Label recommendation method and vehicle

Abstract

The invention relates to the field of data classification, and particularly discloses a tag recommendation method and a vehicle, and the method comprises the steps: extracting multi-dimensional data features of to-be-marked data, the data features comprising text features, image features and structured data features; performing feature fusion on the multi-dimensional data features to obtain a fused feature vector; based on a pre-constructed label association model and the fusion feature vector, determining recommendation scores of the to-be-marked data and each label; and determining a target label corresponding to the to-be-marked data based on the recommendation score. According to the method, content and structure information of data can be comprehensively captured by extracting various modal features including texts, images and structured data, and effective integration and complementation of multi-source information are realized by performing fusion processing on multi-dimensional features and generating a unified fusion feature vector. And the recommendation score is calculated based on the pre-constructed tag association model and the fusion feature vector, so that accurate inference from data features to tag mapping is realized.

Description

Technical Field

[0001] This application relates to the field of data classification, specifically to a label recommendation method and vehicles. Background Technology

[0002] In the current fields of data management, content retrieval, and knowledge organization, adding tags to data objects (such as documents, images, structured datasets, etc.) is a widely used technique. It aims to classify, describe, and index data through semantic keywords, thereby improving the discoverability, manageability, and subsequent utilization value of data.

[0003] Traditional tagging methods rely heavily on manual operation, typically requiring users to manually select and associate one or more tags with the target data from a predefined tag set based on their understanding of the data content. With the rapid growth in data volume and the increasing diversification of data types (such as text, images, and tables coexisting), this manually-driven model has gradually revealed problems such as inefficiency, difficulty in guaranteeing tag quality, and insufficient utilization of the tag system. Summary of the Invention

[0004] In view of the above problems, this disclosure provides a tag recommendation method and vehicle to overcome or at least partially solve the above problems, the technical solution of which is as follows: A tag recommendation method, characterized by comprising: extracting multidimensional data features of data to be tagged, the multidimensional data features including text features, image features, and structured data features; performing feature fusion on the multidimensional data features to obtain a fused feature vector; determining the recommendation score of the data to be tagged and each tag based on a pre-constructed tag association model and the fused feature vector; and determining the target tag corresponding to the data to be tagged based on the recommendation score.

[0005] By integrating modal features from text, images, and structured data, recommendation bias caused by incomplete information from a single modality is avoided. Fusing multimodal features generates a more comprehensive data description. Utilizing a pre-trained label association model, complex real-time calculations are eliminated for each recommendation; instead, historical patterns are quickly matched based on fused features, directly outputting recommendation scores. This method transforms traditional manual feature matching and rule-based judgment into automated model inference based on a unified feature vector, directly reducing manual labeling steps and judgment time, improving labeling efficiency, and ensuring consistency in recommendation logic across different data types.

[0006] In one example, the extraction of multidimensional data features from the data to be labeled specifically includes: extracting keywords from the text content based on the weights of word frequency statistics and inverse document frequency to obtain a first text feature; mapping the text content to a continuous vector space to obtain a contextual semantic representation to obtain a second text feature; merging the first and second text features to obtain a text feature; extracting the low-level visual features of the image and obtaining high-level semantic features by pooling and weighting the output features of the pre-trained model; the low-level visual features include at least one of color features, texture features, and shape features; merging the low-level visual features with the high-level semantic features to obtain image features; encoding structured fields and extracting statistical features between fields to obtain structured data features; field encoding includes at least one of one-hot encoding and field name embedding encoding; the statistical features include at least one of mean, variance, number of unique values, and correlation coefficient between fields.

[0007] In terms of text features, combining keyword and semantic features enables the system to capture both explicit keywords and deeper semantic meanings in documents, thus addressing issues related to technical terms and synonyms. For images, simultaneously extracting low-level features such as color and texture, as well as high-level semantic features, allows the system to recognize both the style of an image and the specific objects within it. For structured data, field encoding and statistical feature extraction effectively process tables, attribute lists, and other data, identifying correlations between fields. This refined feature extraction provides high-quality, complementary raw materials for subsequent fusion, directly improving the representation accuracy of feature vectors.

[0008] In one example, the step of fusing the multidimensional data features to obtain a fused feature vector includes: performing weighted fusion on the multidimensional data features to obtain a weighted feature vector; and / or concatenating feature vectors from different modalities to obtain a concatenated feature vector; and / or calculating an interaction feature vector between feature vectors from different modalities; and generating the fused feature vector based on at least one of the weighted feature vector, the concatenated feature vector, and the interaction feature vector.

[0009] Weighted fusion dynamically adjusts the contribution of different modalities based on the task. For example, when analyzing vehicle exterior images, it automatically assigns higher weights to image features, while when analyzing technical documents, it assigns higher weights to text features, thus generating more targeted feature representations. Concatenation fusion preserves the dimensions of all original features, providing complete input information for subsequent complex deep learning models and allowing the model to internally learn higher-order interactions between features. Calculating interaction features reveals direct correlations across modalities, and these strategies can be selected or combined according to the scenario, achieving optimal information integration during the feature fusion stage and directly improving the input quality of subsequent recommendation steps.

[0010] In one example, before determining the recommendation scores of the data to be labeled and each label based on the pre-built label association model and the fused feature vector, the method further includes: constructing multi-dimensional interaction data based on historical labeling records, the multi-dimensional interaction data being used to represent the interaction strength between users, data, and labels; decomposing the multi-dimensional interaction data into multiple low-dimensional factor matrices, the low-dimensional factor matrices including at least a user factor matrix, a data factor matrix, and a label factor matrix; and constructing a label association model based on the multiple low-dimensional factor matrices.

[0011] By constructing a three-dimensional interaction tensor of user-data-tags and performing tensor decomposition, two core challenges in recommender systems are addressed: data sparsity and high-order association modeling. In large systems, most user-data combinations never interact, resulting in extremely sparse data. Tensor decomposition, by decomposing a high-dimensional sparse tensor into a low-dimensional factor matrix, enables computation and prediction in a low-dimensional dense space, effectively inferring unobserved user-tag preferences. Simultaneously, three-dimensional modeling can capture complex patterns such as "user A frequently labels data containing feature X with label Y," which is more accurate than considering only two-dimensional data-tag associations. The cosine similarity between the vectors in the decomposed label factor matrix can be directly used to measure the semantic relevance between labels, providing additional association knowledge for recommendations.

[0012] In one example, determining the recommendation scores of the data to be labeled and each label based on the pre-built label association model and the fused feature vector specifically includes: constructing a graph structure containing data nodes, label nodes, and user nodes, where edges in the graph structure represent the interaction relationships between nodes; learning representations for the nodes in the graph structure based on the label association model; using the fused feature vector as the representation of the data to be labeled in the latent space corresponding to the label association model; and determining the recommendation scores of the data to be labeled and each label based on the fused feature vector and the learned graph structure.

[0013] This approach constructs the recommendation problem on a heterogeneous graph containing user, data, and tag nodes, and utilizes graph neural networks for representation learning. The core advantage of this method lies in its ability to explicitly leverage the network structure for reasoning. For example, the representation of a tag node is determined not only by its own semantics but also by aggregating information from user nodes that have been tagged with that tag, as well as related data nodes. Through multiple rounds of message passing, the representation of a data node to be tagged can "sense" the tag nodes indirectly connected to it in the graph (e.g., through sharing similar users or associating similar data), thereby discovering tags with low content similarity but highly related usage patterns. This achieves a deep fusion of content filtering and collaborative filtering at the model architecture level, enabling the recommendation of tags that are difficult to discover solely through data content analysis but are widely recognized in community practice.

[0014] In one example, determining the recommendation scores for the data to be labeled and each label based on the fused feature vector and the learned graph structure specifically includes: inputting the fused feature vector into the data node representation in the graph structure; performing message passing and node representation updates through a multi-layer graph convolutional network to obtain updated data node representations and label node representations; calculating the cosine similarity between the data node representation and each label node representation; and converting the cosine similarity into a probability score using a preset function, which is then used as the recommendation score.

[0015] Using fused feature vectors as the initial representation of data nodes ensures that the rich content information of the data itself can be directly input into the model. Multi-layer graph convolution is key; each convolution layer allows a node to absorb information from its one-hop neighbor. After two layers, a data node can receive information from its neighboring users with associated labels, thus learning collaborative patterns of "which user groups prefer this type of data." Finally, by calculating the cosine similarity between data nodes and label nodes in the same graph embedding space, the complex graph structure reasoning results are transformed into a quantifiable matching score. This score combines content matching and collaborative association, and is then converted into a probability, making the recommendation results both accurate and possessing interpretable confidence.

[0016] In one example, after determining the target label corresponding to the data to be labeled based on the recommendation score, the method further includes: in response to receiving a user's confirmation or correction operation on the target label, updating the multidimensional interaction data as a new historical labeling record; incrementally updating the user factor matrix, data factor matrix, and label factor matrix based on the updated multidimensional interaction data; and updating the node representation and parameters of the graph structure based on the updated factor matrix.

[0017] By establishing a feedback and incremental update mechanism, the system acquires online learning capabilities, enabling it to respond to user behavior in real time. When a user confirms or modifies a recommendation tag, the system doesn't simply log it; instead, it immediately uses this interaction as a new training sample, incrementally updating the factor matrix of users, data, and tags. This means that a user's act of adding a new tag to certain data can quickly influence subsequent recommendations for similar users and may affect recommendations for other similar users through shared tag factors. Simultaneously, the node representations and model parameters in the graph structure are also fine-tuned. This process avoids the huge computational overhead of periodically retraining the entire dataset, ensuring the system keeps pace with changes in user behavior and the evolution of the tag system, keeping recommendation results consistently novel and personalized, and improving the consistency of the user experience.

[0018] In one example, determining the target label corresponding to the data to be labeled based on the recommendation score includes: sorting each label in descending order of recommendation score; and selecting the first preset positive integer labels from the sorting results as the target labels.

[0019] In one example, determining the target label corresponding to the data to be labeled based on the recommendation score includes: prompting the user to input a new label in response to the fact that the recommendation scores of all labels are lower than a preset threshold; adding the new label input by the user to the system label set; and updating the multidimensional interaction data as new data by treating the interaction record between the new label and the data to be labeled.

[0020] The system defines specific output strategies and procedures for handling unknown situations. Descending order sorting and prioritizing tags with higher recommendation scores efficiently guides users to focus on the most likely candidates. A key design feature is proactively prompting users to input new tags when all recommendation scores fall below a threshold. This avoids errors that might arise from the system forcing recommendations under low confidence levels and provides users with a channel to create and define new tags, turning the system's "blind spots" into opportunities for knowledge expansion. Subsequently, the system immediately adds the new tag to the tag library and updates its first use record in the interaction data and association model, allowing the system to immediately "understand" the new tag and use it for subsequent recommendations. This closed-loop process ensures that the system's tag library can be dynamically and cost-effectively expanded to meet business needs.

[0021] This application also provides a vehicle, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the tag recommendation method described in any of the above examples.

[0022] By employing the aforementioned technical solutions, the tag recommendation method and vehicle disclosed herein significantly improve the accuracy and efficiency of tag recommendation through the comprehensive utilization of multimodal data features and advanced model inference mechanisms. The system can not only process various types of data but also dynamically adjust feature weights according to specific task requirements, thereby generating recommendation results that are more closely aligned with actual application scenarios. Furthermore, by introducing user feedback and incremental update mechanisms, the system possesses strong adaptive capabilities, continuously optimizing its performance as user behavior changes and the tag system expands. This design makes the tag recommendation process more intelligent and personalized, reducing manual intervention while improving the consistency and reliability of tagging.

[0023] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, specific embodiments of this disclosure are described below. Attached Figure Description

[0024] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this disclosure. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a tag recommendation method according to an embodiment of this disclosure is shown; Figure 2 A schematic diagram of the structure of a label recommendation device according to an embodiment of the present disclosure is shown; Figure 3 A schematic diagram of the structure of a vehicle according to an embodiment of the present disclosure is shown. Detailed Implementation

[0025] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0026] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0027] In the current fields of data management, content retrieval, and knowledge organization, adding tags to data objects (such as documents, images, structured datasets, etc.) is a widely used technique. It aims to classify, describe, and index data through semantic keywords, thereby improving the discoverability, manageability, and subsequent utilization value of data.

[0028] Traditional tagging methods rely heavily on manual operation. Specifically, users typically need to manually select and associate one or more tags with the target data from a predefined tag set based on their understanding of the data content. With the rapid growth of data scale and the increasing diversification of data types (such as the coexistence of text, images, and tables), this manual-driven model has gradually exposed several inherent defects: First, facing massive and continuously growing data, relying entirely on manual content identification and tag matching requires a significant investment of manpower and time, becoming a significant bottleneck in the data governance process.

[0029] Secondly, due to differences in users' knowledge background, cognitive habits, and judgment standards, different users may choose different tags for the same or similar data, and even the same user may make inconsistent choices in different contexts. This subjectivity and arbitrariness can easily lead to confusion, inaccuracy, or incompleteness in tag usage, thereby weakening the consistency and authority of the tagging system.

[0030] Furthermore, faced with a complex system that may contain tens of thousands of tags, users find it difficult to fully grasp its structure and semantic scope. Therefore, when tagging, users may not be able to accurately recall or find the most appropriate tags already existing in the system, leading to misuse of tags, duplicate creation, or failure to fully utilize the existing system, which affects the standardization of data organization and the tightness of knowledge association.

[0031] Furthermore, existing solutions often lack effective learning and adaptive capabilities. They typically store and query historical tagging records merely as static relationships, failing to delve into the underlying user behavior patterns, data features, and deep correlations between tags. The system cannot continuously learn from past interactions and optimize its recommendation logic, resulting in limited intelligence and an inability to adapt to constantly changing data formats and user needs.

[0032] Therefore, the industry urgently needs an automated or semi-automated technical solution that can overcome the aforementioned shortcomings to improve the efficiency and standardization of tag addition, ensure tag quality, and fully utilize existing knowledge systems. Although some tag suggestion methods based on keyword matching or simple collaborative filtering exist, they are generally insufficient in terms of the comprehensiveness of feature coverage, the depth of association modeling, and adaptability to complex multimodal data scenarios.

[0033] Therefore, this application provides a tag recommendation method, such as... Figure 1 The diagram illustrates a label recommendation method provided by one or more embodiments of this specification. This method can be applied to different types of business data, such as internet finance business data, e-commerce business data, instant messaging business data, and government business data. The process can be executed by computing devices in the corresponding field (e.g., risk control servers or smart mobile terminals for payment services). Certain input parameters or intermediate results in the process can be manually adjusted to help improve accuracy.

[0034] The analysis method involved in the embodiments of this application can be implemented by a terminal device or a server. For ease of understanding and description, the following embodiments will all use a server as an example for detailed description. It should be noted that the server can be a single device or a system composed of multiple devices, i.e., a distributed server. This application does not specifically limit it in this way.

[0035] like Figure 1 As shown in the figure, this application provides a tag recommendation method, including: S101: Extract multidimensional data features from the data to be labeled, including text features, image features, and structured data features.

[0036] First, after receiving the data to be labeled, the server extracts multidimensional data features from it. These multidimensional data features include text features, image features, and structured data features. The data to be labeled can be any digital object that needs to be tagged, such as a text document, an image, a data table, or a composite file containing multiple elements. The server processes this data using preset feature extraction methods. For text, such as a documentation document or log file, the server analyzes its character sequences, vocabulary, and sentence structure to extract text features reflecting its theme and key information. For images, such as a product photo or design drawing, the server analyzes the color distribution, texture patterns, and shape contours in its pixel array to extract image features describing its visual content. For structured data, such as a data table containing model numbers, parameters, and production dates, the system parses its row and column organization and field values ​​to extract structured data features reflecting its field types, value range distribution, and internal table relationships. Finally, one or more sets of numerical vectors or feature lists are output; these are collectively referred to as data features.

[0037] Multidimensional data features refer to numerical representations extracted from different analytical dimensions of data (such as text, visual, and structural), used to describe different attributes of the data. Text features are numerical vectors reflecting information such as semantics, keyword frequency, and syntactic structure of text content. Image features are numerical vectors describing visual attributes of images, such as color, texture, shape, and high-level semantic content. Structured data features are numerical vectors reflecting the structure, value distribution, and relationships between fields of data with fixed field formats, such as tables and logs.

[0038] By simultaneously extracting features from text, images, and structured data, the server can cover key descriptive dimensions of different data types, avoiding descriptive biases caused by the incompleteness of a single feature type. For example, when processing a technical report with both text and images, relying solely on the text may overlook key components in the illustrations, while analyzing only images cannot capture the descriptions of technical parameters. Extracting both simultaneously creates complementary information, enabling subsequent steps to analyze based on a more complete data profile.

[0039] In one embodiment, the server first preprocesses the raw text data, including but not limited to: removing irrelevant HTML / XML tags, filtering special characters, standardizing text encoding, segmenting text (for long documents), and performing word segmentation based on a dictionary or statistical model (for non-space-separated languages ​​such as Chinese). This provides standardized and clean text input for subsequent feature extraction.

[0040] During basic statistical feature extraction, the server extracts statistical features from the preprocessed text that intuitively reflect the surface information of the document. First, a first text feature vector is generated using the term frequency-inverse document frequency vector. The server calculates the weight of each term in the text, forming a high-dimensional sparse vector representation of the document using a bag-of-words model. Specifically, for a document set... Documents in and terms Its weight The calculation formula is as follows:

[0041] in, It is a term In the document Frequency of occurrence in It is the total number of documents in the document collection. It contains terms A collection of documents, Indicates included terms The number of documents in the document collection. The server uses this method to generate the first text feature vector of each document, which highlights important keywords within the document.

[0042] The server can also simultaneously calculate a series of document-level statistics, including: total text length (number of characters or words), number of paragraphs, number of sentences, average sentence length, lexical richness (such as the percentage of unique words), etc. These metrics, as supplementary numerical features, together with the weights, constitute the basic set of statistical features.

[0043] When performing deep semantic feature extraction, the server first converts the preprocessed text into the standard input format of the BERT model. The specific construction method is as follows:

[0044] in, It is a classification label, and its corresponding output vector is usually used to represent the semantics of the entire sequence; It is a sequence of tokens obtained by segmenting the original text into sub-word units using the WordPiece tokenizer; This is a separator used to distinguish different input sequences (placed at the end of the sentence in single-sentence tasks). During semantic vector extraction, the server inputs the constructed BERT input sequence into the pre-trained BERT model. The semantic vectors are extracted from the hidden states of the last layer of the model. The output vector of the tags, or the average pooling of the output of the last layer of all tags in the sequence, yields a dense vector of fixed dimensions. This vector encodes the comprehensive semantic information of the words in a specific context, forming the second text feature vector of the document.

[0045] The server can also use pre-trained word embedding models to map each word in a document to a dense vector of fixed dimensions, and generate a semantic representation of the document as supplementary features through aggregation operations (such as average pooling). The server can also apply topic models (such as LDA – Latent Dirichlet Allocation) to perform unsupervised learning on the document set, obtaining the probability distribution of each document across multiple latent topics. This distribution vector reflects the macroscopic content composition of the document.

[0046] After obtaining the aforementioned text feature vectors, the server can merge the extracted text features from different levels to form a unified text feature representation. Taking the first and second text feature vectors as an example, the server can perform a concatenation operation between the first text feature vector (statistical features such as TF-IDF) and the second text feature vector (semantic vectors based on BERT). For example, if the first vector has a dimension of... The second vector dimension is The dimension of the concatenated vector is then... Optionally, the server can incorporate document statistical metrics as additional features into the final vector, or as a separate metadata feature stream for later integration in the cross-modal fusion stage. Ultimately, the server outputs a text feature vector that integrates keyword weights, statistical attributes, and deep semantics; this vector will serve as one of the inputs to the subsequent multimodal feature fusion module.

[0047] Through the above steps, the server achieves comprehensive feature generation of text data from surface statistics to deep semantics. In particular, through standardized BERT input construction and semantic extraction, it ensures the quality and consistency of semantic features, providing high-quality and information-rich text-side input for subsequent intelligent label recommendation.

[0048] In one embodiment, image features include low-level visual features and high-level semantic features based on deep learning. The server synthesizes these two levels of features to construct a robust feature vector that comprehensively describes the image content and semantics. The specific steps are as follows: The server first preprocesses the input image, including size normalization and color space conversion (e.g., converting to RGB or HSV space). Then, it calculates at least one of the following low-level visual features: color features, texture features, and shape features. When calculating color features, the distribution of different color values ​​can be statistically analyzed by calculating the image's color histogram; or the global distribution characteristics of colors can be described by calculating color moments (such as the first-order moment mean, second-order moment variance, and third-order moment skewness). When calculating texture features, the roughness and regularity of the texture can be described by calculating the image's gray-level co-occurrence matrix and extracting statistics such as contrast, correlation, energy, and homogeneity. When calculating shape features, edge detection operators can be used to extract image contours, or the oriented gradient histogram can be calculated to describe the edge or shape information of the target.

[0049] To extract high-level semantic features, servers can use pre-trained deep convolutional neural network models (such as ResNet50 and VGG16) as feature extractors to obtain features with greater semantic abstraction capabilities. When extracting feature maps, the pre-processed image is input into the pre-trained CNN model, and feature maps are obtained from specific layers of the model (e.g., the output of the last convolutional layer in ResNet50). These feature maps contain high-level visual patterns learned by the network. The extracted original feature map can be represented as...

[0050] in, The input image tensor, This represents the forward propagation function of the pre-trained CNN model. The image shows the extracted original feature map. To aggregate the two-dimensional feature map into a fixed-length vector and enhance the translation invariance of the features, the server performs a global pooling operation on the feature map.

[0051] Global average pooling or global max pooling is typically used. Global average pooling can be expressed by the following formula:

[0052] in, It is a channel index. and These are the height and width of the feature map, respectively. Global max pooling can be expressed by the following formula:

[0053] The server denotes the pooled vector as... ,in This represents the number of feature channels (e.g., 2048 dimensions for ResNet50).

[0054] To further focus on key regions in the image, the server can introduce an attention mechanism. First, it uses a learnable weight matrix... The pooled features are transformed, and attention weights are calculated:

[0055] in, This is the normalized attention weight vector. Then, the weights are applied to the original pooling features:

[0056] in, This indicates element-wise multiplication. This is the attention-weighted feature vector. The resulting low-level visual feature vector (denoted as...) ) and high-level semantic feature vector (denoted as After that, the features need to be merged. For example, a splicing method can be used to merge them, resulting in the final image feature vector. for: The server outputs image feature vectors. This serves as one of the inputs for the subsequent feature fusion module.

[0057] In one embodiment, structured data typically refers to data with clearly defined fields and a fixed format, such as database tables, CSV files, JSON objects, or XML documents. When the server extracts structured data features from the data to be tagged, it first identifies and parses the structured data. Specifically, the server first receives the data to be tagged uploaded by the user and identifies whether it is in a structured data format. If so, it parses the structure of the data and extracts the field names, data types, and corresponding values ​​for each field. For example, for a vehicle information table, after parsing, fields such as "vehicle model," "year of production," and "engine power" can be obtained, along with the corresponding text or numerical content for each field.

[0058] Then, the server extracts metadata features for each parsed field. Specifically, it treats the field name as a short text and maps it to a low-dimensional real-valued vector using a word embedding model to capture the semantic information of the field name. It identifies the data type of each field, such as "integer," "floating-point," "string," and "date," and generates a binary vector for that type using one-hot encoding. For numeric fields, the server calculates the statistical features of all its values, including mean, variance, minimum, maximum, quantiles, number of unique values, and field relevance coefficient. For categorical fields, it calculates the frequency of occurrence of different categories and the total number of categories.

[0059] After extracting metadata features, the server further analyzes the relationships between fields and extracts relational features. Specifically, for numeric field pairs, it calculates the Pearson correlation coefficient or Spearman rank correlation coefficient between them to obtain a scalar value representing linear or monotonic correlation. For key-value pairs or tabular data, the server extracts features such as key name string length, value type diversity, total number of rows in the table, total number of columns, and data sparsity.

[0060] Finally, the server standardizes and vectorizes the extracted features, performing Z-score standardization or min-max normalization on numerical statistical features and correlation coefficients to bring them into similar numerical ranges. The field name embedding vector, data type one-hot encoded vector, standardized statistical feature vector, and relational feature vector are concatenated in a preset order to form a high-dimensional intermediate feature vector. To reduce dimensionality and remove redundant information, the server uses principal component analysis or a pre-trained autoencoder to reduce the dimensionality of this intermediate feature vector, ensuring that most of the original information is retained. Ultimately, the server outputs a fixed-dimensional real-number vector as the feature representation of the structured data, denoted as . This is used for subsequent feature fusion and label recommendation.

[0061] S102: Perform feature fusion on the multidimensional data features to obtain a fused feature vector.

[0062] After acquiring various types of data features, the server fuses these features. Specifically, the server can use methods such as weighted summation, vector concatenation, or dynamic weighting based on attention mechanisms to combine features from different modalities into a unified numerical vector, i.e., the fused feature vector. This step aims to integrate feature information from different sources and semantic levels into a common mathematical representation, eliminating heterogeneity between modalities and facilitating unified processing by subsequent models. Feature fusion refers to the process of merging feature vectors from multiple sources or types into a single vector through mathematical operations. The fused feature vector is the single vector formed after fusion processing, carrying integrated information from multi-dimensional features.

[0063] Feature fusion addresses the issues of inconsistent representations and information isolation among multimodal features. By generating a unified fused feature vector, the system maps information originally scattered across different feature spaces into a common space. This enables subsequent models to learn and infer based on a structured input, simplifying the processing flow and enhancing the collaborative expressive power among different features.

[0064] By fusing structured data, the limitations of text and image features can be compensated for, enhancing the system's ability to understand the overall data and improving the accuracy, diversity, and interpretability of recommendations. This is particularly suitable for enterprise-level data management scenarios containing rich metadata and logical relationships. Specifically, structured data (such as metadata, field information, and relational information) contains rich logical relationships and contextual information, such as data type, field names, and data distribution. Fusing these features can more comprehensively describe the inherent structure and attributes of the data, thereby improving the expressive power of feature vectors. For example, in a dataset, structured information such as "vehicle model," "year of production," and "sensor type" has clear guiding significance for label recommendations. In addition, structured data often contains information such as correlations between fields, key-value pair structures, and table structures, which are not easily extracted from plain text or images. By fusing relational features (such as Pearson correlation coefficient, number of rows / columns, sparsity, etc.), the model can better understand the structured dependencies between data, improving the logical consistency and interpretability of the recommendation system. Structured features typically contain information such as category, enumeration, time, and numerical distribution, which helps the model discover implicit patterns across modalities. For example, the same car model might correspond to different tag trends in datasets from different years. Integrating year information can improve the novelty and diversity of recommendations. In enterprise systems, structured data is often standardized and normalized. Integrating these features helps models better adapt to the enterprise's data governance system and improves the standardization and consistency of tags.

[0065] In one embodiment, when performing feature fusion, the multidimensional data features can be weighted and fused to obtain a weighted feature vector; and / or, feature vectors of different modalities can be concatenated to obtain a concatenated feature vector; and / or, interaction feature vectors between feature vectors of different modalities can be calculated; and the fused feature vector is generated based on at least one of the weighted feature vector, the concatenated feature vector, and the interaction feature vector.

[0066] Specifically, after the server receives the data to be labeled, it first obtains its multidimensional data features through a feature extraction module. These features may include text feature vectors. Image feature vectors and structured data feature vectors Subsequently, the server's feature fusion engine will perform one or more of the following fusion operations to generate a fused feature vector: During weighted fusion, the server is configured with a learnable attention weight matrix. And a fusion controller. For multiple modal feature vectors as input, the server first calculates a dynamic weight for each feature vector. Specifically, for the ... Feature vectors of each modality Its attention weight It can be calculated using the following formula:

[0067] The softmax function ensures that the sum of all weights is 1. This is a learnable attention weight matrix. Subsequently, the server calculates the weighted sum of each feature vector and its corresponding weight to obtain the weighted feature vector. :

[0068] in, This refers to the number of feature modalities participating in the fusion. This weighting method allows the server to automatically adjust the contribution of different modal features in the final representation based on the characteristics of the current data. For example, for a text-based report, the weight of text features may be automatically increased.

[0069] During feature concatenation, the server concatenates feature vectors from different modalities along the feature dimension to generate a concatenated feature vector. For example, if the text feature dimension is 768, the image feature dimension is 2048, and the structured feature dimension is 512, then the dimension of the concatenated feature vector is... .

[0070]

[0071] Here, [;] represents the vector concatenation operation. This operation preserves all original feature information, providing complete input for potentially complex models (such as deep neural networks) that may be used later.

[0072] When calculating interactive features, in order to explicitly model the intrinsic relationships between different modal features, the server calculates element-wise products between different feature vectors as interactive feature vectors. For example, calculating the interaction between text features and image features. :

[0073] in, This represents element-wise multiplication. Interactive features can capture cross-modal cooperative or co-occurrence patterns; for example, the word "sky" appearing in text is strongly correlated with blue area features in an image.

[0074] Finally, when generating the fused feature vector, the server selects one or more of the aforementioned intermediate feature vectors (weighted feature vectors) according to a preset configuration strategy. , Concatenate feature vectors Interactive feature vectors This is used to generate the final fused feature vector. .

[0075] The strategy could be: to use only the results of the weighted fusion, i.e. Suitable for scenarios requiring lightweight models and high real-time performance.

[0076] Alternatively, the weighted feature vector can be concatenated with one or more interactive feature vectors. For example, ,in It is an interaction between text and structured features, which takes into account both dynamic weights and explicit associations.

[0077] Alternatively, the concatenated feature vector can be concatenated with all the important interactive feature vectors. This provides the richest feature representation, but with higher dimensionality.

[0078] In generation Subsequently, the server can selectively pass it through a fully connected layer and perform a nonlinear transformation (such as ReLU activation) to further integrate information and reduce dimensionality, obtaining a standardized fused feature vector. : .in, and These are learnable parameters. Ultimately, the server will use this. The data is fed into the downstream label association model to calculate recommendation scores. Through the aforementioned flexible and configurable fusion mechanism, the server can adapt to the needs of different types of data (text-oriented, image-oriented, or multimodal balanced), balancing the richness of feature representation with computational efficiency.

[0079] In one embodiment, before determining the recommendation scores of the data to be labeled and each label based on the pre-built label association model and the fused feature vector, the server constructs multi-dimensional interaction data based on historical labeling records. The multi-dimensional interaction data is used to represent the interaction strength between users, data, and labels. The multi-dimensional interaction data is decomposed into multiple low-dimensional factor matrices, which include at least a user factor matrix, a data factor matrix, and a label factor matrix. Based on the multiple low-dimensional factor matrices, a label association model is constructed.

[0080] Specifically, the server retrieves historical tagging records from the system database. Each historical tagging record contains at least one triple: user identifier, data identifier, and the tag identifier that the user assigned to that data. Optionally, the record may also include interaction timestamps, confidence levels (e.g., whether the user confirmed a recommendation or entered the data themselves), and other information. Based on these historical records, the server constructs a three-dimensional, multi-dimensional interaction data (or interaction tensor). .in, This represents the total number of users in the system. This represents the total number of historical data objects. This represents the total number of tags in the system's tag library. (Tensor) medium elements The value is used to quantify the user ,data and tags The interaction strength among the three. Specifically, the interaction strength. It can be determined in at least one of the following ways: if a user exists in the history. For data Label The behavior, Otherwise Or, The value can be user For data Label The number of times, or the frequency over a period of time. Or, it can be calculated by combining a time decay factor or behavior type weight (e.g., the weight of a user "confirming" a recommendation tag is higher than the weight of "ignoring" it).

[0081] Since the vast majority of users, data, and tag combinations have never interacted, the constructed tensor It is highly sparse. In this case, it is necessary to decompose the multidimensional interactive data into multiple low-dimensional factor matrices. To address the computational and modeling difficulties caused by high-dimensional sparse tensors, the server employs tensor decomposition technology to decompose the tensor... It can be decomposed into a product of multiple low-dimensional factor matrices.

[0082] In a preferred embodiment, parallel factor analysis decomposition can be used. Specifically, the server will decompose the tensor... Decomposed into three low-dimensional factor matrices: user factor matrix each of the lines User exist Vector representation in the 3D latent space. Data factor matrix. each of the lines It is data exist Vector representation in the dimensional latent space. Label factor matrix. each of the lines It is a tag exist Vector representation in the dimensional latent space. Where, As a preset latent factor dimension, and The goal of the decomposition is to find three factor matrices such that they can be used to approximately reconstruct the original tensor through an outer product. This is achieved by optimizing the following loss function:

[0083] in, It is a tensor The set of indices of all non-zero elements (i.e., observed interactions). It is a regularization coefficient used to control model complexity and prevent overfitting. Let Frobenius norm represent the matrix. The server can use optimization algorithms such as alternating least squares or stochastic gradient descent to solve the above problem, ultimately obtaining a convergent user factor matrix. Data factor matrix and label factor matrix .

[0084] After obtaining the low-dimensional factor matrix, the server completes the core part of the tag association model. This model encapsulates the ability to derive various associations from the factor matrix. Specifically, this tag association model supports at least one of the following core operations: for any given user ,data and tags (Including combinations that have not appeared in history), the predicted interaction strength It can be calculated using the following formula: This score can be used to evaluate users. For data Select tabs The possibility.

[0085] For the given data and tags The cosine similarity of their vector representations can be calculated. As the semantic correlation between the two in the latent space.

[0086] For any two tags and It can be calculated As a semantic similarity, this reflects their co-occurrence relationship or semantic similarity in historical usage patterns.

[0087] In subsequent recommendation systems based on graph neural networks, the label factor matrix The row vectors can be directly used as the initial feature vectors of the label nodes in the graph, and the data factor matrix. The row vectors can be used as the initial feature vectors of historical data nodes, thereby injecting the knowledge learned from tensor decomposition into the graph model.

[0088] The server will generate the user factor matrix Data factor matrix and label factor matrix and latent factor dimensions The model parameters are persistently stored to form a usable tag association model. This model provides the semantic representation and prior knowledge of the association between data and tags in a unified low-dimensional space for subsequent tag recommendation processes (such as neural network recommendation).

[0089] S103: Based on the pre-built label association model and the fused feature vector, determine the recommendation scores of the data to be labeled and each label.

[0090] After obtaining the fused feature vector, the recommendation score for each tag can be determined based on the pre-built tag association model and the fused feature vector. The tag association model reflects the relationship between users, tags, and data, while the recommendation score reflects the degree of matching between the data to be labeled and each tag.

[0091] In one embodiment, when determining the recommendation score, the server constructs a graph structure containing data nodes, tag nodes, and user nodes, where edges represent the interaction relationships between nodes; based on the tag association model, representation learning is performed on the nodes in the graph structure; the fused feature vector is used as the representation of the data to be labeled in the latent space corresponding to the tag association model; and based on the fused feature vector and the learned graph structure, the recommendation scores of the data to be labeled and each tag are determined.

[0092] Specifically, the server constructs a heterogeneous graph based on the historical tagging records stored in the system. This heterogeneous graph contains three types of nodes: Data nodes: Each piece of historical data (such as documents, images, datasets, etc.) corresponds to a data node. Tag nodes: Each existing tag in the system corresponds to a tag node. User nodes: Each user who has performed a tagging operation corresponds to a user node. Edges between nodes represent interaction relationships, specifically including: Edges between user nodes and tag nodes: If a user has used a tag, an edge is established between that user node and that tag node; the edge weight can be set based on usage frequency or a time decay factor. Edges between data nodes and tag nodes: If data has been tagged, an edge is established between that data node and that tag node; the weight can be set to 1 or adjusted according to the tagging confidence level. Edges between user nodes and data nodes: Optionally, if a user has performed an operation on data (such as uploading, viewing, or tagging), an edge can also be established to enrich the graph structure. This graph structure is stored in the server's memory or a graph database in the form of an adjacency matrix or adjacency list.

[0093] The server can utilize a pre-built label association model to learn representations of nodes in the heterogeneous graph. This label association model can be a graph neural network-based model, such as a graph convolutional network or a graph attention network. The specific process is as follows: the initial feature vector of the label node can be derived from pre-trained word embeddings or vectors learned based on label co-occurrence relationships. The initial feature vector of the user node can be aggregated based on its historical labeling behavior (e.g., the average of used label vectors). The initial feature vector of the historical data nodes is a fused feature vector obtained using the feature extraction and fusion method described in claim 1.

[0094] The server performs multi-layer graph convolution operations. For the... The update formula for a layer and node can be expressed as: in: For the first The node feature matrix of the layer To add the adjacency matrix after adding self-loops, for The degree matrix, For the first The trainable weight matrix of the layer, The activation function is (e.g., ReLU). Through multiple layers of propagation, the final representation of each node... It aggregates structural and semantic information within its multi-hop neighborhood.

[0095] When data to be labeled is input, the server first performs an extraction and fusion step to obtain the fused feature vector of the data. The server will The data to be labeled is represented in the latent space corresponding to the label association model, which is considered as the initial feature vector of a new data node. The server temporarily adds the new data node to the learned graph structure, and based on the trained graph neural network model, determines its recommendation score with each label node in the following way: the initial feature vector of the new data node is used as the initial feature vector of the new data node. The input is fed into a pre-trained GNN model, and through forward propagation, its final node representation in the graph context is obtained. At this point, there's no need to retrain the entire graph; only forward inference is required. For each label node in the system... To obtain its final representation obtained through learning. Calculate the representation of the new data node. Represented by label nodes Cosine similarity between them: The cosine similarity is then converted into a probability value using the sigmoid function, which serves as the recommendation score for that tag. The formula is expressed as follows: .

[0096] S104: Determine the target label corresponding to the data to be labeled based on the recommended score.

[0097] After obtaining the recommendation scores, labels can be determined based on these scores. For example, labels with higher recommendation scores can be selected as target labels for the data to be labeled. The target labels are then associated with the data to be labeled, and the association results are stored in the system for later use. In practice, the system filters recommendation scores based on preset thresholds to ensure that only labels with high confidence are selected as target labels. This approach not only improves the accuracy of label recommendations but also reduces noise interference introduced by low-quality labels.

[0098] Furthermore, to enhance the system's interpretability, this application also includes a visualization module to display the calculation basis and feature contribution of the recommendation scores. Through this module, users can intuitively understand the specific reasons for recommending each tag, such as which text keywords, image regions, or structured fields significantly influenced the recommendation results. This transparent approach enhances user trust in the system and provides clearer operational guidance.

[0099] In one embodiment, after returning the target label to the client, the server continuously learns and optimizes. In response to receiving confirmation or correction from the user regarding the target label, the user's action is treated as a new historical tagging record and updated in the multidimensional interaction data. Based on the updated multidimensional interaction data, the user factor matrix, data factor matrix, and label factor matrix are incrementally updated. Based on the updated factor matrix, the node representations and parameters of the graph structure are updated.

[0100] Specifically, after the server receives user feedback from the client, the feedback includes at least the user identifier, the data identifier to be labeled, and the user's final action result on the recommended tags. The action result is divided into confirmation (accepting one or more recommended tags) and correction (rejecting the recommended tags and manually entering or selecting other tags). The server stores this action as a new historical labeling record triple (user identifier, data identifier, tag identifier) ​​in the historical labeling record database. Simultaneously, the server synchronously updates this triple to the corresponding position in the multidimensional interaction tensor in memory, marking the interaction strength value at that position as observed (e.g., set to 1 or a weighted value based on user behavior). If the user's correction action introduces a completely new tag, the server first creates the new tag in the system tag set and assigns a unique tag identifier.

[0101] To respond to the addition of new data and avoid the costly computation of full retraining, the server can employ an incremental tensor decomposition algorithm to process the updated multidimensional interaction tensor and update the user factor matrix, data factor matrix, and label factor matrix. Specifically, the server first fixes most of the existing factor matrix parameters, and then fine-tunes only the row vectors of the factor matrices corresponding to users, data, and labels directly related to the new interactions, as well as the global latent factor representation, by performing one or more gradient descent iterations. The regularization coefficient λ continues to play a role in this process to prevent overfitting to individual new samples.

[0102] Updating the factor matrix signifies a change in the vector representations of users, data, and labels in the latent semantic space. Therefore, the server needs to synchronously update the Graph Neural Network (GNN) model. This step involves two parts: For users, data, and labels that have changed in the factor matrix, the server uses the latest factor vectors as the new initial feature representations of these nodes in the GNN (or as components of their input features). For data nodes to be labeled, their fused feature vectors remain unchanged, but the contextual information of their associated users and labels has been updated through the factor matrix. Once system resources allow or a certain amount of updates have accumulated, the server can initiate a lightweight retraining process. Using the updated factor vectors as the initial features of the nodes, the server fine-tunes the weight parameters of the graph convolutional neural network for a limited number of rounds using the latest multidimensional interaction data (or recent data) as supervision. The training objective remains minimizing the difference between predicted and actual interactions. In this way, the message passing and aggregation rules of the graph neural network can also adaptively adjust with changes in data distribution, further optimizing recommendation performance. After completing the above steps, when the same user or other users request tag recommendations for similar data again, the server will use the updated association model and graph structure to perform calculations, thereby providing more accurate and personalized recommendation results. This closed-loop mechanism ensures that the system has the ability to continuously evolve.

[0103] In one embodiment, when determining the target label corresponding to the data to be labeled based on the recommendation score, the server sorts each label in descending order of recommendation score; selects the first preset positive integer labels from the sorting results as the target labels; in response to the fact that the recommendation scores of all labels are lower than a preset threshold, prompts the user to input a new label; adds the new label input by the user to the system label set; and updates the multidimensional interaction data as new data by using the interaction record between the new label and the data to be labeled.

[0104] Specifically, after calculating the recommended scores for the data to be labeled and all candidate labels using a graph neural network, the server sorts all scores in descending order. Then, the server reads a preset positive integer parameter K (e.g., K=5) and selects the top K label identifiers and their corresponding recommended scores from the sorted list. The server encapsulates these label identifiers and scores into a recommendation result list and sends it to the client's user interface for display. The results are typically arranged in descending order of score so that users can prioritize the most relevant suggestions.

[0105] While generating the recommendation list, the server checks the recommendation scores of all candidate tags. If the highest score is also lower than the system's preset confidence threshold (e.g., 0.3), the confidence of the recommendation is deemed too low. In this case, the server includes a special status code in the message returned to the client and generates a prompt message outside the recommendation list, guiding the user to manually enter one or more tags they deem suitable. The client then displays the input box to the user accordingly.

[0106] When the server receives a new tag text submitted by the client, it standardizes the new tag text (e.g., converts case and removes leading and trailing spaces), and then searches and compares it in the system's global tag collection. If no matching tag exists, a globally unique tag identifier is created for the new tag, and its text description, creation time, and the user who created it are stored in the tag library.

[0107] The server immediately generates a new historical tagging record, containing the current user identifier, the data identifier to be tagged, and the newly created tag identifier. This record marks the first use of the new tag. The server writes this new interaction record to the historical database and synchronously updates the multidimensional interaction tensor T in memory. For the newly added tag dimension in the tensor, it is expanded, and the interaction intensity is marked at the corresponding user, data, and the position of the new tag.

[0108] In one embodiment, when generating the final set of recommended tags, the server does not rely solely on a single matching score for ranking, but instead employs a multi-objective optimization framework to balance three core metrics: relevance, diversity, and novelty.

[0109] When calculating relevance scores, for the candidate label set Each tag in Its relationship with the data to be labeled The recommended score is recorded as ,gather correlation score The relevance score is defined as the average of the recommendation scores for all tags within the set, and can be determined using the following formula:

[0110] When calculating diversity scores, the server measures the semantic differences between tags using semantic vector representations of the tags (e.g., tag factor matrices derived from tensor decomposition or pre-trained language models). (Calculation set) The diversity score is calculated by averaging the cosine similarities between all pairwise label vectors and subtracting 1 from the average.

[0111] in, To score for diversity, For tags The semantic vector.

[0112] When calculating the novelty score, the server calculates the tags based on historical data. The "popularity" or "commonness" within the entire system. For example, it can be defined as the reciprocal of the frequency of a tag's occurrence in historical tagging records (normalized). For a set, its novelty score can be defined as the average of the novelty scores of all tags within the set. :

[0113] During optimization, the server presets the optimization target weight coefficients. And satisfy The specific weight can be configured according to the application scenario; for example, it can be increased in a rigorous archiving scenario. Weighting can be increased in creative inspiration scenarios. and Weights. The set of all candidate labels obtained by the server from the graph neural network model. In a algorithm (e.g., Top-N), optimization methods such as greedy algorithms, reinforcement learning, or evolutionary algorithms are used to find a value of size [value missing]. subset of This maximizes the following expression:

[0114] This collection This is the final set of recommended tags displayed to the user.

[0115] In addition, such as Figure 2 As shown, Figure 2 This is a schematic diagram of a label recommendation device provided in an embodiment of this application. The device includes: The data feature extraction module 201 extracts multidimensional data features of the data to be labeled, including text features, image features, and structured data features.

[0116] The data feature fusion module 202 performs feature fusion on the multidimensional data features to obtain a fused feature vector.

[0117] The recommendation score determination module 203 determines the recommendation scores of the data to be labeled and each label based on the pre-built label association model and the fused feature vector.

[0118] The target label determination module 204 determines the target label corresponding to the data to be labeled based on the recommendation score.

[0119] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0120] Figure 3 This is a schematic diagram of the structure of a vehicle provided in an embodiment of this application.

[0121] For example, such as Figure 3 As shown, the vehicle includes a memory 301 and a processor 302. The memory 301 stores executable program code 3011, and the processor 302 is used to call and execute the executable program code 3011 to execute the tag recommendation method.

[0122] This embodiment can divide the vehicle into functional modules according to the above method example. For example, each function can be assigned to a separate module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0123] When each functional module is divided according to its corresponding function, the vehicle may include: The data feature extraction module extracts multidimensional data features from the data to be labeled, including text features, image features, and structured data features.

[0124] The data feature fusion module performs feature fusion on the multidimensional data features to obtain a fused feature vector.

[0125] The recommendation score determination module determines the recommendation scores of the data to be labeled and each label based on the pre-built label association model and the fused feature vector.

[0126] The target label determination module determines the target label corresponding to the data to be labeled based on the recommendation score.

[0127] It should be noted that all relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0128] The vehicle provided in this embodiment is used to execute the above-described tag recommendation method, and therefore can achieve the same effect as the above-described implementation method.

[0129] When using integrated units, the vehicle may include a processing module and a storage module. The processing module is used to control and manage the vehicle's actions. The storage module supports the vehicle in executing program code and data.

[0130] The processing module may be a processor or a controller, which can implement or execute various exemplary logic blocks, modules, and circuits as disclosed in this application. The processor may also be a combination of computing functions, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and a microprocessor, etc., and the storage module may be a memory.

[0131] This embodiment also provides a computer-readable storage medium (including but not limited to disk storage, CD-ROM, optical storage, etc.) storing computer program code. When the computer program code is run on a computer, the computer executes the above-described related method steps to implement the label recommendation method provided in the above embodiment.

[0132] This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to implement the tag recommendation method provided in the above embodiment.

[0133] The beneficial effects of the above embodiments can be referred to the beneficial effects of the corresponding methods provided above, and will not be repeated here.

[0134] Through the above description of the embodiments, those skilled in the art will understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0135] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0136] In the description of this disclosure, it should be understood that if the terms "upper", "lower", "front", "rear", "left" and "right" are used to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the position or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this disclosure.

[0137] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0138] The above are merely embodiments of this disclosure and are not intended to limit the scope of this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of the claims of this disclosure.

Claims

1. A tag recommendation method, characterized in that, include: Extract multidimensional data features from the data to be labeled, including text features, image features, and structured data features; The multidimensional data features are fused to obtain a fused feature vector; Based on the pre-built tag association model and the fused feature vector, the recommendation scores of the data to be labeled and each tag are determined. The target label corresponding to the data to be labeled is determined based on the recommended score.

2. The method according to claim 1, characterized in that, The extraction of multidimensional data features from the data to be labeled specifically includes: Keyword extraction is performed on the text content based on the weights of word frequency statistics and inverse document frequency to obtain the first text feature; the text content is mapped to a continuous vector space to obtain the contextual semantic representation to obtain the second text feature; the first text feature and the second text feature are merged to obtain the text feature. The low-level visual features of the image are extracted, and the high-level semantic features are obtained by pooling and weighting the output features of the pre-trained model; the low-level visual features include at least one of color features, texture features, and shape features; the low-level visual features and the high-level semantic features are merged to obtain the image features; Encode structured fields and extract statistical features between fields to obtain structured data features; The field encoding includes at least one of one-hot encoding and field name embedding encoding; the statistical features include at least one of mean, variance, number of unique values, and correlation coefficient between fields.

3. The method according to claim 1, characterized in that, The step of fusing the multidimensional data features to obtain a fused feature vector includes: The multidimensional data features are weighted and fused to obtain a weighted feature vector; And / or, concatenate feature vectors from different modalities to obtain a concatenated feature vector; And / or, calculate the interaction feature vectors between feature vectors of different modalities; The fused feature vector is generated based on at least one of the weighted feature vector, the concatenated feature vector, and the interactive feature vector.

4. The method according to claim 1, characterized in that, Before determining the recommendation scores of the data to be labeled and each label based on the pre-built label association model and the fused feature vector, the method further includes: Based on historical tagging records, multidimensional interactive data is constructed, which is used to represent the interaction intensity between users, data, and tags. The multidimensional interactive data is decomposed into multiple low-dimensional factor matrices, and the low-dimensional factor matrices include at least a user factor matrix, a data factor matrix, and a tag factor matrix. A label association model is constructed based on the aforementioned multiple low-dimensional factor matrices.

5. The method according to claim 4, characterized in that, The determination of the recommendation scores for the data to be labeled and each label based on the pre-built label association model and the fused feature vector specifically includes: Construct a graph structure containing data nodes, label nodes, and user nodes, where edges in the graph structure are used to represent the interaction relationships between the nodes; Based on the label association model, representation learning is performed on the nodes in the graph structure; The fused feature vector is used as the representation of the data to be labeled in the latent space corresponding to the label association model; Based on the fused feature vector and the learned graph structure, the recommendation scores of the data to be labeled and each label are determined.

6. The method according to claim 5, characterized in that, The step of determining the recommendation scores of the data to be labeled and each label based on the fused feature vector and the learned graph structure specifically includes: The fused feature vector is input into the data node representation in the graph structure; By using a multi-layer graph convolutional network for message passing and node representation updating, the updated data node representation and label node representation are obtained. Calculate the cosine similarity between the data node representation and the tag node representation; The cosine similarity is converted into a probability score using a preset function, which is then used as the recommendation score.

7. The method according to claim 5, characterized in that, After determining the target label corresponding to the data to be labeled based on the recommendation score, the method further includes: In response to receiving a user's confirmation or correction operation on the target label, the user's operation is recorded as a new historical tagging record and updated in the multidimensional interactive data; Based on the updated multidimensional interaction data, the user factor matrix, data factor matrix, and tag factor matrix are incrementally updated. Based on the updated factor matrix, the node representations and parameters of the graph structure are updated.

8. The method according to claim 5, characterized in that, The step of determining the target label corresponding to the data to be labeled based on the recommendation score includes: Sort the tags in descending order of their recommendation scores; Select the first preset positive integer number of labels from the sorting results as the target labels and output them.

9. The method according to claim 5, characterized in that, The step of determining the target label corresponding to the data to be labeled based on the recommendation score includes: If the recommended scores for all tags are below a preset threshold, the user is prompted to enter a new tag. Add the new tag entered by the user to the system tag collection; The interaction records between the new label and the data to be labeled are added as new data and updated to the multidimensional interaction data.

10. A vehicle, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the tag recommendation method as described in any one of claims 1 to 8.

Images