents the decoder network, with parameters control, is the abbreviation of decoder function; and Represent the learnable parameters of the encoder and decoder respectively; is the total loss function of the variational autoencoder; N is the total number of item samples in the batch; i is the sample index number; is the square of the Euclidean distance; is the balance coefficient; is the Kullback–Leibler divergence; is the approximate posterior distribution; is the prior distribution.

[0013] Optionally, the low-dimensional potential vector is clustered into the following categories: K (1) The following formula is used for each major cluster: ; in, The number of preset categories; For the j A collection of sample indexes of major categories; is a low-dimensional latent vector, For large categories The center of mass; is the square of the Euclidean distance; For each large cluster, secondary clustering is performed again and subdivided into K (2) The following formula is used to obtain the hierarchical label of each item: ; in, Preset the number of sub-category clusters to be subdivided within each major category; is a small cluster sample set, k is the subcategory index number, For the In the major categories The centroid of each small cluster; The hierarchical label of each item is ,in, For the i The category labels to which the items are assigned in the first-level cluster; It is the subdivision label of the same item in the secondary cluster.

[0014] Optionally, for a cluster that requires a new model, the following formula is used to generate candidate names based on the characteristic content or domain terms of the cluster: ; in, is the cosine similarity; is the secondary cluster center; is the name vector; is a candidate name; To select the item with the greatest similarity among all candidate models, arg Indicates the value of the variable that corresponds to the maximum (or minimum) value of an expression.

[0015] Optionally, calculate the structured attribute feature set corresponding to the baseline model F p and structured attribute feature sets of clusters F c The overlap rate of the clusters is determined based on the overlap rate, and the category model that matches the cluster cluster is determined using the following formula: ; in, for F p and F c The overlap rate of When , a reusable model is recommended; when When , the extended model is recommended; when It is recommended to create a new model.

[0016] Optionally, the classification result of each item is represented as a tuple as follows: ; in, It is the item identifier. For the i The category labels to which the items are assigned in the first-level cluster; is the subdivision label of the same item in the secondary cluster, is the finalized benchmark name. Indicates the matching status of the category with the benchmark model, including reuse, extension or creation. It is the referenced or newly created attribute model. If it is newly created, it is the new model ID. The overlap ratio calculated during matching.

[0017] In a second aspect, the present application provides a large-scale item classification system, the large-scale item classification system comprising: An original item set acquisition module is used to acquire an original item set; the original item set includes: images, text descriptions, and structured attribute fields; A vector extraction module, configured to extract visual feature vectors, semantic embedding vectors, and attribute field vectors from the original item set; A splicing and fusion module, configured to splice and fuse the visual feature vector, the semantic embedding vector, and the attribute field vector to obtain a multimodal vector of the object; A dimensionality reduction module, configured to reduce the dimensionality of the multimodal vector to obtain a low-dimensional latent vector; The first-level clustering module is used to perform first-level clustering on the low-dimensional potential vector, which is roughly divided into K (1) Large clusters; The secondary clustering module is used to perform secondary clustering on each large cluster, which is further subdivided into K (2) small clusters to obtain the hierarchical label of each item; the large clusters and small clusters constitute clusters; A base name library construction module is used to construct a base name library; the base name library includes a base model and a structured attribute feature set corresponding to the base model F p ; A clustering cluster structured attribute feature set acquisition module, used to obtain the clustering cluster structured attribute feature set F c ; Overlap rate calculation module, used to calculate the structured attribute feature set corresponding to the benchmark modelF p and structured attribute feature sets of clusters F c and determining a category model that matches the cluster based on the overlap rate; the category model includes: a reuse model, an extended model, and a new model; A candidate name determination module is used to generate candidate names for clusters that require a new model based on the characteristic content or domain terms of the clusters; The retrieval module is used to check whether the candidate name already exists in the benchmark name library. If the candidate name already exists, it indicates a naming conflict. The system will prompt the name conflict and use the edit distance or synonym library to calculate a list of candidate names closest to the candidate name for selection; if the candidate name does not exist, the candidate name will be directly used, a new unique ID will be assigned to it and registered in the library to obtain the classification result of each item.

[0018] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method and system for large-scale object classification. First, this application can solve the problem of strong subjectivity and high error rate of manual classification. It adopts two-level clustering + cosine semantic matching and field overlap rate judgment to provide quantitative classification for each category. and indicators, improving overall classification accuracy; second, this application can solve the problem of low manual processing efficiency, and end-to-end batch processing of ≥10,000 items only requires one offline pipeline run. Compared with traditional manual entry one by one, the classification speed is increased by dozens of times, achieving high efficiency. Third, this application can solve the problem of lagging updates in classification and coding standards, and automatically connect the old and new standards through the "reuse / expansion / new construction" three-state model decision; at the same time, the name uniqueness check is introduced to ensure cross-system coding consistency; fourth, this application can solve the problem of poor generalization ability and difficulty in identifying new categories of existing methods, and use unsupervised clustering to first discover potential new clusters, and then automatically generate attribute drafts to achieve dynamic expansion capabilities for unknown items; fifth, this application can solve the problem of lack of interpretability of black box models, output cluster centers, similarity heat maps and field difference lists, forming an auditable and traceable explanation chain, and enhancing the credibility of the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 A flowchart of a large-scale object classification method provided in one embodiment of the present application; Figure 2 A flowchart of a large-scale item classification method provided in one embodiment of the present application; Figure 3 A structural diagram of a large-scale item classification system provided in one embodiment of the present application. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0022] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0023] In an exemplary embodiment, Figure 1 and Figure 2 As shown, a large-scale item classification is provided. Figure 2 The system begins with multimodal item data input and, through feature processing, dimensionality reduction, and cluster analysis, forms preliminary categories. Based on this, it compares each cluster against a library of benchmark category models to determine whether it corresponds to an existing standard category model. For clusters that match a highly similar existing category, the system directly adopts the corresponding benchmark name and reuses its attribute model. If there is only a partial match, the system triggers the "suggest expansion" process, recommending adjustments to the existing model or the addition of new attribute fields to complete the category definition. For clusters that do not match any existing categories, the system identifies a new category need and enters the "new model generation" process, which involves automatically extracting shared attributes from the clusters to construct candidate attribute models and recommending a name for the new category. The system then verifies the uniqueness of the new name and resolves naming conflicts (suggesting alternative names if duplicates exist). Finally, the system outputs a result file or data table containing information about each item's category and its benchmark name / model, and updates the internal knowledge base to support subsequent classification tasks. This entire process automatically converts data into information, reducing manual intervention while ensuring the scalability and standardization of the classification system.

[0024] See also Figure 1 The large-scale object classification method in this application includes the following steps 201 to 208. Among them: Step 101: Acquire an original item set; the original item set includes: images, text descriptions, and structured attribute fields.

[0025] Suppose the original item set is: ; in, Is the item serial number, the value range is , each Corresponding to an original item record, For images, is a text description, It is a structured attribute field.

[0026] The raw item data is then collected and cleaned. Specifically, the input is the original item record, which includes images, text descriptions, and structured attributes, and the output is a standardized data format. For example, images are resized and noise removed, text is segmented, stop words are removed, and encoding is performed, and missing structured fields are filled in or marked to ensure the input data quality for subsequent processing.

[0027] Step 102: extracting visual feature vectors, semantic embedding vectors, and attribute field vectors from the original item set.

[0028] Three types of deep network models are used to extract image, text and attribute features respectively, and the feature vectors are expressed as: ; in, is the visual feature vector: the object image is transformed by the visual Transformer or CNN Extract,single-dimensional response image local texture, edge direction or high-order appearance semantics; Semantic Embedding Vectors: Contrastive Distillation-BERT for Text Description Encoding is obtained, and geometric distance describes the similarity of texts in semantic space; For attribute field vector: the structured attribute The numeric fields are normalized, the categorical fields are uniquely hot / embedded, and the dimensions correspond to the real physical or engineering attribute scales. , but they are completely different in feature sources, physical interpretations, and scale distributions. Subsequently, through splicing and VAE compression, these three signals are fused into a unified "multimodal representation".

[0029] Step 103: Concatenate and fuse the visual feature vector, semantic embedding vector, and attribute field vector to obtain a multimodal vector of the object.

[0030] After obtaining three complementary and overlapping vectors of image, text, and attributes, they are directly concatenated in the column direction to generate a multi-modal representation of the item. It refers to an "integrated vector label" that can simultaneously depict the appearance, text semantics, and engineering attributes of an item, and is expressed as: ; Among them, is the vector concatenation symbol, which sequentially connects three sub-vectors in the column direction; is the multi-modal representation vector, an integrated label that simultaneously carries information in three aspects of vision, semantics, and attributes, and is used for subsequent dimensionality reduction, clustering, and semantic retrieval. is the dimension of the visual sub-vector, which is determined by the visual network architecture and the feature extraction layer; is the dimension of the semantic sub-vector, which is determined by the size of the hidden layer; is the dimension of the attribute sub-vector, which is determined by the number of attribute fields and the coding method; is the sum of the dimensions of the three types of sub-vectors, and this dimension is compressed to the latent dimension in the dimensionality reduction coding .

[0031] Step 104: Perform dimensionality reduction on the multi-modal vector to obtain a low-dimensional latent vector.

[0032] Multi-modal representation The dimension can reach up to hundreds at most, which is not conducive to calculation and storage. Therefore, a variational autoencoder is introduced, and its purpose is to compress the high-dimensional multi-modal representation into the latent space (k << d), so that has good properties of "continuous, interpolable, and probability measurable". While ensuring information recoverability, learn the non-linear mapping from high dimension to low dimension, and its training objective is: Encoding: ; Decoding: ; Loss: ; Among them, is the low-dimensional latent representation, which is used as the basis for clustering in the subsequent process; represents that the encoder network Enc (Encoder) takes the parameter as the control, inputs the sample , and outputs a latent space vector , representing the compressed representation of the sample; is the latent mean vector output by the encoder; is the i th high-dimensional multi-modal representation vector of the item; represents the standard deviation vector of the latent distribution of the encoder output, which is uncertainty scale mapped to the latent space; is a noise vector that obeys the standard normal distribution; represents a random vector sampled from a standard normal distribution; represents the decoder network, with parameters control; Is the representation vector reconstructed by the decoder, if the result is the same as The closer they are, the more The more information you have; and They represent the learnable parameters of the encoder and decoder, respectively, which are jointly optimized by gradient descent during the training phase; To achieve differentiable random sampling for noise vectors that obey the standard normal distribution; is an approximate posterior distribution, used to characterize uncertainty; is the prior distribution to ensure consistent sampling of the latent space; is the Kullback–Leibler divergence, which is used to measure the difference between two distributions; is the balance coefficient.

[0033] Step 105: Perform first-level clustering on the low-dimensional potential vector, roughly dividing it into K (1) A large cluster.

[0034] First-level clustering (large category): Use improved K-means to cluster all potential vectors Roughly marked major categories , such as electrical appliances, machinery, textiles, etc. Its optimization goals are: ; in, The number of preset categories; is the sample index set of the jth category, all categories are pairwise disjoint and their union is the complete set; For large categories The center of mass; is the square of the Euclidean distance, which is used to measure the closeness of the sample to the centroid.

[0035] Among them, the explanation of the intersection and union of all major categories is as follows: In the latent space, clustering multimodal vectors can form several large categories (clusters). To include all The sample space of samples. The clustering result will be Divided into major categories, corresponding to Sample subset (Right now ), For the j The sample index set of each major category, formally speaking, the above subsets meet the following two conditions: Pairwise disjointness: For any , both In other words, no single sample belongs to two different subsets at the same time.

[0036] Combine into a complete set: That is, the elements of these subsets add up to cover the entire set , there are no missed samples.

[0037] Satisfying the above properties means that each sample falls into a certain subset It is neither repeated in multiple subsets nor omitted from all subsets. Therefore, the subset set Constitutes the complete sample set A division of .

[0038] For example, assuming the total sample set , which is divided into Class, desirable , , Obviously They do not intersect with each other, and Therefore, each sample belongs to only one of the above three subsets, and all five samples are included in the divided subsets.

[0039] Step 106: Perform secondary clustering again for each large cluster, and subdivide it into K (2) Small clusters are formed to obtain the hierarchical labels of each item; the large clusters and small clusters constitute clusters.

[0040] Secondary clustering (small categories): K-means is performed again within each large category. Subdivided into Small categories , the objective function is the same as above but constrained to local fine-grained clusters, and the optimization goal is: ; in, For each major category, preset the number of sub-category clusters to be subdivided. For example, if the "major category = charger" is divided into 5 sub-categories, then =5; is a small cluster sample set, k is the subcategory index number, For the In the major categories The centroid (mean vector) of each small cluster; The hierarchical label of each item is ,in, For the i The category labels to which items are assigned in the first-level cluster; Subdivision labels for the same item in the secondary cluster To this end, each item gets a pair of hierarchical labels , which lays the structural foundation for subsequent name and attribute matching and can effectively reduce the problem of long-tail categories being neglected.

[0041] Step 107: Construct a base name library; the base name library includes base models and structured attribute feature sets corresponding to the base models. F p .

[0042] The system presets a set of manually reviewed benchmark category names as a standard name library ; each Uniquely corresponds to an industry or national standard category, Indicates the name library capacity, that is, the total number of registered base names.

[0043] Step 108: Obtain the structured attribute feature set of the cluster F c .

[0044] Step 109: Calculate the structured attribute feature set corresponding to the baseline model F p and structured attribute feature sets of clusters F c and determining a category model that matches the cluster based on the overlap rate; the category model includes: a reuse model, an extended model, and a new model.

[0045] Step 110: For the clusters for which a new model is required, candidate names are generated based on the characteristic content or domain terms of the clusters.

[0046] Indicates the name The goal is to embed the semantic embedding vector of the benchmark name (a short text / term) is mapped into a k-dimensional real vector so that it can express the semantics of the name in the vector space and can be compared with the cluster center Do similarity comparison. Secondary clustering gets several small classes, each of which is recorded as , for the secondary cluster centers with name vector Artificial cosine similarity , find the largest As a candidate benchmark name, the calculation process is expressed as: ; like Then directly use the cosine similarity The biggest name As a subcategory Otherwise, it is marked as "pending new name". This step quickly aligns the unsupervised clustering results to the existing semantic system and promptly discovers potential new categories.

[0047] Step 111: Check whether the candidate name already exists in the benchmark name library. If the candidate name already exists, it indicates a naming conflict. The system will prompt a name conflict and use the edit distance or synonym library to calculate a list of candidate names closest to the candidate name for selection. If the candidate name does not exist, this candidate name can be used directly, assigned a new unique ID and registered in the library to obtain the classification result of each item.

[0048] Candidate benchmark names Calling the hash index: If the name is already occupied by other categories in the standard name library, conflict resolution is triggered, and a list of alternative names is generated using the edit distance and semantic distance in turn; if there is no occupied category, a globally unique ID is immediately assigned, and a batch of cluster labels and attribute model association records generated by the new name are recorded. Based on the threshold rules, it is decided to reuse, expand, or create a new model to ensure that the attribute structure is aligned with the standard system. For example, the process of finding the most similar old template for a new item: if the similarity is more than 80%, the old template is directly used; if the similarity is 60-80%, adjustments are made based on the old template; if the similarity is less than 60%, the new content must be revised and included in the template. With the model library Model Field Set Calculate overlap ratio Expressed as: ; when Consider reusable models ; As an extended scenario, Add fields based on Determine whether a new model is needed This mechanism ensures that the attribute structure inherits history while covering new item characteristics.

[0049] Finally, the classification result of each item is represented as a tuple, specifically as follows: ; in, Item ID, used to uniquely locate the item in the database; For the i The category labels to which the items are assigned in the first-level cluster; is the subdivision label of the same item in the secondary cluster; The final benchmark name is aligned with the industry standard database; The result status of the name-model matching indicates how the item (or its sub-category cluster) handles the attribute model. The possible values ​​are "reuse", "extend", and "create". The ID of the referenced attribute model: if match_status=Reuse, it points to the existing model; if it is "Extend", it points to the derived model; if it is "New", it is the new model number generated by the system for the cluster; is the overlap ratio, which takes values ​​of [0,1] and can quantify the degree of fit between the model and the actual data. Parquet data sets are written in batches through Spark and pushed to the review end, achieving a fast closed loop of "algorithm clustering first, expert confirmation later".

[0050] Based on the same inventive concept, embodiments of the present application also provide a large-scale item classification system for implementing the aforementioned large-scale item classification method. The solution provided by this system is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the large-scale item classification system provided below can be found in the above-mentioned limitations of the large-scale item classification method and will not be further elaborated here.

[0051] In an exemplary embodiment, Figure 3 As shown, a large-scale item classification system is provided, which includes: an original item set acquisition module, a vector extraction module, a splicing and fusion module, a dimensionality reduction module, a first-level clustering module, a second-level clustering module, a reference name library construction module, a structured attribute feature set acquisition module for clusters, an overlap rate calculation module, a candidate name determination module, and a retrieval module; The original item set acquisition module is used to obtain an original item set, including images, text descriptions, and structured attribute fields. It is also used to collect and clean original item data. The input is the original item record, which includes images, text descriptions, and structured attributes, and the output is a standardized data format. For example, images are resized and noise removed; text is segmented, stop-word removed, and encoded; missing structured fields are filled in or marked, ensuring the quality of input data for subsequent processing.

[0052] The vector extraction module is used to extract visual feature vectors, semantic embedding vectors, and attribute field vectors from the original object set. This module includes an image feature extraction submodule, a text feature extraction submodule, and a structured attribute encoding submodule. The input is the preprocessed image. , text description and attribute fields , the output is the corresponding image feature vector , text feature vector and attribute feature vector The output of this module can be used as the pre-training model result.

[0053] The splicing and fusion module is used to splice and fuse the visual feature vector, semantic embedding vector and attribute field vector to obtain the multimodal vector of the object. The input of this module is the image, text and attribute feature vector of the same object. The above vectors are spliced ​​and fused according to the specified dimension and the comprehensive feature vector of the object is output. ,in For example, let the total number of items be , No. The image features of an item are , the text features are , the attribute characteristics are , then: ; In the above formula, ” represents the vector concatenation operation. After fusion, the output vector of this module contains multiple aspects of the item’s visual, semantic, and attribute information, serving as the representation space for dimensionality reduction and clustering.

[0054] The dimensionality reduction module is used to reduce the dimensionality of the multimodal vector to obtain a low-dimensional latent vector. The input of this module is the fused feature vector , the output is a low-dimensional compact representation The autoencoder contains the encoder and decoder Two parts, through unsupervised learning Map to The low-dimensional latent space of , while reconstructing the original input as much as possible. The parameters of the encoder are , the encoding process is: , decoded and reconstructed into The training goal of the autoencoder is to minimize the reconstruction error: ; By minimizing the above loss function, the model can learn a low-dimensional representation of multimodal data. , effectively remove redundant information. In practical applications, the appropriate dimensionality reduction dimension can be selected based on experience. (For example, the original Dimension compression to dimensional) to balance information fidelity and computational cost. The output of the dimensionality reduction encoding module is a low-dimensional representation of each item , these representations will be used for cluster analysis.

[0055] The first-level clustering module is used to perform first-level clustering on the low-dimensional potential vector, which is roughly divided into K (1) A large cluster.

[0056] The secondary clustering module is used to perform secondary clustering on each large cluster, which is subdivided into K (2) Small clusters are formed to obtain the hierarchical labels of each item; the large clusters and small clusters constitute clusters.

[0057] The above-mentioned first-level clustering module and second-level clustering module are used to perform unsupervised cluster analysis on the item representation after dimensionality reduction, so as to automatically classify similar items. Specifically, a two-level clustering architecture is adopted, firstly clustering the "big category" and then further subdividing the subcategories. The input is the set of representation vectors of all items. , output each item to which the major category and minor category. First, the first level clustering (major category) stage uses -means algorithm will all The vector of items is divided into Initial clusters; in the subsequent secondary clustering (small class) stage, samples within each large cluster continue to use -means is subdivided into several subclusters. The output of the clustering module includes the clustering results of each item: major category label and subcategory labels , the output of this module It can be broken down into numbers under the corresponding major categories.

[0058] The base name library construction module is used to construct a base name library; the base name library includes a base model and a structured attribute feature set corresponding to the base model F p .

[0059] The structured attribute feature set acquisition module of the cluster is used to acquire the structured attribute feature set of the cluster F c .

[0060] The overlap calculation module is used to calculate the structured attribute feature set corresponding to the benchmark model F p and structured attribute feature sets of clusters F cand determining a category model that matches the cluster based on the overlap rate; the category model includes: a reuse model, an extended model, and a new model.

[0061] Specifically, first obtain the structured attribute feature set of the cluster and the system's pre-existing benchmark attribute model library For comparison, each benchmark model A set of structured fields is predefined , corresponding to the attribute template of a standard item category. By calculating the overlap rate between the cluster attribute set and the model attribute set , which measures the similarity between the cluster and each known model. Then, according to The value is determined by the set threshold rule: when there is When the model is used, it is considered that the current cluster is highly consistent with an existing category model, and the matching model is output. , that is, it is recommended to reuse the existing model; when the highest overlap rate falls on When the cluster characteristics are in the interval, it is considered that the cluster characteristics are partially consistent with the existing model, and the output is the closest model. and difference fields, it is recommended to modify and expand the model based on this; if all , the cluster is determined to have failed to match a suitable existing model, and a new model needs to be created to define the new category of the cluster. In the new model creation scenario, the system automatically counts the most common important attribute fields of the cluster items and forms a preliminary draft attribute structure for subsequent manual confirmation and improvement.

[0062] The candidate name determination module is used to generate candidate names for clusters that need to be newly modeled based on the characteristic content or domain terms of the clusters. The input of this module is the category cluster to be named and its matching status, and the output is the recommended name and unique identifier of the category. For clusters that can be directly matched to existing models, the standard name of the model is directly inherited. As the category name; for the category cluster that needs to be newly modeled, the system will generate a candidate name based on the cluster's characteristic content or domain terminology For example, based on the most frequent keywords in the cluster, or temporarily named "New Benchmark", etc. The system then searches the candidate name in the benchmark name library: if the name If it already exists, it means naming conflict, the system will prompt the name conflict and use edit distance or synonym library to calculate the conflict. A list of similar candidate names to choose from; if the name If it does not exist, the name can be used directly, a new unique ID can be assigned to it and registered in the database. This mechanism ensures the standardization and uniqueness of the newly added category name and avoids confusion with existing categories.

[0063] The retrieval module is used to check whether the candidate name already exists in the benchmark name library. If the candidate name already exists, it indicates a naming conflict. The system will prompt the name conflict and use the edit distance or synonym library to calculate a list of candidate names closest to the candidate name for selection; if the candidate name does not exist, this candidate name can be used directly, assigned a new unique ID and registered in the library to obtain the classification result of each item.

[0064] The following is a clustering example based on the medical supplies field, which clearly shows how the system can complete a large-scale item classification task at one time through clustering: This case study, based on the medical supplies industry, addresses the need for standardized classification of medical devices and their components. A multimodal clustering classification system capable of processing over 10,000 records per batch was built. The system's core goal was to address the challenges of diverse medical product categories, inconsistent naming, and ambiguous definitions despite high structural similarity. This system enabled standardized classification, archiving of baseline names, and the generation of attribute models.

[0065] The data source includes a set of 12,000 medical parts samples, covering a variety of typical categories, including syringes, respiratory masks, infusion sets, and surgical forceps. Each sample data contains three dimensions: image (product photo), text (product name and description), and structured parameter fields (such as material, model, caliber, connector standard, etc.). The system first uses ResNet to extract 256-dimensional visual embeddings for the images. The text is then processed using the BERT model to extract 768-dimensional semantic vectors. The attribute fields are then normalized to form a 64-dimensional structured field vector. These three are then concatenated into a unified 1088-dimensional multimodal feature vector for subsequent clustering and classification tasks.

[0066] The system uses an autoencoder to compress and reduce the dimensionality of multimodal features to a 64-dimensional latent space, and then uses a two-level KMeans clustering algorithm to group items. The first-level clustering divides items into 15 broad categories, such as "respiratory equipment," "injection parts," and "surgical aids." The second-level clustering further divides each broad category into subcategories, resulting in a total of 80 subcategorized clusters. The center vector of each subcategory represents the semantic features of typical items in that category, serving as a key reference for subsequent name and model matching.

[0067] During the baseline name recommendation phase, the system introduces a standard name knowledge base containing existing names, aliases, and their embedded representations. By calculating the semantic cosine similarity between the center vector of each small cluster and the baseline name vector, the system can determine whether there are highly similar matches. If the match is higher than 0.85, the system recommends retaining the original name; if it is lower than the threshold, the system automatically recommends three highly similar reference names for manual selection and generates a suggested new baseline name and unique code. This mechanism effectively solves the problem of inconsistent names and non-standard expressions of medical items in multiple scenarios.

[0068] The system also counts the common structured fields within each cluster (such as ventilation resistance, interface diameter, and tightness level) and compares them with existing attribute models. If the overlap rate of structure fields exceeds 85%, the system recommends reusing the existing model. If it falls between 65% and 85%, the system suggests adding or modifying fields based on a specific template. If the overlap rate is low, the system automatically summarizes the fields and outputs a draft structure for a "suggested new model." This ensures both structural continuity and scalability of the new model.

[0069] Ultimately, the system integrates the classification suggestions, baseline name recommendations, and attribute model comparison results for each cluster into an "audit package" and submits it in a structured format to the manual review platform. Reviewers can confirm or correct naming suggestions, structural fields, and coding conflicts. The corrections are written to the official database and used for subsequent model optimization. By periodically feeding manual feedback results back into the training data, the system achieves adaptive optimization capabilities, ensuring that the model continues to evolve as the data scale grows and adapts to the ever-expanding medical category system.

[0070] Database table structure design (MySQL) Table 1 Category table – classification level (major category / minor category)

[0071] Table 2 BenchmarkName table – Benchmark name and unique code

[0072] Table 3 Alias ​​table – name alias management

[0073] Table 4 AttributeModel table – attribute model structure

[0074] Table 5 ClusterSuggestion table – clustering results and suggestion records

[0075] Recommendation system evaluation metrics: In addition to commonly used indicators such as precision, recall, F1, and manual review acceptance rate, a comprehensive multi-dimensional evaluation should be conducted: 1. Multi-objective ranking quality: NDCG (Normalized Discounted Cumulative Gain) and MAP (Average Average Precision) are used to measure the rationality and accuracy of the ranking of recommendation results in the list.

[0076] 2. Coverage: Recommendation coverage is defined as the ratio of recommended items to the total number of items. A high coverage means that the recommendation system can serve a wider user group, reducing the situation where users feel that the system does not respond to their needs, thereby improving user satisfaction and the overall utility of the system. This indicator reflects the system's exposure ability for long-tail products.

[0077] 3. Diversity: This measures the extent to which recommendations are drawn from diverse categories or topics. Recommendation systems can easily lead to the clustering of similar content, resulting in bland content, a narrow range of information access, and a "information cocoon." This can also lead to poor user experience due to the lack of clear or broad interests, and the inability to generate recommendations quickly. High diversity helps uncover potential user interests and avoids a single recommendation result.

[0078] 4. Novelty: Recommendation systems must focus on recommending items to users they haven't encountered before, stimulating their interest in exploration by expanding their cognitive boundaries. The system must dynamically optimize its recommendation strategy while maintaining accuracy, achieving a balance between novelty and precision. Novelty also requires that the item cataloging system possess dynamic update capabilities and continuously optimize the system's item cataloging: regularly integrating and categorizing low-frequency matching items, recommending the generation of cataloging base names or attribute models that meet data standard requirements and classification patterns for new items, and continuously adjusting and optimizing its re-clustering algorithm.

[0079] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for large-scale object classification, characterized by: The large-scale object classification method includes: Obtaining an original item set; the original item set includes: images, text descriptions, and structured attribute fields; Extracting visual feature vectors, semantic embedding vectors, and attribute field vectors from the original item set; Concatenating and fusing the visual feature vector, the semantic embedding vector, and the attribute field vector to obtain a multimodal vector of the object; Performing dimensionality reduction on the multimodal vector to obtain a low-dimensional latent vector; The low-dimensional potential vector is clustered into the following categories: K (1) Large clusters; For each large cluster, secondary clustering is performed again and subdivided into K (2) small clusters to obtain the hierarchical label of each item; the large clusters and small clusters constitute clusters; Constructing a benchmark name library; the benchmark name library includes benchmark models and structured attribute feature sets corresponding to the benchmark models F p ; Get the structured attribute feature set of the cluster F c ; Compute the structured attribute feature set corresponding to the baseline model F p and structured attribute feature sets of clusters F c and determining a category model that matches the cluster based on the overlap rate; the category model includes: a reuse model, an extended model, and a new model; For clusters that require new models, candidate names are generated based on the characteristic content or domain terms of the clusters; Check whether the candidate name already exists in the benchmark name library. If the candidate name already exists, it indicates a naming conflict. The system will prompt a name conflict and use the edit distance or synonym library to calculate a list of candidate names closest to the candidate name for selection; if the candidate name does not exist, directly use this candidate name, assign it a new unique ID and register it in the library to obtain the classification result of each item.

2. The large-scale object classification method according to claim 1, characterized in that: The large-scale object classification method further includes, after the last step: The classification result of each item is represented as a tuple.

3. The large-scale object classification method according to claim 1, characterized in that: The following formula is used to extract the visual feature vector, semantic embedding vector and attribute field vector from the original item set: ； in, is the visual feature vector, is the semantic embedding vector, is the attribute field vector, For images, is a text description, is a structured attribute field, is the visual feature space, is the text semantic embedding space, Encoding space for structured attributes.

4. The large-scale object classification method according to claim 1, characterized in that: The visual feature vector, semantic embedding vector, and attribute field vector are concatenated and fused to obtain the multimodal vector of the item using the following formula: ； in, is the vector splicing symbol, is the visual feature vector, is the semantic embedding vector, is the attribute field vector, is the Euclidean space where the three types of vectors are located, For the i A high-dimensional multimodal representation vector of an item, is the visual sub-vector dimension; is the semantic sub-vector dimension; is the attribute sub-vector dimension; is the sum of the dimensions of the three types of sub-vectors.

5. The large-scale object classification method according to claim 1, characterized in that: The multimodal vector is reduced in dimension to obtain a low-dimensional latent vector. Specifically, an autoencoder is introduced to reduce the dimension of the multimodal vector. The specific formula is as follows: coding: ; decoding: ; loss: ; in, is a low-dimensional latent vector; Represents the encoder network Enc with parameters For control, input sample , outputs a latent space vector , represents the compressed representation of the sample; is the potential mean vector output by the encoder; For the i A high-dimensional multimodal representation vector for each item; represents the standard deviation vector of the latent distribution of the encoder output, which is uncertainty scale mapped to the latent space; is a noise vector that obeys the standard normal distribution; represents a random vector sampled from a standard normal distribution; is the representation vector reconstructed by the decoder, if the result is the same as The closer they are, the more The more information you have; represents the decoder network, with parameters control; is the abbreviation of decoder function; and Represent the learnable parameters of the encoder and decoder respectively; is the total loss function of the variational autoencoder; N is the total number of item samples in the batch; i is the sample index number; is the square of the Euclidean distance; is the balance coefficient; is the Kullback–Leibler divergence; is the approximate posterior distribution; is the prior distribution.

6. The large-scale object classification method according to claim 1, characterized in that: The low-dimensional potential vector is clustered into the following categories: K (1) The following formula is used for each major cluster: ； in, The number of preset categories; For the j A collection of sample indexes of major categories; is a low-dimensional latent vector, For large categories The center of mass; is the square of the Euclidean distance; For each large cluster, secondary clustering is performed again and subdivided into K (2) The following formula is used to obtain the hierarchical label of each item: ； in, Preset the number of sub-category clusters to be subdivided within each major category; is a small cluster sample set, k is the subcategory index number, For the In the major categories The centroid of each small cluster; The hierarchical label of each item is ,in, For the i The category labels to which the items are assigned in the first-level cluster; It is the subdivision label of the same item in the secondary cluster.

7. The large-scale object classification method according to claim 1, characterized in that: For clusters that require a new model, the following formula is used to generate candidate names based on the characteristic content or domain terms of the cluster: ； in, is the cosine similarity; is the secondary cluster center; is the name vector; is a candidate name; To select the item with the greatest similarity among all candidate models, arg Indicates the corresponding variable value when a certain expression takes the maximum or minimum value.

8. The large-scale object classification method according to claim 1, characterized in that: Compute the structured attribute feature set corresponding to the baseline model F p and structured attribute feature sets of clusters F c The overlap rate of the clusters is determined based on the overlap rate, and the category model that matches the cluster cluster is determined using the following formula: ； in, for F p and F c The overlap rate of When , a reusable model is recommended; when When , the extended model is recommended; when It is recommended to create a new model.

9. The large-scale object classification method according to claim 1, characterized in that: The classification results of each item are expressed as tuples as follows: ； in, It is the item identifier. For the i The category labels to which the items are assigned in the first-level cluster; is the subdivision label of the same item in the secondary cluster, is the finalized benchmark name. Indicates the matching status of the category with the benchmark model, including reuse, extension or creation. It is the referenced or newly created attribute model. If it is newly created, it is the new model ID. The overlap ratio calculated during matching.

10. A large-scale object classification system, characterized by: The large-scale object classification system includes: An original item set acquisition module is used to acquire an original item set; the original item set includes: images, text descriptions, and structured attribute fields; A vector extraction module, configured to extract visual feature vectors, semantic embedding vectors, and attribute field vectors from the original item set; A splicing and fusion module, configured to splice and fuse the visual feature vector, the semantic embedding vector, and the attribute field vector to obtain a multimodal vector of the object; A dimensionality reduction module, configured to reduce the dimensionality of the multimodal vector to obtain a low-dimensional latent vector; The first-level clustering module is used to perform first-level clustering on the low-dimensional potential vector, which is roughly divided into K (1) Large clusters; The secondary clustering module is used to perform secondary clustering on each large cluster, which is further subdivided into K (2) small clusters to obtain the hierarchical label of each item; the large clusters and small clusters constitute clusters; A base name library construction module is used to construct a base name library; the base name library includes a base model and a structured attribute feature set corresponding to the base model F p ; A clustering cluster structured attribute feature set acquisition module, used to obtain the clustering cluster structured attribute feature set F c ; Overlap rate calculation module, used to calculate the structured attribute feature set corresponding to the benchmark model F p and structured attribute feature sets of clusters F c and determining a category model that matches the cluster based on the overlap rate; the category model includes: a reuse model, an extended model, and a new model; A candidate name determination module is used to generate candidate names for clusters that require a new model based on the characteristic content or domain terms of the clusters; The retrieval module is used to check whether the candidate name already exists in the benchmark name library. If the candidate name already exists, it indicates a naming conflict. The system will prompt the name conflict and use the edit distance or synonym library to calculate a list of candidate names closest to the candidate name for selection; if the candidate name does not exist, the candidate name will be directly used, a new unique ID will be assigned to it and registered in the library to obtain the classification result of each item.

Images